Research on preference optimization often varies the training objective while holding the data fixed. We instead ask whether a small, high-confidence set of on-policy responses can provide a reliable learning signal. Our method, DMAPO (Data-centric Multi-evaluator Agreement for Preference Optimization), generates candidate responses from the target policy, evaluates helpfulness, factuality, and conciseness with rubric-specialized evaluators, applies a process-critic correction, and retains only high-consensus desirable or undesirable examples. This procedure accepts 1,871 of 54,236 Mistral-7B candidates (3.45%). KTO trained on this set reaches 7.50 on MT-Bench, 95.5% length-controlled win rate against a text-davinci-003 reference, and 57.3% IFEval prompt accuracy. Independent pairwise evaluation also favors DMAPO over SimPO: GPT-4o yields a net win rate of 23.3 points on 129 held-out prompts and 24.0 points on 200 out-of-distribution LMSYS-Chat prompts; Claude Opus 4.7 yields 24.1 points on the held-out set. Changing the evaluator model or rubric alters the selected examples but has little effect on downstream performance. A second-backbone study yields a similar 3.41% acceptance rate, although its performance gains are more modest. Across these experiments, consensus filtering offers a data-efficient route to preference optimization for general instructions, at the cost of additional curation compute and dependence on evaluator judgments.
展开完整摘要收起摘要↓
Research on preference optimization often varies the training objective while holding the data fixed. We instead ask whether a small, high-confidence set of on-policy responses can provide a reliable learning signal. Our method, DMAPO (Data-centric Multi-evaluator Agreement for Preference Optimization), generates candidate responses from the target policy, evaluates helpfulness, factuality, and conciseness with rubric-specialized evaluators, applies a process-critic correction, and retains only high-consensus desirable or undesirable examples. This procedure accepts 1,871 of 54,236 Mistral-7B candidates (3.45%). KTO trained on this set reaches 7.50 on MT-Bench, 95.5% length-controlled win rate against a text-davinci-003 reference, and 57.3% IFEval prompt accuracy. Independent pairwise evaluation also favors DMAPO over SimPO: GPT-4o yields a net win rate of 23.3 points on 129 held-out prompts and 24.0 points on 200 out-of-distribution LMSYS-Chat prompts; Claude Opus 4.7 yields 24.1 points on the held-out set. Changing the evaluator model or rubric alters the selected examples but has little effect on downstream performance. A second-backbone study yields a similar 3.41% acceptance rate, although its performance gains are more modest. Across these experiments, consensus filtering offers a data-efficient route to preference optimization for general instructions, at the cost of additional curation compute and dependence on evaluator judgments.
Language model alignment aims to make model behavior reliably reflect desirable properties such as helpfulness, safety, and instruction following. Current approaches typically use supervised fine-tuning on demonstrations or reinforcement learning with rewards derived from verifiers or human feedback. These paradigms leave an important question underexplored: can demonstrations alone yield an implicit reward that can be inspected, reused, and optimized on-policy to align AI? Motivated by inverse reinforcement learning, we introduce Projected Alignment Reward Estimated from Demonstrations (PARED). PARED recovers the implicit reward underlying expert demonstrations as an explicit function over a small set of response-level features, learned by a lightweight discriminator that separates demonstrations from the policy's own samples in this feature space. Unlike a standard reward model, PARED requires no task-specific preference annotations: demonstrations provide the task-specific supervision, which can be augmented with AI feedback as additional dimensions of supervision. Through experiments involving inference-time reranking and adversarial on-policy RL, we show that the recovered reward improves a base policy without a supervised loss and yields further gains when optimized after standard supervised fine-tuning. Additionally, we demonstrate that PARED can be used for contextual alignment, in which a single policy can be tailored to the preferences of different audiences.
展开完整摘要收起摘要↓
Language model alignment aims to make model behavior reliably reflect desirable properties such as helpfulness, safety, and instruction following. Current approaches typically use supervised fine-tuning on demonstrations or reinforcement learning with rewards derived from verifiers or human feedback. These paradigms leave an important question underexplored: can demonstrations alone yield an implicit reward that can be inspected, reused, and optimized on-policy to align AI? Motivated by inverse reinforcement learning, we introduce Projected Alignment Reward Estimated from Demonstrations (PARED). PARED recovers the implicit reward underlying expert demonstrations as an explicit function over a small set of response-level features, learned by a lightweight discriminator that separates demonstrations from the policy's own samples in this feature space. Unlike a standard reward model, PARED requires no task-specific preference annotations: demonstrations provide the task-specific supervision, which can be augmented with AI feedback as additional dimensions of supervision. Through experiments involving inference-time reranking and adversarial on-policy RL, we show that the recovered reward improves a base policy without a supervised loss and yields further gains when optimized after standard supervised fine-tuning. Additionally, we demonstrate that PARED can be used for contextual alignment, in which a single policy can be tailored to the preferences of different audiences.
The alignment of Small Language Models (SLMs) in the 70--500M parameter range using reinforcement learning is often considered unstable, though the underlying failure mechanisms have not been systematically investigated. In the State-of-the-Art (SOTA) research, fifteen (model, corpus) configurations were trained using Proximal Policy Optimization (PPO). The experiments included Pythia-70M, 160M, 410M and SmolLM2-135M, 360M on the TinyStories, CNN/DailyMail, and Wikitext-103 corpora. Three reproducible failure modes were identified in small-scale language models: silent LoRA parameter freezing in standard PEFT/TRL pipelines, numerical overflow in importance ratios when using bfloat16, and catastrophic policy collapse due to reward-model error. These issues were addressed using a merge-and-reinitialize adapter technique, float32 precision during PPO updates, and a three-layer safety mechanism comprising reward whitening, importance-ratio guarding, and weight rollback. In this paper, a capacity-headroom hypothesis is proposed, which states that PPO performance at the SLM scale depends on both a fluent supervised model ($\text{PPL}<20$) and a discriminative reward signal, rather than on the number of model parameters. The proposed system converged stably in all experiments and improved preference win rate over the SFT baseline in configurations with a fluent prior and an informative reward signal. Furthermore, it outperformed instruction-tuned baselines while requiring significantly less training data. All checkpoints, preference datasets, and training scripts are publicly released$^§$.
展开完整摘要收起摘要↓
The alignment of Small Language Models (SLMs) in the 70--500M parameter range using reinforcement learning is often considered unstable, though the underlying failure mechanisms have not been systematically investigated. In the State-of-the-Art (SOTA) research, fifteen (model, corpus) configurations were trained using Proximal Policy Optimization (PPO). The experiments included Pythia-70M, 160M, 410M and SmolLM2-135M, 360M on the TinyStories, CNN/DailyMail, and Wikitext-103 corpora. Three reproducible failure modes were identified in small-scale language models: silent LoRA parameter freezing in standard PEFT/TRL pipelines, numerical overflow in importance ratios when using bfloat16, and catastrophic policy collapse due to reward-model error. These issues were addressed using a merge-and-reinitialize adapter technique, float32 precision during PPO updates, and a three-layer safety mechanism comprising reward whitening, importance-ratio guarding, and weight rollback. In this paper, a capacity-headroom hypothesis is proposed, which states that PPO performance at the SLM scale depends on both a fluent supervised model ($\text{PPL}<20$) and a discriminative reward signal, rather than on the number of model parameters. The proposed system converged stably in all experiments and improved preference win rate over the SFT baseline in configurations with a fluent prior and an informative reward signal. Furthermore, it outperformed instruction-tuned baselines while requiring significantly less training data. All checkpoints, preference datasets, and training scripts are publicly released$^§$.
Standard vision-language models (VLMs) suffer from Moravec's paradox: they excel at complex offline visual reasoning but struggle with simple streaming perception tasks and process them inefficiently. We present Mage-VL, an efficient codec-native streaming foundation model for real-time multimodal understanding and interaction. At its core, our custom tokenizer, Mage-ViT, replaces uniform frame sampling by selectively encoding dynamic, entropy-rich regions using motion vectors and residual energy across sparse anchor (I) and predicted (P) frames. Operating at a 16 x 16 patch level, this reduces visual token consumption by over 75% while preserving spatiotemporal context. Trained from scratch on approximately 560M unlabeled images and 100M unlabeled video frames, Mage-ViT matches or outperforms flagship encoders trained on billions of image-text pairs. We establish AI4AI data pipelines encompassing prompt-code joint optimization for multimodal captioning and AI-driven performance diagnosis to guide training recipes. Furthermore, through a bio-inspired dual-system architecture - a lightweight System 1 event gate and a causal System 2 decoder - Mage-VL enables proactive streaming perception. Extensive evaluations show that Mage-VL-4B matches Qwen3-VL-4B on static tasks while achieving strong gains in video understanding and 2D/3D spatial reasoning, with up to a 3.5x wall-clock inference speedup, and comprehensively surpasses the 15B Phi-4-reasoning-vision baseline. Beyond model artifacts, we deliver seven key empirical findings covering pre-training data efficiency, variable-resolution scaling, codec system acceleration, VideoQA SFT redundancy, motion-spatial synergy, AI4AI data pipelines, and Zero-Vision SFT for multimodal RL.
