作者Zhongjie Shi, Rongjie Lai, Alexander Cloninger, Wenjing Liao
Transformers achieve remarkable performance by jointly learning broad families of tasks during pretraining and adapting to unseen tasks from only a short prompt. Yet a rigorous mathematical and statistical understanding of this phenomenon remains limited. This paper aims to study how Transformers exploit shared cross-task structure and how this structure affects the sample complexity of in-context learning (ICL). Specifically, we characterize task-space complexity through covering numbers under a prescribed metric, thereby quantifying the low-dimensional cross-task structure without requiring an explicit parametric representation. The resulting cover provides a set of anchor functions, which we use to introduce a task-identification-and-evaluation procedure: context observations localize an unseen task among the anchor functions, and the response at a query is predicted by aggregating the corresponding anchor function query evaluations. For approximation, we explicitly construct a Transformer with Softmax attention to approximate this procedure. For generalization, we derive an error bound that separates the effects of the number of pretraining tasks and the prompt length. The scaling with respect to the number of pretraining tasks is governed by the intrinsic dimensions of the task space and input domain; once sufficiently many tasks are available, the dependence on the prompt context length becomes dimension-free. To the best of our knowledge, this is the first work to quantify cross-task complexity for general nonlinear task families and explicitly construct a Transformer that exploits their low-dimensional structure to perform ICL. Our theory provides a quantitative explanation of how joint pretraining across related tasks improves in-context generalization.
展开完整摘要收起摘要↓
Transformers achieve remarkable performance by jointly learning broad families of tasks during pretraining and adapting to unseen tasks from only a short prompt. Yet a rigorous mathematical and statistical understanding of this phenomenon remains limited. This paper aims to study how Transformers exploit shared cross-task structure and how this structure affects the sample complexity of in-context learning (ICL). Specifically, we characterize task-space complexity through covering numbers under a prescribed metric, thereby quantifying the low-dimensional cross-task structure without requiring an explicit parametric representation. The resulting cover provides a set of anchor functions, which we use to introduce a task-identification-and-evaluation procedure: context observations localize an unseen task among the anchor functions, and the response at a query is predicted by aggregating the corresponding anchor function query evaluations. For approximation, we explicitly construct a Transformer with Softmax attention to approximate this procedure. For generalization, we derive an error bound that separates the effects of the number of pretraining tasks and the prompt length. The scaling with respect to the number of pretraining tasks is governed by the intrinsic dimensions of the task space and input domain; once sufficiently many tasks are available, the dependence on the prompt context length becomes dimension-free. To the best of our knowledge, this is the first work to quantify cross-task complexity for general nonlinear task families and explicitly construct a Transformer that exploits their low-dimensional structure to perform ICL. Our theory provides a quantitative explanation of how joint pretraining across related tasks improves in-context generalization.
作者Zeyu Michael Li, William Xingxu Chen, Bingshuo Qian, Jiayin Liu, Xiang Cheng
Fully continuous diffusion language models (dLMs) denoise continuous representations without intermediate discretization, then decode all response tokens in parallel at the final step. Their performance on challenging reasoning tasks remains less established than that of autoregressive (AR) LLMs and masked dLMs. We scale Embedded Language Flows (ELF) to mathematical reasoning and code generation on GSM8K, MATH-500, HumanEval, and MBPP. We introduce ELF-REG, which improves learning with representation alignment and entanglement (REPA+REG), where a frozen AR teacher supervises intermediate denoiser features and supplies a global representation that is jointly denoised with the response. ELF-REG-L achieves 55.96% pass@1 on GSM8K at 64 network function evaluations (NFE), and 13.39% on MATH-500 and 22.56% on HumanEval at 128 NFE. It outperforms the evaluated comparable-scale dLMs in pass@1 on GSM8K and code, and improves MATH-500 pass@1 from 10.55% for the ELF-L baseline to 13.39% with ELF-REG-L. Without few-step training, the same task-specific checkpoints support strong low-NFE performance through early-stop, which decodes an intermediate clean prediction without completing the denoising trajectory. At 16 NFE, ELF-REG-L reaches 41.21% HumanEval pass@10, outperforming recent continuous dLMs of comparable scale.