展开完整摘要收起摘要↓
Standard vision-language models (VLMs) suffer from Moravec's paradox: they excel at complex offline visual reasoning but struggle with simple streaming perception tasks and process them inefficiently. We present Mage-VL, an efficient codec-native streaming foundation model for real-time multimodal understanding and interaction. At its core, our custom tokenizer, Mage-ViT, replaces uniform frame sampling by selectively encoding dynamic, entropy-rich regions using motion vectors and residual energy across sparse anchor (I) and predicted (P) frames. Operating at a 16 x 16 patch level, this reduces visual token consumption by over 75% while preserving spatiotemporal context. Trained from scratch on approximately 560M unlabeled images and 100M unlabeled video frames, Mage-ViT matches or outperforms flagship encoders trained on billions of image-text pairs. We establish AI4AI data pipelines encompassing prompt-code joint optimization for multimodal captioning and AI-driven performance diagnosis to guide training recipes. Furthermore, through a bio-inspired dual-system architecture - a lightweight System 1 event gate and a causal System 2 decoder - Mage-VL enables proactive streaming perception. Extensive evaluations show that Mage-VL-4B matches Qwen3-VL-4B on static tasks while achieving strong gains in video understanding and 2D/3D spatial reasoning, with up to a 3.5x wall-clock inference speedup, and comprehensively surpasses the 15B Phi-4-reasoning-vision baseline. Beyond model artifacts, we deliver seven key empirical findings covering pre-training data efficiency, variable-resolution scaling, codec system acceleration, VideoQA SFT redundancy, motion-spatial synergy, AI4AI data pipelines, and Zero-Vision SFT for multimodal RL.
State-of-the-art toxicity detectors for text-to-image generation adopt a one-size-fits-all approach: a single universal model applying fixed safety guidelines to all users. Our empirical evidence shows that these detectors fail to shield marginalized communities: approximately 35% of generated images labeled safe are considered harmful by disability communities. In this position paper, we argue for community-specific toxicity detection (CTD). To demonstrate its feasibility, we collaborate with disability experts to develop safety guidelines for two communities: dwarfism and blind/low vision. Using a dataset of 2,400 annotated T2I-generated images we demonstrate that both large vision-language models and existing general-purpose toxicity detectors catastrophically fail to recognize harmful content under these guidelines in zero-shot settings with F1 score lower than random guessing (F1 0.32 and 0.37). Promisingly, prompt-based adaptation methods (ICL, VQA) substantially improve harm detection performance (GPT-4o: F1 0.50 and 0.78), while parameter-efficient fine-tuning improves smaller models (0.5b-7b with best F1 0.48 and 0.59) with less than 100 demonstrations, but remains sensitive to evolving guidelines. Despite these gains, CTD performance remains far below F1 $\approx 0.9$ achieved for general-purpose toxicity detection, highlighting the challenge and the need for sustained research effort.
展开完整摘要收起摘要↓
State-of-the-art toxicity detectors for text-to-image generation adopt a one-size-fits-all approach: a single universal model applying fixed safety guidelines to all users. Our empirical evidence shows that these detectors fail to shield marginalized communities: approximately 35% of generated images labeled safe are considered harmful by disability communities. In this position paper, we argue for community-specific toxicity detection (CTD). To demonstrate its feasibility, we collaborate with disability experts to develop safety guidelines for two communities: dwarfism and blind/low vision. Using a dataset of 2,400 annotated T2I-generated images we demonstrate that both large vision-language models and existing general-purpose toxicity detectors catastrophically fail to recognize harmful content under these guidelines in zero-shot settings with F1 score lower than random guessing (F1 0.32 and 0.37). Promisingly, prompt-based adaptation methods (ICL, VQA) substantially improve harm detection performance (GPT-4o: F1 0.50 and 0.78), while parameter-efficient fine-tuning improves smaller models (0.5b-7b with best F1 0.48 and 0.59) with less than 100 demonstrations, but remains sensitive to evolving guidelines. Despite these gains, CTD performance remains far below F1 $\approx 0.9$ achieved for general-purpose toxicity detection, highlighting the challenge and the need for sustained research effort.
作者Taimoor Rizwan, Sara Atito, Zhenhua Feng, Muhammad Awais, Josef Kittler
Face morphing attacks create synthetic images verifiable against multiple identities, threatening border control and identity verification systems. We introduce MorphUNet, a diffusion morphing framework formulating two-parent generation as alpha-controlled biometric transport: each parent is decomposed into CLIP appearance and ArcFace identity evidence, aligned into a CLIP-compatible token space, with the two contributors preserved as separate identity-aware token banks. To our knowledge, MorphUNet is the first diffusion-based morphing framework using trainable parent-separated dual cross-attention inside the denoising U-Net: a Biometric Transport Layer carrying parent-specific identity evidence through denoising, attending to each parent separately before combining residuals via the morphing parameter alpha. DDIM-inverted latent interpolation gives a coherent denoising start, while weaker-parent-guided selection favours morphs maximising the lower parent-similarity score, reducing collapse toward one contributor. We evaluate MorphUNet against three state-of-the-art baselines (StableMorph, MIPGAN-II, and MorDIFF) on FEI and FRLL using six recognition systems, and propose CFD-based unseen-identity stress testing across gender and ethnicity pairing, demographic shifts, and parent-similarity extremes. MorphUNet achieves the best Morphing Attack Potential (MAP) when at least three of six systems are fooled by one morph, reaching 0.919 on FEI and 0.886 on FRLL, and obtains the best FID on both datasets (35.19 FEI, 44.86 FRLL). It also gives the highest APCER at 5% BPCER in the same-dataset setting, and remains highly difficult to detect under cross-dataset transfer, with APCER 0.996 on FEI and 0.946 on FRLL. The full evaluation analyses MAP, MAD, per-system vulnerability, identity balance, image quality, top/bottom-similarity stress tests, and CFD unseen-identity robustness.
展开完整摘要收起摘要↓
Face morphing attacks create synthetic images verifiable against multiple identities, threatening border control and identity verification systems. We introduce MorphUNet, a diffusion morphing framework formulating two-parent generation as alpha-controlled biometric transport: each parent is decomposed into CLIP appearance and ArcFace identity evidence, aligned into a CLIP-compatible token space, with the two contributors preserved as separate identity-aware token banks. To our knowledge, MorphUNet is the first diffusion-based morphing framework using trainable parent-separated dual cross-attention inside the denoising U-Net: a Biometric Transport Layer carrying parent-specific identity evidence through denoising, attending to each parent separately before combining residuals via the morphing parameter alpha. DDIM-inverted latent interpolation gives a coherent denoising start, while weaker-parent-guided selection favours morphs maximising the lower parent-similarity score, reducing collapse toward one contributor. We evaluate MorphUNet against three state-of-the-art baselines (StableMorph, MIPGAN-II, and MorDIFF) on FEI and FRLL using six recognition systems, and propose CFD-based unseen-identity stress testing across gender and ethnicity pairing, demographic shifts, and parent-similarity extremes. MorphUNet achieves the best Morphing Attack Potential (MAP) when at least three of six systems are fooled by one morph, reaching 0.919 on FEI and 0.886 on FRLL, and obtains the best FID on both datasets (35.19 FEI, 44.86 FRLL). It also gives the highest APCER at 5% BPCER in the same-dataset setting, and remains highly difficult to detect under cross-dataset transfer, with APCER 0.996 on FEI and 0.946 on FRLL. The full evaluation analyses MAP, MAD, per-system vulnerability, identity balance, image quality, top/bottom-similarity stress tests, and CFD unseen-identity robustness.
作者Farzana Yasmin Ahmad, Vanamala Venkataswamy, Geoffrey Fox
Monte Carlo simulation of calorimeter showers is a principal bottleneck for the High-Luminosity LHC, and diffusion models have emerged as fast, high-fidelity surrogates. Their denoising objective is purely statistical, however: a model can minimize it while placing the physics wrong. Existing physics-informed generative methods cannot close this gap, because they assume a closed-form law, a governing PDE residual or a hard per-sample constraint, that a shower does not supply: no per-sample PDE governs a stochastic cascade, and energy conservation fixes only one scalar per shower. Standard metrics ignore the correlation structure across calorimeter layers and voxels, comparing showers only in a physics feature space. We address both gaps. We introduce the Correlation Frobenius Distance (CFD), a single normalized score for correlation fidelity at layer-wise and voxel-wise scales. We then encode the soft per-sample structure available in a shower as two physics-aware auxiliary losses: a variance-stabilized voxel residual loss grounded in counting statistics, and a graph Laplacian loss over the detector geometry. We combine both with denoising through GradBlend, which anchors the step magnitude to the denoising gradient while letting the auxiliary steer its direction, yielding Lantern, a physics-guided diffusion surrogate. On CaloChallenge Dataset 2, injecting the physics losses through task-symmetric rules such as PCGrad, GradNorm, IMTL-G, and ConFIG inflates FPD by 2-100x relative to denoising alone, whereas GradBlend admits the same signal without regression and, with the Laplacian loss, Lantern improves both FPD and CFD. Our ablation on the auxiliary loss scheduler shows that the voxel residual loss, whose gradient conflicts with denoising, requires a terminal denoising-only phase to preserve shower fidelity, whereas the non-conflicting Laplacian loss is insensitive to the schedule.
展开完整摘要收起摘要↓
Monte Carlo simulation of calorimeter showers is a principal bottleneck for the High-Luminosity LHC, and diffusion models have emerged as fast, high-fidelity surrogates. Their denoising objective is purely statistical, however: a model can minimize it while placing the physics wrong. Existing physics-informed generative methods cannot close this gap, because they assume a closed-form law, a governing PDE residual or a hard per-sample constraint, that a shower does not supply: no per-sample PDE governs a stochastic cascade, and energy conservation fixes only one scalar per shower. Standard metrics ignore the correlation structure across calorimeter layers and voxels, comparing showers only in a physics feature space. We address both gaps. We introduce the Correlation Frobenius Distance (CFD), a single normalized score for correlation fidelity at layer-wise and voxel-wise scales. We then encode the soft per-sample structure available in a shower as two physics-aware auxiliary losses: a variance-stabilized voxel residual loss grounded in counting statistics, and a graph Laplacian loss over the detector geometry. We combine both with denoising through GradBlend, which anchors the step magnitude to the denoising gradient while letting the auxiliary steer its direction, yielding Lantern, a physics-guided diffusion surrogate. On CaloChallenge Dataset 2, injecting the physics losses through task-symmetric rules such as PCGrad, GradNorm, IMTL-G, and ConFIG inflates FPD by 2-100x relative to denoising alone, whereas GradBlend admits the same signal without regression and, with the Laplacian loss, Lantern improves both FPD and CFD. Our ablation on the auxiliary loss scheduler shows that the voxel residual loss, whose gradient conflicts with denoising, requires a terminal denoising-only phase to preserve shower fidelity, whereas the non-conflicting Laplacian loss is insensitive to the schedule.
Generative semantic communication receivers deliver high perceptual quality but suffer from prohibitive decoding latency. This bottleneck arises because diffusion receivers rely on stochastic iterative decoding, while existing flow matching receivers employ independent endpoint coupling that ignores the physical source--channel link, yielding unnecessarily long and curved sampling trajectories. In this paper, we reformulate receiver-side recovery as a realization-coupled bridge flow matching problem under explicit bandwidth and power constraints. Specifically, we propose Realization-Coupled Bridge Flow Matching (RC-BFM), where the decoder initializes from a channel-induced semantic state rather than isotropic noise. Crucially, training pairs are linked via a realization-coupled entropic optimal transport (RC-OT) plan that preserves the physical channel realization of each transmission while maintaining robustness to stochastic fading. Furthermore, we identify independent coupling as the fundamental source of a conditional train--test distribution shift in conditional flow matching-based receivers, and derive an end-to-end distortion bound whose discretization error decays as \(O(K^{-2})\). Experiments on CIFAR-10 and FFHQ-64\(\times\)64 over AWGN and Rayleigh fading channels demonstrate that RC-BFM achieves a superior fidelity--perception trade-off, reducing decoding latency by over 10\(\times\) compared to diffusion-based receivers.
展开完整摘要收起摘要↓
Generative semantic communication receivers deliver high perceptual quality but suffer from prohibitive decoding latency. This bottleneck arises because diffusion receivers rely on stochastic iterative decoding, while existing flow matching receivers employ independent endpoint coupling that ignores the physical source--channel link, yielding unnecessarily long and curved sampling trajectories. In this paper, we reformulate receiver-side recovery as a realization-coupled bridge flow matching problem under explicit bandwidth and power constraints. Specifically, we propose Realization-Coupled Bridge Flow Matching (RC-BFM), where the decoder initializes from a channel-induced semantic state rather than isotropic noise. Crucially, training pairs are linked via a realization-coupled entropic optimal transport (RC-OT) plan that preserves the physical channel realization of each transmission while maintaining robustness to stochastic fading. Furthermore, we identify independent coupling as the fundamental source of a conditional train--test distribution shift in conditional flow matching-based receivers, and derive an end-to-end distortion bound whose discretization error decays as \(O(K^{-2})\). Experiments on CIFAR-10 and FFHQ-64\(\times\)64 over AWGN and Rayleigh fading channels demonstrate that RC-BFM achieves a superior fidelity--perception trade-off, reducing decoding latency by over 10\(\times\) compared to diffusion-based receivers.
作者Taimoor Rizwan, Sara Atito, Muhammad Awais, Zhenhua Feng, Josef Kittler
Generative diffusion models have revolutionized facial image synthesis, yet robust identity preservation in high resolution outputs remains a critical challenge. This issue is especially vital for security systems, biometric authentication, and privacy sensitive applications, where any drift in identity integrity can undermine trust and functionality. We introduce Diff-ID, a diffusion based framework that enforces identity consistency while delivering photorealistic quality. Central to our approach is a custom 210K image dataset synthesized from CelebA-HQ, FFHQ, and LAION-Face and captioned via a fine tuned BLIP model to bolster identity awareness during training. Diff-ID integrates ArcFace and CLIP embeddings through a dual cross attention adapter within a fine tuned Stable Diffusion UNet. To further reinforce identity fidelity, we propose a pseudo discriminator loss based on ArcFace cosine similarity with exponential timestep weighting. Experiments on held out and unseen faces show that Diff-ID does not exceed InstantID in raw ArcFace Face Similarity, but achieves substantially lower FID and the strongest FIQ based identity--realism trade off among the evaluated methods. We also present a unified DDIM based morphing pipeline that enables qualitative facial interpolation without per identity fine tuning. We further argue that identity preservation and photorealism should be evaluated jointly rather than in isolation, as high identity similarity alone does not guarantee realistic outputs. To make this trade off explicit, we report Face Image Quality (FIQ) as a complementary ratio based score that combines identity similarity and perceptual realism while keeping FS and FID as the primary metrics.
展开完整摘要收起摘要↓
Generative diffusion models have revolutionized facial image synthesis, yet robust identity preservation in high resolution outputs remains a critical challenge. This issue is especially vital for security systems, biometric authentication, and privacy sensitive applications, where any drift in identity integrity can undermine trust and functionality. We introduce Diff-ID, a diffusion based framework that enforces identity consistency while delivering photorealistic quality. Central to our approach is a custom 210K image dataset synthesized from CelebA-HQ, FFHQ, and LAION-Face and captioned via a fine tuned BLIP model to bolster identity awareness during training. Diff-ID integrates ArcFace and CLIP embeddings through a dual cross attention adapter within a fine tuned Stable Diffusion UNet. To further reinforce identity fidelity, we propose a pseudo discriminator loss based on ArcFace cosine similarity with exponential timestep weighting. Experiments on held out and unseen faces show that Diff-ID does not exceed InstantID in raw ArcFace Face Similarity, but achieves substantially lower FID and the strongest FIQ based identity--realism trade off among the evaluated methods. We also present a unified DDIM based morphing pipeline that enables qualitative facial interpolation without per identity fine tuning. We further argue that identity preservation and photorealism should be evaluated jointly rather than in isolation, as high identity similarity alone does not guarantee realistic outputs. To make this trade off explicit, we report Face Image Quality (FIQ) as a complementary ratio based score that combines identity similarity and perceptual realism while keeping FS and FID as the primary metrics.
Generative models are increasingly adopted in distributionally robust optimization (DRO), but existing approaches trade off model compatibility and adversarial structure: methods that accept arbitrary samplers do not restrict worst-case laws to a generator family, while generator-parameterized adversaries rely on model-specific access such as likelihoods, scores, or training data. We propose Generative Distributionally Robust Optimization (GDRO), a principled framework that accepts any sampleable conditional generator as the nominal model and restricts worst-case laws to a chosen conditional generator family. The key is the sampler-Sinkhorn pairing: samplers represent the conditional laws exactly, while Sinkhorn divergence compares their induced distributions without likelihood access and can be estimated from samples alone. The resulting population problem admits a direct finite-sample approximation and differentiable primal-dual implementation at the active decision context. For Lipschitz losses, the population Sinkhorn radius bounds downstream degradation. Across explicit and implicit generators, our method reduces rare-context inventory regret by 60% and SocialGAN navigation collisions by 50% relative to nominal decisions.
展开完整摘要收起摘要↓
Generative models are increasingly adopted in distributionally robust optimization (DRO), but existing approaches trade off model compatibility and adversarial structure: methods that accept arbitrary samplers do not restrict worst-case laws to a generator family, while generator-parameterized adversaries rely on model-specific access such as likelihoods, scores, or training data. We propose Generative Distributionally Robust Optimization (GDRO), a principled framework that accepts any sampleable conditional generator as the nominal model and restricts worst-case laws to a chosen conditional generator family. The key is the sampler-Sinkhorn pairing: samplers represent the conditional laws exactly, while Sinkhorn divergence compares their induced distributions without likelihood access and can be estimated from samples alone. The resulting population problem admits a direct finite-sample approximation and differentiable primal-dual implementation at the active decision context. For Lipschitz losses, the population Sinkhorn radius bounds downstream degradation. Across explicit and implicit generators, our method reduces rare-context inventory regret by 60% and SocialGAN navigation collisions by 50% relative to nominal decisions.
Autonomous Underwater Vehicles (AUVs) operate in complex, unstructured environments where efficient and safe path planning is critical for mission success and energy conservation. This paper presents a comprehensive comparative evaluation of thirteen path planning algorithms, ranging from classical graph-search methods (A*, Dijkstra) and sampling-based approaches (RRT*) to metaheuristics (PSO, GA, ACO, BCO) and learning-based architectures. Special emphasis is placed on a proposed hybrid approach combining Artificial Potential Fields (APF) with a Spatio-Temporal (ST) Transformer. Evaluated across five navigation scenarios on high-resolution underwater terrain maps, all algorithms achieved 100% task completion; however, significant trade-offs emerged in path optimality, collision avoidance, and computational load. The Hybrid APF + ST-Transformer demonstrated superior balanced performance, achieving the shortest average path length (943.15 units), a low collision rate (0.031), and efficient computation time (0.96 s), outperforming standalone learning models, which required fallback mechanisms and classical methods that incurred higher latency. While classical algorithms guaranteed collision-free paths, their excessive path lengths and processing times render them less suitable for dynamic underwater operations. Conversely, metaheuristic approaches introduced trajectory complexity unsuitable for strict energy constraints. Based on these findings, the Hybrid APF + ST framework is recommended as a principal approach for real-time AUV navigation, offering a robust solution that harmonizes reactive obstacle avoidance with global path optimality in resource-constrained underwater systems.
展开完整摘要收起摘要↓
Autonomous Underwater Vehicles (AUVs) operate in complex, unstructured environments where efficient and safe path planning is critical for mission success and energy conservation. This paper presents a comprehensive comparative evaluation of thirteen path planning algorithms, ranging from classical graph-search methods (A*, Dijkstra) and sampling-based approaches (RRT*) to metaheuristics (PSO, GA, ACO, BCO) and learning-based architectures. Special emphasis is placed on a proposed hybrid approach combining Artificial Potential Fields (APF) with a Spatio-Temporal (ST) Transformer. Evaluated across five navigation scenarios on high-resolution underwater terrain maps, all algorithms achieved 100% task completion; however, significant trade-offs emerged in path optimality, collision avoidance, and computational load. The Hybrid APF + ST-Transformer demonstrated superior balanced performance, achieving the shortest average path length (943.15 units), a low collision rate (0.031), and efficient computation time (0.96 s), outperforming standalone learning models, which required fallback mechanisms and classical methods that incurred higher latency. While classical algorithms guaranteed collision-free paths, their excessive path lengths and processing times render them less suitable for dynamic underwater operations. Conversely, metaheuristic approaches introduced trajectory complexity unsuitable for strict energy constraints. Based on these findings, the Hybrid APF + ST framework is recommended as a principal approach for real-time AUV navigation, offering a robust solution that harmonizes reactive obstacle avoidance with global path optimality in resource-constrained underwater systems.
Long-running autonomous agents plan, act, and judge their own completion without human intervention. When an agent grades its own work, self-evaluation bias takes hold: plausible changes are accepted as progress while real-world outcomes stagnate or regress. We name this failure mode the progress mirage and show, with controlled measurement, that it is a question of what the evaluator is grounded in. We built a testbed that holds the agent and its tool surface fixed and manipulates only the information-channel type of the evaluator that gates the loop. A world-state oracle, unfakeable in principle, is enforced by container and network isolation and verified at every run. Across 54 cycles a frontier agent claimed improvement every time, yet 56 percent had a measured delta of zero or below. Self-report was thus uninformative, and the self-verdict gate degenerated into accept-all, eroding the best deployed state it had reached by 19 percent. Even the strongest in-band judge, reading the full artifact text, the change diff, and its own verdict history, accepted cycles of which 44 percent were real-world regressions and rejected 38 percent of real improvements; the preregistered adversarial hypothesis that a strong judge closes the gap was rejected. On a boundary task whose success specification is verifiable from the artifact itself, the same judge's mirage vanished to zero and the gap collapsed within the registered threshold, showing that the gap depends on where the success signal resides. A sign-only variant returning only the acceptance verdict kept real-world output similar to full feedback (110.0 versus 113.0), locating the benefit in the gate's grounding rather than in feedback content. For open-ended objectives whose success signal lives outside the transcript, scaling up the judge is not enough; out-of-band evaluation with real-world access is a structural requirement.
展开完整摘要收起摘要↓
Long-running autonomous agents plan, act, and judge their own completion without human intervention. When an agent grades its own work, self-evaluation bias takes hold: plausible changes are accepted as progress while real-world outcomes stagnate or regress. We name this failure mode the progress mirage and show, with controlled measurement, that it is a question of what the evaluator is grounded in. We built a testbed that holds the agent and its tool surface fixed and manipulates only the information-channel type of the evaluator that gates the loop. A world-state oracle, unfakeable in principle, is enforced by container and network isolation and verified at every run. Across 54 cycles a frontier agent claimed improvement every time, yet 56 percent had a measured delta of zero or below. Self-report was thus uninformative, and the self-verdict gate degenerated into accept-all, eroding the best deployed state it had reached by 19 percent. Even the strongest in-band judge, reading the full artifact text, the change diff, and its own verdict history, accepted cycles of which 44 percent were real-world regressions and rejected 38 percent of real improvements; the preregistered adversarial hypothesis that a strong judge closes the gap was rejected. On a boundary task whose success specification is verifiable from the artifact itself, the same judge's mirage vanished to zero and the gap collapsed within the registered threshold, showing that the gap depends on where the success signal resides. A sign-only variant returning only the acceptance verdict kept real-world output similar to full feedback (110.0 versus 113.0), locating the benefit in the gate's grounding rather than in feedback content. For open-ended objectives whose success signal lives outside the transcript, scaling up the judge is not enough; out-of-band evaluation with real-world access is a structural requirement.
Modern coding agents are usually evaluated by whether they eventually produce a correct patch, but patch generation depends on an earlier context-acquisition stage: finding the repository files needed for the task. We introduce Agent Retrieval Bench, a file-level benchmark for this upstream retrieval problem. Samples are built from real coding-workflow signals and evaluated against frozen base-commit repositories, with relevance defined by what an agent needs next rather than direct query-file semantic similarity. The benchmark covers four positive-retrieval tasks: code2test, comment2context, trace2code, and edit2ripple; a fifth subset evaluates selective retrieval using natural evidence-backed no-gold cases and counterfactual wrong-repository controls. Agent Retrieval Bench contains 427 samples across 25 repositories: 345 positive examples, 50 natural no-gold examples, and 32 counterfactual controls. The corpus includes 308 base-commit snapshots, 392,000 files, and 7.9 million chunks. We evaluate lexical retrieval, RepoMap, open-source embeddings, selective abstention, and logged agent context selection. No single retrieval family dominates: Qwen3-Embedding-4B has the best sample-weighted MRR on positive samples, Qwen3-Embedding-8B the best Recall@20, and RepoMap the best budgeted context yield at 8K tokens, with task-level winners differing substantially. Selective thresholds calibrated with counterfactual controls do not improve selective success on natural no-gold cases, revealing a calibration gap. Logged trajectories also miss every gold file on 27-35 percent of samples. A controlled seed-intervention pilot finds that retrieval-derived initial context yields higher file F1 with less post-seed exploration than random non-gold context, while oracle gold context shows substantial remaining headroom.
展开完整摘要收起摘要↓
Modern coding agents are usually evaluated by whether they eventually produce a correct patch, but patch generation depends on an earlier context-acquisition stage: finding the repository files needed for the task. We introduce Agent Retrieval Bench, a file-level benchmark for this upstream retrieval problem. Samples are built from real coding-workflow signals and evaluated against frozen base-commit repositories, with relevance defined by what an agent needs next rather than direct query-file semantic similarity. The benchmark covers four positive-retrieval tasks: code2test, comment2context, trace2code, and edit2ripple; a fifth subset evaluates selective retrieval using natural evidence-backed no-gold cases and counterfactual wrong-repository controls. Agent Retrieval Bench contains 427 samples across 25 repositories: 345 positive examples, 50 natural no-gold examples, and 32 counterfactual controls. The corpus includes 308 base-commit snapshots, 392,000 files, and 7.9 million chunks. We evaluate lexical retrieval, RepoMap, open-source embeddings, selective abstention, and logged agent context selection. No single retrieval family dominates: Qwen3-Embedding-4B has the best sample-weighted MRR on positive samples, Qwen3-Embedding-8B the best Recall@20, and RepoMap the best budgeted context yield at 8K tokens, with task-level winners differing substantially. Selective thresholds calibrated with counterfactual controls do not improve selective success on natural no-gold cases, revealing a calibration gap. Logged trajectories also miss every gold file on 27-35 percent of samples. A controlled seed-intervention pilot finds that retrieval-derived initial context yields higher file F1 with less post-seed exploration than random non-gold context, while oracle gold context shows substantial remaining headroom.
Agent Skills are an emerging way to extend large language model agents with reusable procedural knowledge that the agent loads on demand. Anthropic introduced Agent Skills and published the format as an open specification supported across several agent tools. This note argues that a skill is a software artefact and that its construction should follow software-engineering principles, with qualifications: single responsibility, separation of interface from implementation, low coupling, and economy in a shared token budget, together with behavioural evaluation in place of deterministic testing. Using Claude Code as the reference implementation, it describes how a skill is structured, how its contents are loaded in stages, and how to write the description on which selection depends. It places skills against the other mechanisms a developer can use to shape agent behaviour, like project memory files, slash commands, subagents, external tool connections, and hooks, and gives a rule for choosing between them based on who decides that a mechanism runs and what guarantee it provides. It then sets out an evaluation-driven authoring process, a set of patterns and faults commonly encountered in authoring, and the trust question raised by using skills from third parties. We illustrate the comparison drawn in UML class style, the loading model, the anatomy of a skill, the relative position of each mechanism, and the points at which skills and hooks act during a session.
展开完整摘要收起摘要↓
Agent Skills are an emerging way to extend large language model agents with reusable procedural knowledge that the agent loads on demand. Anthropic introduced Agent Skills and published the format as an open specification supported across several agent tools. This note argues that a skill is a software artefact and that its construction should follow software-engineering principles, with qualifications: single responsibility, separation of interface from implementation, low coupling, and economy in a shared token budget, together with behavioural evaluation in place of deterministic testing. Using Claude Code as the reference implementation, it describes how a skill is structured, how its contents are loaded in stages, and how to write the description on which selection depends. It places skills against the other mechanisms a developer can use to shape agent behaviour, like project memory files, slash commands, subagents, external tool connections, and hooks, and gives a rule for choosing between them based on who decides that a mechanism runs and what guarantee it provides. It then sets out an evaluation-driven authoring process, a set of patterns and faults commonly encountered in authoring, and the trust question raised by using skills from third parties. We illustrate the comparison drawn in UML class style, the loading model, the anatomy of a skill, the relative position of each mechanism, and the points at which skills and hooks act during a session.
Every major wave of platform software follows the same arc: an initial period of experimentation with competing frameworks and ad-hoc implementations, followed by the articulation of a small set of stable abstractions with well-defined semantics, and finally consolidation around those abstractions into a platform that applications can portably target. POSIX did this for classical operating systems; Kubernetes did it for the cloud. Agentic AI systems - autonomous, LLM-driven agents that plan, use tools, maintain memory, and collaborate - are currently in the experimentation phase of the third such wave. dozens of frameworks and protocols have emerged, but no community consensus exists on what the core abstractions are or what guarantees they carry. Without that consensus, agentic applications cannot be written portably, platforms cannot compose reliably, and the field cannot advance beyond prototype deployments. We argue that the path forward is to follow the prior-wave methodology: derive new agentic abstractions by extending classical OS and cloud OS primitives to stochastic, natural-language-mediated execution, specify their semantics precisely, and consolidate around them - just as POSIX and Kubernetes consolidated their respective waves.
展开完整摘要收起摘要↓
Every major wave of platform software follows the same arc: an initial period of experimentation with competing frameworks and ad-hoc implementations, followed by the articulation of a small set of stable abstractions with well-defined semantics, and finally consolidation around those abstractions into a platform that applications can portably target. POSIX did this for classical operating systems; Kubernetes did it for the cloud. Agentic AI systems - autonomous, LLM-driven agents that plan, use tools, maintain memory, and collaborate - are currently in the experimentation phase of the third such wave. dozens of frameworks and protocols have emerged, but no community consensus exists on what the core abstractions are or what guarantees they carry. Without that consensus, agentic applications cannot be written portably, platforms cannot compose reliably, and the field cannot advance beyond prototype deployments. We argue that the path forward is to follow the prior-wave methodology: derive new agentic abstractions by extending classical OS and cloud OS primitives to stochastic, natural-language-mediated execution, specify their semantics precisely, and consolidate around them - just as POSIX and Kubernetes consolidated their respective waves.
作者Rushi Qiang, Changhao Li, Haotian Sun, Yuchen Zhuang, Chao Zhang, Bo Dai
Machine learning engineering (MLE) tasks require long-horizon decision making over iterative solution debugging and refinement, under expensive and feedback-driven environment interactions. Developing and training a monolithic agent for such tasks is fundamentally challenging, as it must simultaneously manage extremely long and noisy contexts, explore vast solution spaces, and remain effective under limited model capacity and computational budgets. To address these challenges, we propose Matryoshka Agent, a unified hierarchical agent framework for complex long-horizon tasks. Matryoshka Agent decomposes agentic problem solving into a coordinated hierarchy of decision making and execution: a high-level Orchestrator maintains compact, long-horizon exploration states and issues strategic instructions, while lower-level Sub-Agents execute concrete solution attempts through direct environment interaction, mediated by standardized Tool interface. This design decouples strategic exploration from costly execution, substantially reducing the burden of long-context reasoning and enabling efficient iterative refinement. We further develop an efficient training paradigm for Matryoshka Agent. Experimental results on a broad range of MLE tasks with diverse model types and scales demonstrate that Matryoshka Agent is an effective and scalable paradigm for long-horizon MLE tasks and complex agentic problem solving. Notably, Matryoshka Agent enables Qwen3-4B-Instruct to reach Orchestrator performance comparable to o4-mini. Applying Matryoshka Agent to Qwen3-30B-Coder results in at most 36.7% relative performance gain.
展开完整摘要收起摘要↓
Machine learning engineering (MLE) tasks require long-horizon decision making over iterative solution debugging and refinement, under expensive and feedback-driven environment interactions. Developing and training a monolithic agent for such tasks is fundamentally challenging, as it must simultaneously manage extremely long and noisy contexts, explore vast solution spaces, and remain effective under limited model capacity and computational budgets. To address these challenges, we propose Matryoshka Agent, a unified hierarchical agent framework for complex long-horizon tasks. Matryoshka Agent decomposes agentic problem solving into a coordinated hierarchy of decision making and execution: a high-level Orchestrator maintains compact, long-horizon exploration states and issues strategic instructions, while lower-level Sub-Agents execute concrete solution attempts through direct environment interaction, mediated by standardized Tool interface. This design decouples strategic exploration from costly execution, substantially reducing the burden of long-context reasoning and enabling efficient iterative refinement. We further develop an efficient training paradigm for Matryoshka Agent. Experimental results on a broad range of MLE tasks with diverse model types and scales demonstrate that Matryoshka Agent is an effective and scalable paradigm for long-horizon MLE tasks and complex agentic problem solving. Notably, Matryoshka Agent enables Qwen3-4B-Instruct to reach Orchestrator performance comparable to o4-mini. Applying Matryoshka Agent to Qwen3-30B-Coder results in at most 36.7% relative performance gain.
We implement an agentic AI workflow built around a large language model (LLM) agent for autonomous experiments with nitrogen-vacancy (NV) centers in diamond. NV centers are a widely used platform for quantum sensing, and the ability to control many measurements from a computer makes NV experiments a natural setting for autonomous workflows. We make two main contributions. First, we demonstrate an autonomous NV experiment workflow that combines persistent project records, quantitative calculation and data analysis tools, and deterministic experiment control. In one autonomous experiment, the agent selected a single NV center, calibrated its resonant frequency, measured \(T_2^\ast\) with Ramsey measurements, and added a Carr--Purcell--Meiboom--Gill (CPMG) measurement to check a weak feature that could be related to nearby \(^{13}\mathrm{C}\). Second, we introduce two offline benchmarks that evaluate the agent's reasoning separately from laboratory execution. We evaluated both benchmarks with GPT-5.4, GPT-5.5, and GPT-5.6 Sol. In the Ramsey checkpoint benchmark, greater reasoning effort generally improved recognition of a residual resonance calibration offset. By contrast, in the pulsed optically detected magnetic resonance (pODMR) data evaluation benchmark, pulse sequence information alone produced more false positive resonance judgments at higher reasoning effort. Requiring an expected signal calculation kept false positive rates low across all three models and reasoning settings. The results suggest a clear division of labor for autonomous experiments. The agent forms scientific hypotheses and uses quantitative tools to evaluate data, while deterministic code controls the hardware and enforces safety constraints.
展开完整摘要收起摘要↓
We implement an agentic AI workflow built around a large language model (LLM) agent for autonomous experiments with nitrogen-vacancy (NV) centers in diamond. NV centers are a widely used platform for quantum sensing, and the ability to control many measurements from a computer makes NV experiments a natural setting for autonomous workflows. We make two main contributions. First, we demonstrate an autonomous NV experiment workflow that combines persistent project records, quantitative calculation and data analysis tools, and deterministic experiment control. In one autonomous experiment, the agent selected a single NV center, calibrated its resonant frequency, measured \(T_2^\ast\) with Ramsey measurements, and added a Carr--Purcell--Meiboom--Gill (CPMG) measurement to check a weak feature that could be related to nearby \(^{13}\mathrm{C}\). Second, we introduce two offline benchmarks that evaluate the agent's reasoning separately from laboratory execution. We evaluated both benchmarks with GPT-5.4, GPT-5.5, and GPT-5.6 Sol. In the Ramsey checkpoint benchmark, greater reasoning effort generally improved recognition of a residual resonance calibration offset. By contrast, in the pulsed optically detected magnetic resonance (pODMR) data evaluation benchmark, pulse sequence information alone produced more false positive resonance judgments at higher reasoning effort. Requiring an expected signal calculation kept false positive rates low across all three models and reasoning settings. The results suggest a clear division of labor for autonomous experiments. The agent forms scientific hypotheses and uses quantitative tools to evaluate data, while deterministic code controls the hardware and enforces safety constraints.
作者Qihang Wu, Taizun Jafri, Aman Arora, Vidya A. Chhabria
CAD tools typically apply the same fixed, hand-designed algorithms across circuits with widely different structural and timing characteristics. A common way to specialize these one-size-fits-all flows to a target design is to tune the CAD tool's hyperparameters. However, hyperparameter tuning can only select among behaviors already implemented by the fixed algorithm, limiting the achievable quality of results while requiring many expensive place-and-route evaluations. We present VPR-Evolve, a multi-agent framework that specializes Versatile Place and Route (VPR), the open-source FPGA pack-place-and-route engine in the Verilog-to-Routing (VTR) flow, by evolving its source code for each design. VPR-Evolve uses LLM agents to propose, implement, and evaluate code-level modifications, while a shared memory records prior outcomes and guides subsequent evolution. Every candidate is evaluated through a complete VPR build and run, directly optimizing a composite score measured as a weighted function of critical-path delay (CPD), routed wirelength (WL), and tool runtime (RT). Across five VTR-9 benchmark circuits, VPR-Evolve improves the composite score by up to 2.7% over stock VPR in VTR-9. Relative to stock VPR, it reduces CPD by up to 9.8%, routed WL by up to 18.1%, and tool RT by up to 79.3%. VPR-Evolve reduces CPD by up to 6.0%, routed WL by up to 2.2%, and tool RT by up to 7.8% compared with a hyperparameter-tuning baseline.
展开完整摘要收起摘要↓
CAD tools typically apply the same fixed, hand-designed algorithms across circuits with widely different structural and timing characteristics. A common way to specialize these one-size-fits-all flows to a target design is to tune the CAD tool's hyperparameters. However, hyperparameter tuning can only select among behaviors already implemented by the fixed algorithm, limiting the achievable quality of results while requiring many expensive place-and-route evaluations. We present VPR-Evolve, a multi-agent framework that specializes Versatile Place and Route (VPR), the open-source FPGA pack-place-and-route engine in the Verilog-to-Routing (VTR) flow, by evolving its source code for each design. VPR-Evolve uses LLM agents to propose, implement, and evaluate code-level modifications, while a shared memory records prior outcomes and guides subsequent evolution. Every candidate is evaluated through a complete VPR build and run, directly optimizing a composite score measured as a weighted function of critical-path delay (CPD), routed wirelength (WL), and tool runtime (RT). Across five VTR-9 benchmark circuits, VPR-Evolve improves the composite score by up to 2.7% over stock VPR in VTR-9. Relative to stock VPR, it reduces CPD by up to 9.8%, routed WL by up to 18.1%, and tool RT by up to 79.3%. VPR-Evolve reduces CPD by up to 6.0%, routed WL by up to 2.2%, and tool RT by up to 7.8% compared with a hyperparameter-tuning baseline.
作者Harrison J. Goldwyn, Graham Johnson, Christopher Ibarra, Lace Padilla, Kenny Gruchalla
Empirical user studies are essential for evaluating visual encodings and can reveal perceptual and cognitive mechanisms, but they do not by themselves provide causal, predictive accounts of interpretation errors. Evaluations are therefore often post hoc: they measure performance after a design has been specified rather than predicting how attention, uncertainty, memory, and bias may produce accurate or erroneous judgments. To address this mechanistic gap, we translate a cognitive theory of visualization interpretation into executable simulation using Active Inference, a probabilistic framework for perception, learning, and action. We model chart reading as dynamic visual search in which agents update beliefs and choose actions that balance uncertainty reduction against cognitive effort. As a proof of concept, we implement Fast, heuristic (Type 1) and Slow, analytic (Type 2) agents for a bar-chart average-estimation task. The Fast agent is vulnerable to tick-salience bias, whereas the Slow agent is more vulnerable to working-memory decay. Both produce inspectable cognitive traces, including evolving belief uncertainty and fixation sequences. By expressing these hypothesized failure mechanisms as interpretable parameters, the architecture provides a framework for formalizing and testing mechanistic hypotheses about visualization interpretation. Empirical studies can then parameterize, refine, or falsify these simulations, supporting earlier and more predictive in silico evaluation of visualization efficacy.
展开完整摘要收起摘要↓
Empirical user studies are essential for evaluating visual encodings and can reveal perceptual and cognitive mechanisms, but they do not by themselves provide causal, predictive accounts of interpretation errors. Evaluations are therefore often post hoc: they measure performance after a design has been specified rather than predicting how attention, uncertainty, memory, and bias may produce accurate or erroneous judgments. To address this mechanistic gap, we translate a cognitive theory of visualization interpretation into executable simulation using Active Inference, a probabilistic framework for perception, learning, and action. We model chart reading as dynamic visual search in which agents update beliefs and choose actions that balance uncertainty reduction against cognitive effort. As a proof of concept, we implement Fast, heuristic (Type 1) and Slow, analytic (Type 2) agents for a bar-chart average-estimation task. The Fast agent is vulnerable to tick-salience bias, whereas the Slow agent is more vulnerable to working-memory decay. Both produce inspectable cognitive traces, including evolving belief uncertainty and fixation sequences. By expressing these hypothesized failure mechanisms as interpretable parameters, the architecture provides a framework for formalizing and testing mechanistic hypotheses about visualization interpretation. Empirical studies can then parameterize, refine, or falsify these simulations, supporting earlier and more predictive in silico evaluation of visualization efficacy.
This paper studies the behavior of language models in a multi-agent crowd simulation, focusing on how affect propagates among agents that perceive and appraise one another. Each agent perceives its neighbors through visual, auditory, and tactile channels, then appraises these perceptions in light of its prompted personality profile, memory, current affective state, and situational context. Appraisal is carried out by an LLM, which updates the agent's internal affective state and selects its outward expression. The architecture contains no hand-authored mechanism for directly transferring affective state between agents; instead, inter-agent influence arises through the perception-appraisal-expression loop. The agent representation draws on the Big Five personality model and Russell's circumplex model of affect. To limit latency, low-level steering and navigation are handled by a conventional crowd simulator operating independently of the LLM-based cognitive layer. We evaluate the architecture across five scenario environments spanning alarming, joyful, and neutral situations in different spatial layouts. The results show that the system produces emotional contagion dynamics with spatial, temporal, and personality-dependent structure in sparse, small crowds. Alarm spreads from seeded agents as a traveling front, the mean alarmed fraction settles at a nonzero plateau, and the distribution of prompted personality profiles determines whether an ambiguous alarm ignites panic and whether a provocation is interpreted as anger or fear. We further evaluate the appraisal step through controlled experiments across prompt variants, sampling temperatures, and four model backends, showing that the dynamics are backend-dependent.
展开完整摘要收起摘要↓
This paper studies the behavior of language models in a multi-agent crowd simulation, focusing on how affect propagates among agents that perceive and appraise one another. Each agent perceives its neighbors through visual, auditory, and tactile channels, then appraises these perceptions in light of its prompted personality profile, memory, current affective state, and situational context. Appraisal is carried out by an LLM, which updates the agent's internal affective state and selects its outward expression. The architecture contains no hand-authored mechanism for directly transferring affective state between agents; instead, inter-agent influence arises through the perception-appraisal-expression loop. The agent representation draws on the Big Five personality model and Russell's circumplex model of affect. To limit latency, low-level steering and navigation are handled by a conventional crowd simulator operating independently of the LLM-based cognitive layer. We evaluate the architecture across five scenario environments spanning alarming, joyful, and neutral situations in different spatial layouts. The results show that the system produces emotional contagion dynamics with spatial, temporal, and personality-dependent structure in sparse, small crowds. Alarm spreads from seeded agents as a traveling front, the mean alarmed fraction settles at a nonzero plateau, and the distribution of prompted personality profiles determines whether an ambiguous alarm ignites panic and whether a provocation is interpreted as anger or fear. We further evaluate the appraisal step through controlled experiments across prompt variants, sampling temperatures, and four model backends, showing that the dynamics are backend-dependent.
Large Language Models (LLMs) are increasingly used to generate executable software environments from repository artifacts. However, functional executability does not necessarily imply conformity with architectural, security, workflow, and production intent. This study investigates whether a frontier LLM can generate Dockerfiles and Docker Compose configurations for multi-service applications using repository contents without access to developer-authored deployment artifacts. Three heterogeneous repositories combining Python, Node.js, .NET, React, Rust, Java, Redis, PostgreSQL, and MySQL-compatible infrastructure were evaluated using deterministic end-to-end HTTP oracles and manual structural comparison. All three generated environments became functionally operational, although one required a Rust base-image update from version 1.85 to 1.88. The model correctly reconstructed service topology, application ports, infrastructure dependencies, service hostnames, a background worker, hidden proxy configurations, and a file-based Docker secrets mechanism. However, it consistently omitted network segmentation, multi-stage builds, dependency-layer caching, live-reload volumes, production frontend serving, restrictive backend-port policies, and cross-platform build logic. Based on these observations, the study formalizes the distinction between functional correctness and deployment-intent fidelity and analytically derives a minimal explicit deployment specification for information that cannot be reliably inferred from repository artifacts.
展开完整摘要收起摘要↓
Large Language Models (LLMs) are increasingly used to generate executable software environments from repository artifacts. However, functional executability does not necessarily imply conformity with architectural, security, workflow, and production intent. This study investigates whether a frontier LLM can generate Dockerfiles and Docker Compose configurations for multi-service applications using repository contents without access to developer-authored deployment artifacts. Three heterogeneous repositories combining Python, Node.js, .NET, React, Rust, Java, Redis, PostgreSQL, and MySQL-compatible infrastructure were evaluated using deterministic end-to-end HTTP oracles and manual structural comparison. All three generated environments became functionally operational, although one required a Rust base-image update from version 1.85 to 1.88. The model correctly reconstructed service topology, application ports, infrastructure dependencies, service hostnames, a background worker, hidden proxy configurations, and a file-based Docker secrets mechanism. However, it consistently omitted network segmentation, multi-stage builds, dependency-layer caching, live-reload volumes, production frontend serving, restrictive backend-port policies, and cross-platform build logic. Based on these observations, the study formalizes the distinction between functional correctness and deployment-intent fidelity and analytically derives a minimal explicit deployment specification for information that cannot be reliably inferred from repository artifacts.
作者Thanh Cong Le, Michal Szczepanski, Martyna Poreba
Vision Transformer detectors now approach the accuracy of CNNs but remain difficult to deploy on NPUs and microcontrollers because key components, including deformable attention, feature fusion, and nonlinear activation functions, are not natively compatible with integer arithmetic. Existing quantized detectors either retain operators such as Softmax, GELU, and LayerNorm or focus on heavyweight backbones, leaving lightweight detection transformers without an end-to-end integer implementation. We address this gap with I-LW-DETR, the first fully integer-only lightweight DETR, in which every operation in the forward pass, including transformer nonlinearities, is executed in integer arithmetic. I-LW-DETR is built upon three key components: a scale-preserving split convolution that assigns independent activation scale to each branch of the multi-scale projector; SD-ShiftGELU, a sign-dependent GELU approximation that preserves element-wise behavior while avoiding the accuracy degradation; and a constrained Shiftmax that maintains stable Softmax normalization. Experimental results demonstrate that the proposed quantization pipeline consistently produces efficient fully integer-only models across different model scales. Across all model scales, the proposed pipeline incurs only a moderate accuracy degradation while reducing the model size by approximately $3.6\times$ and the computational cost by more than one order of magnitude.
展开完整摘要收起摘要↓
Vision Transformer detectors now approach the accuracy of CNNs but remain difficult to deploy on NPUs and microcontrollers because key components, including deformable attention, feature fusion, and nonlinear activation functions, are not natively compatible with integer arithmetic. Existing quantized detectors either retain operators such as Softmax, GELU, and LayerNorm or focus on heavyweight backbones, leaving lightweight detection transformers without an end-to-end integer implementation. We address this gap with I-LW-DETR, the first fully integer-only lightweight DETR, in which every operation in the forward pass, including transformer nonlinearities, is executed in integer arithmetic. I-LW-DETR is built upon three key components: a scale-preserving split convolution that assigns independent activation scale to each branch of the multi-scale projector; SD-ShiftGELU, a sign-dependent GELU approximation that preserves element-wise behavior while avoiding the accuracy degradation; and a constrained Shiftmax that maintains stable Softmax normalization. Experimental results demonstrate that the proposed quantization pipeline consistently produces efficient fully integer-only models across different model scales. Across all model scales, the proposed pipeline incurs only a moderate accuracy degradation while reducing the model size by approximately $3.6\times$ and the computational cost by more than one order of magnitude.
Early ranking stages in recommendation systems precompute item embeddings and cache them in-model for scoring within strict latency constraints. Because this cache exists only at serving time, outside the training loop, training and serving use different item representations, a structural discrepancy that limits quality and adds operational fragility. We show that co-designing the training and serving paths removes this representation discrepancy at its source. We introduce the memory layer, an in-model key-value embedding cache co-trained with the model: the item tower writes embeddings during training and the model reads them at serving, one source of truth for item representations by construction. Always-on embeddings cover items not yet cached, so every item receives a prediction, and the design consolidates three separate trainer-to-predictor update paths into a single self-contained pipeline. Deployed in production on Instagram Reels, the memory layer raises prediction coverage from 96% to 100%, improves embedding freshness from $O(5\text{ min})$ to $O(20\text{ s})$, and narrows the training-serving Normalized Entropy (NE) gap by up to 86%, yielding over $2\times$ recall for the freshest content and a 5-6% cold start engagement lift. Because embeddings are produced during training, the system needs no separate bulk-evaluation or publish-time recomputation, cutting training-and-publish computational cost by 30% at neutral serving computational cost.
展开完整摘要收起摘要↓
Early ranking stages in recommendation systems precompute item embeddings and cache them in-model for scoring within strict latency constraints. Because this cache exists only at serving time, outside the training loop, training and serving use different item representations, a structural discrepancy that limits quality and adds operational fragility. We show that co-designing the training and serving paths removes this representation discrepancy at its source. We introduce the memory layer, an in-model key-value embedding cache co-trained with the model: the item tower writes embeddings during training and the model reads them at serving, one source of truth for item representations by construction. Always-on embeddings cover items not yet cached, so every item receives a prediction, and the design consolidates three separate trainer-to-predictor update paths into a single self-contained pipeline. Deployed in production on Instagram Reels, the memory layer raises prediction coverage from 96% to 100%, improves embedding freshness from $O(5\text{ min})$ to $O(20\text{ s})$, and narrows the training-serving Normalized Entropy (NE) gap by up to 86%, yielding over $2\times$ recall for the freshest content and a 5-6% cold start engagement lift. Because embeddings are produced during training, the system needs no separate bulk-evaluation or publish-time recomputation, cutting training-and-publish computational cost by 30% at neutral serving computational cost.
作者Barbara Kitchenham, Sebastián Pizard, Lech Madeyski, Ronnie de Souza Santos, Martin Shepperd, David Budgen
Context: Generative AI (GenAI) and Large Language Models (LLMs) are increasingly used for academic tasks in software engineering and beyond, including systematic literature reviews (SLRs). However, while capable of summarizing text, there is no guarantee they can meet the rigour, reliability, and transparency that SLRs require. Objectives: To support researchers intending to conduct SLRs using GenAI or those conducting empirical studies evaluating how well GenAI supports SLR tasks. Methods: First, we conducted a rapid review to identify studies that propose guidelines for evaluating and using GenAI and LLMs to support SLRs. Second, we drew on thought experiments, relevant guidance from the literature, and our own experience conducting SLRs and evaluating tools to develop recommendations for how to use and assess GenAI in the context of SLRs. Results: We discuss the problems researchers face when evaluating GenAI for SLRs. We identify and explain process issues to consider when planning, conducting, and reporting both SLRs using GenAI and evaluations of GenAI tools. Finally, we summarize our results as a set of process recommendations, which we name GUEST (GenAI Use and Evaluation in SLR Tasks). Conclusion: We argue that GenAI requires human oversight and is not currently capable of unsupervised systematic studies. However, it offers the prospect of cost-effective assistance for some repetitive tasks and for additional validation of some complex tasks. Our GUEST recommendations should help software engineering researchers both to conduct and report trustworthy SLRs using GenAI and to provide rigorous independent evaluation studies.
展开完整摘要收起摘要↓
Context: Generative AI (GenAI) and Large Language Models (LLMs) are increasingly used for academic tasks in software engineering and beyond, including systematic literature reviews (SLRs). However, while capable of summarizing text, there is no guarantee they can meet the rigour, reliability, and transparency that SLRs require. Objectives: To support researchers intending to conduct SLRs using GenAI or those conducting empirical studies evaluating how well GenAI supports SLR tasks. Methods: First, we conducted a rapid review to identify studies that propose guidelines for evaluating and using GenAI and LLMs to support SLRs. Second, we drew on thought experiments, relevant guidance from the literature, and our own experience conducting SLRs and evaluating tools to develop recommendations for how to use and assess GenAI in the context of SLRs. Results: We discuss the problems researchers face when evaluating GenAI for SLRs. We identify and explain process issues to consider when planning, conducting, and reporting both SLRs using GenAI and evaluations of GenAI tools. Finally, we summarize our results as a set of process recommendations, which we name GUEST (GenAI Use and Evaluation in SLR Tasks). Conclusion: We argue that GenAI requires human oversight and is not currently capable of unsupervised systematic studies. However, it offers the prospect of cost-effective assistance for some repetitive tasks and for additional validation of some complex tasks. Our GUEST recommendations should help software engineering researchers both to conduct and report trustworthy SLRs using GenAI and to provide rigorous independent evaluation studies.
Despite rapid progress in Multi-modal Large Language Models (MLLMs), understanding long-form videos is still bottlenecked by limited context windows. While recent keyframe sampling methods attempt to mitigate this by distilling video inputs into a compact set of query-relevant frames, navigating the vast spatio-temporal search space remains challenging, as spatial detail and temporal coverage often conflict. To address this, we introduce LENS, a training-free keyframe sampling framework that dynamically decides when to zoom in for fine-grained details and when to zoom out for broader context based on the text query. Concretely, LENS adaptively allocates a limited frame budget between spatial zoom-ins, which highlight query-relevant regions within individual frames, and temporal zoom-outs, which expand the temporal scope through multi-frame aggregation, enabling the model to reason across multiple granularities while capturing both high-fidelity details and long-range context. Across diverse long-form video benchmarks, LENS consistently outperforms prior state-of-the-art keyframe sampling methods and delivers substantial gains over uniform sampling, improving Video-MME accuracy from 53.3% to 60.7% with Qwen2.5-VL.Code is available at https://github.com/zhangce01/LENS.
展开完整摘要收起摘要↓
Despite rapid progress in Multi-modal Large Language Models (MLLMs), understanding long-form videos is still bottlenecked by limited context windows. While recent keyframe sampling methods attempt to mitigate this by distilling video inputs into a compact set of query-relevant frames, navigating the vast spatio-temporal search space remains challenging, as spatial detail and temporal coverage often conflict. To address this, we introduce LENS, a training-free keyframe sampling framework that dynamically decides when to zoom in for fine-grained details and when to zoom out for broader context based on the text query. Concretely, LENS adaptively allocates a limited frame budget between spatial zoom-ins, which highlight query-relevant regions within individual frames, and temporal zoom-outs, which expand the temporal scope through multi-frame aggregation, enabling the model to reason across multiple granularities while capturing both high-fidelity details and long-range context. Across diverse long-form video benchmarks, LENS consistently outperforms prior state-of-the-art keyframe sampling methods and delivers substantial gains over uniform sampling, improving Video-MME accuracy from 53.3% to 60.7% with Qwen2.5-VL.Code is available at https://github.com/zhangce01/LENS.
Multimodal large language models (MLLMs) can connect visualization patterns to external causes, consequences, and domain knowledge, but the evidential basis of these interpretations is often unclear. We present an exploratory study of 102 visualizations from four sources, three MLLMs, and four input conditions that vary access to the image, source-specific accessible chart context, and withheld-context framing. Across 1,224 descriptions, we analyze model-attributed DIRECT, DERIVED, and SPECULATIVE labels and conduct an automated audit of numeric agreement. Accessible chart context shifted Gemini and GPT toward DIRECT claims and improved numeric agreement for some models. Adding the image to the full context did not yield a consistent numeric benefit, and the withheld-context prompt did not reliably increase cautious language. The prompt-defined Real-World Significance section remained predominantly SPECULATIVE. These results motivate accessible description systems that distinguish claims supported by supplied evidence from model-supplied interpretation
展开完整摘要收起摘要↓
Multimodal large language models (MLLMs) can connect visualization patterns to external causes, consequences, and domain knowledge, but the evidential basis of these interpretations is often unclear. We present an exploratory study of 102 visualizations from four sources, three MLLMs, and four input conditions that vary access to the image, source-specific accessible chart context, and withheld-context framing. Across 1,224 descriptions, we analyze model-attributed DIRECT, DERIVED, and SPECULATIVE labels and conduct an automated audit of numeric agreement. Accessible chart context shifted Gemini and GPT toward DIRECT claims and improved numeric agreement for some models. Adding the image to the full context did not yield a consistent numeric benefit, and the withheld-context prompt did not reliably increase cautious language. The prompt-defined Real-World Significance section remained predominantly SPECULATIVE. These results motivate accessible description systems that distinguish claims supported by supplied evidence from model-supplied interpretation
Urban decision-making requires integrating heterogeneous spatial data. While current GIS tools handle geometric computation efficiently, they lack the semantic reasoning to guide complex workflows. Analysts manually manage data discovery, spatial boundaries, and measurement semantics, risking aggregation errors. We present UrbanTrace, a visual analytics system that transforms manual spatial data-wrangling into a transparent, node-based collaborative workflow with context-aware AI agents. Using an offline profiler to extract semantic and geometric metadata, UrbanTrace grounds LLMs in real-world data distributions. This enables specialized agents to retrieve datasets based on high-level goals and automatically enforce valid spatial aggregations. To make harmonization explicit, three interactive views: an Integration Provenance Graph, Multivariate Priority Map, and Spatial Delta Map, allow users to explore how conclusions shift across spatial configurations. We evaluate UrbanTrace on 28 urban scenarios spanning 112 datasets. Quantitative ablations show our profiling significantly outperforms baseline LLMs in data discovery, achieving 100% semantic and 87% geometric validity in spatial mapping. Through real-world case studies and expert interviews, we demonstrate that UrbanTrace turns spatial aggregation sensitivity from a methodological burden into an exploratory visual asset.
展开完整摘要收起摘要↓
Urban decision-making requires integrating heterogeneous spatial data. While current GIS tools handle geometric computation efficiently, they lack the semantic reasoning to guide complex workflows. Analysts manually manage data discovery, spatial boundaries, and measurement semantics, risking aggregation errors. We present UrbanTrace, a visual analytics system that transforms manual spatial data-wrangling into a transparent, node-based collaborative workflow with context-aware AI agents. Using an offline profiler to extract semantic and geometric metadata, UrbanTrace grounds LLMs in real-world data distributions. This enables specialized agents to retrieve datasets based on high-level goals and automatically enforce valid spatial aggregations. To make harmonization explicit, three interactive views: an Integration Provenance Graph, Multivariate Priority Map, and Spatial Delta Map, allow users to explore how conclusions shift across spatial configurations. We evaluate UrbanTrace on 28 urban scenarios spanning 112 datasets. Quantitative ablations show our profiling significantly outperforms baseline LLMs in data discovery, achieving 100% semantic and 87% geometric validity in spatial mapping. Through real-world case studies and expert interviews, we demonstrate that UrbanTrace turns spatial aggregation sensitivity from a methodological burden into an exploratory visual asset.
Automated verification of numerical claims is a challenging problem, as it requires both language understanding and quantitative reasoning. This paper describes our system for CLEF 2026 CheckThat! Task 2, which focuses on ranking reasoning traces generated by large language models (LLMs) and predicting a final verdict for numerical claims in English and Arabic. We explore two approaches. The first approach fine-tunes an LLM-based verifier using LoRA to score each reasoning trace independently as a binary classification problem, and selects the final verdict using Best-of-N selection. We further experiment with adaptive sub-claim decomposition to break complex claims into simpler parts before verification. The second approach uses a lightweight TF-IDF reward model with handcrafted numeric and temporal overlap features to score traces, and aggregates scores by verdict group to determine the final prediction. For Arabic, we compare a general multilingual model against AraBERT, a language-specific model pretrained on Arabic text. Our results show that the LLM-based approach outperforms the lightweight reward model on most metrics, particularly Recall@5, while the reward-based approach shows stronger performance on the Conflicting class. Sub-claim decomposition did not improve performance, suggesting that claim splitting introduces noise rather than aiding reasoning. For Arabic, AraBERT outperforms the multilingual baseline across most metrics.
展开完整摘要收起摘要↓
Automated verification of numerical claims is a challenging problem, as it requires both language understanding and quantitative reasoning. This paper describes our system for CLEF 2026 CheckThat! Task 2, which focuses on ranking reasoning traces generated by large language models (LLMs) and predicting a final verdict for numerical claims in English and Arabic. We explore two approaches. The first approach fine-tunes an LLM-based verifier using LoRA to score each reasoning trace independently as a binary classification problem, and selects the final verdict using Best-of-N selection. We further experiment with adaptive sub-claim decomposition to break complex claims into simpler parts before verification. The second approach uses a lightweight TF-IDF reward model with handcrafted numeric and temporal overlap features to score traces, and aggregates scores by verdict group to determine the final prediction. For Arabic, we compare a general multilingual model against AraBERT, a language-specific model pretrained on Arabic text. Our results show that the LLM-based approach outperforms the lightweight reward model on most metrics, particularly Recall@5, while the reward-based approach shows stronger performance on the Conflicting class. Sub-claim decomposition did not improve performance, suggesting that claim splitting introduces noise rather than aiding reasoning. For Arabic, AraBERT outperforms the multilingual baseline across most metrics.
This paper studies what discriminatively trained reward models (RMs) memorize by measuring counterfactual memorization on two human preference datasets. We show that RMs 1) misallocate memorization to easy, high margin preference pairs, 2) memorize dataset-specific shortcuts (e.g., model identity, user sampling strategy), and 3) overgeneralize simple heuristic correlates of human preference (e.g., length, compliance) when confronted with unseen preference pairs. Overall, our findings indicate that discriminative training of RMs from human preference data results in biased RMs not yet capable of judging response quality in context-dependent scenarios.
展开完整摘要收起摘要↓
This paper studies what discriminatively trained reward models (RMs) memorize by measuring counterfactual memorization on two human preference datasets. We show that RMs 1) misallocate memorization to easy, high margin preference pairs, 2) memorize dataset-specific shortcuts (e.g., model identity, user sampling strategy), and 3) overgeneralize simple heuristic correlates of human preference (e.g., length, compliance) when confronted with unseen preference pairs. Overall, our findings indicate that discriminative training of RMs from human preference data results in biased RMs not yet capable of judging response quality in context-dependent scenarios.
Large language models increasingly solve scientific-computing tasks, but executable feedback from one problem rarely becomes durable capability on subsequent problems. We study scientific-computing experience consolidation: converting verified runtime experience into transferable procedural knowledge and persistent model improvement. This setting presents two challenges: trajectory-derived artifacts may encode source-specific repairs rather than cross-task computational mechanisms; and a weaker target model may be unable to operationalize an otherwise valid abstract procedure - an abstraction-execution gap. We introduce SciConsolidate, which contrasts verified successes and failures to induce cross-task procedures, selects them through a development-validation gate, and uses failure-informed, answer-free query synthesis to expand the consolidation data without requiring pre-existing reference answers. Because the target model may not directly execute these abstractions, a stronger model concretizes them into executable code supervision for standard, procedure-free SFT; a matched no-procedure teacher branch isolates the value of procedural guidance. On SciCode, runtime procedure injection improves Qwen3.6-27B by +3.85/+6.26 sub-step/main-problem points, but yields almost no aggregate main-problem gain for Qwen3.5-9B, providing operational evidence of the abstraction-execution gap. After procedure-guided concretization, the 9B student improves under procedure-free deployment by +3.89/+6.25 points over the no-procedure SFT control and by +5.62/+11.25 over the original 9B model. These results establish an experience-to-capability pathway for scientific computing and provide a practical starting point for scaling self-improving scientific assistance.
展开完整摘要收起摘要↓
Large language models increasingly solve scientific-computing tasks, but executable feedback from one problem rarely becomes durable capability on subsequent problems. We study scientific-computing experience consolidation: converting verified runtime experience into transferable procedural knowledge and persistent model improvement. This setting presents two challenges: trajectory-derived artifacts may encode source-specific repairs rather than cross-task computational mechanisms; and a weaker target model may be unable to operationalize an otherwise valid abstract procedure - an abstraction-execution gap. We introduce SciConsolidate, which contrasts verified successes and failures to induce cross-task procedures, selects them through a development-validation gate, and uses failure-informed, answer-free query synthesis to expand the consolidation data without requiring pre-existing reference answers. Because the target model may not directly execute these abstractions, a stronger model concretizes them into executable code supervision for standard, procedure-free SFT; a matched no-procedure teacher branch isolates the value of procedural guidance. On SciCode, runtime procedure injection improves Qwen3.6-27B by +3.85/+6.26 sub-step/main-problem points, but yields almost no aggregate main-problem gain for Qwen3.5-9B, providing operational evidence of the abstraction-execution gap. After procedure-guided concretization, the 9B student improves under procedure-free deployment by +3.89/+6.25 points over the no-procedure SFT control and by +5.62/+11.25 over the original 9B model. These results establish an experience-to-capability pathway for scientific computing and provide a practical starting point for scaling self-improving scientific assistance.