展开完整摘要收起摘要↓
Fully continuous diffusion language models (dLMs) denoise continuous representations without intermediate discretization, then decode all response tokens in parallel at the final step. Their performance on challenging reasoning tasks remains less established than that of autoregressive (AR) LLMs and masked dLMs. We scale Embedded Language Flows (ELF) to mathematical reasoning and code generation on GSM8K, MATH-500, HumanEval, and MBPP. We introduce ELF-REG, which improves learning with representation alignment and entanglement (REPA+REG), where a frozen AR teacher supervises intermediate denoiser features and supplies a global representation that is jointly denoised with the response. ELF-REG-L achieves 55.96% pass@1 on GSM8K at 64 network function evaluations (NFE), and 13.39% on MATH-500 and 22.56% on HumanEval at 128 NFE. It outperforms the evaluated comparable-scale dLMs in pass@1 on GSM8K and code, and improves MATH-500 pass@1 from 10.55% for the ELF-L baseline to 13.39% with ELF-REG-L. Without few-step training, the same task-specific checkpoints support strong low-NFE performance through early-stop, which decodes an intermediate clean prediction without completing the denoising trajectory. At 16 NFE, ELF-REG-L reaches 41.21% HumanEval pass@10, outperforming recent continuous dLMs of comparable scale.
Benchmarks for vision-language models offer their answer choices in some convention: a letter, a color name, a pixel coordinate. That convention is treated as neutral. We find it is not, and that the limits a benchmark reports can belong to the readout rather than to the model. On 200 COCO photographs, Qwen3-VL-4B picks the correct one of nine locations for a named object 68.5% of the time when the locations are given in English and 20.0% when the same locations are given as pixel coordinates. Chance is 11.1%. The cost arises when the answer options are coordinates; giving the model a coordinate in the question instead costs 3.5 points and is not significant. The gap holds on a 4x4 grid, under 8-bit rather than 4-bit quantization, and in every slice by object size, boundary distance and category. It also decides which model wins. Two models that tie under English names differ by 39 points in one coordinate system and by 54 in the other, in opposite directions. On the color task, three of the four open models capable of the task show the penalty; on photographs, two of three open models do, and so does Gemini, at 11.1 points on parseable answers (p = 1e-4). GPT-4o does not. To ask whether a model reads a coordinate at all, we attach the wrong name to each one and record which the model follows. Color options written as hue angles are followed below chance; a normalized pixel convention is followed at four times chance. This tells apart conventions a model can use from ones it cannot, though it did not predict accuracy on two untried conventions. Five models also name the same color wheel five different ways, so a fixed answer vocabulary is not neutral across models either. Five times during this work we measured a capable model as incapable because our scorer and the model disagreed about what an answer looks like. We report each case. They are the phenomenon in miniature.
展开完整摘要收起摘要↓
Benchmarks for vision-language models offer their answer choices in some convention: a letter, a color name, a pixel coordinate. That convention is treated as neutral. We find it is not, and that the limits a benchmark reports can belong to the readout rather than to the model. On 200 COCO photographs, Qwen3-VL-4B picks the correct one of nine locations for a named object 68.5% of the time when the locations are given in English and 20.0% when the same locations are given as pixel coordinates. Chance is 11.1%. The cost arises when the answer options are coordinates; giving the model a coordinate in the question instead costs 3.5 points and is not significant. The gap holds on a 4x4 grid, under 8-bit rather than 4-bit quantization, and in every slice by object size, boundary distance and category. It also decides which model wins. Two models that tie under English names differ by 39 points in one coordinate system and by 54 in the other, in opposite directions. On the color task, three of the four open models capable of the task show the penalty; on photographs, two of three open models do, and so does Gemini, at 11.1 points on parseable answers (p = 1e-4). GPT-4o does not. To ask whether a model reads a coordinate at all, we attach the wrong name to each one and record which the model follows. Color options written as hue angles are followed below chance; a normalized pixel convention is followed at four times chance. This tells apart conventions a model can use from ones it cannot, though it did not predict accuracy on two untried conventions. Five models also name the same color wheel five different ways, so a fixed answer vocabulary is not neutral across models either. Five times during this work we measured a capable model as incapable because our scorer and the model disagreed about what an answer looks like. We report each case. They are the phenomenon in miniature.
作者Beomgu Kang, SoJin Yun, Hojoon Kim, Hyunseok Seo
KV-cache eviction is typically evaluated through average quality-memory trade-offs, yet a small average loss can hide requests whose utility degrades materially. We reformulate eviction as a deployment risk-control problem: a material degradation occurs when eviction lowers task utility by more than a deployment-specified tolerance relative to full-KV inference on the same request, and deployment risk is the population frequency of such events. Given a reliability contract specifying a target risk level and confidence requirement, we use a compressor-agnostic post-hoc certification procedure to select a retention policy from calibration data with a finite-sample guarantee, falling back to full KV when no compressed policy is certified. Across multiple eviction methods, Llama and Mistral models, and LongBench and RULER-32K, the same contract supports substantially different levels of eviction: on Llama, it certifies SnapKV at 75% retention on LongBench but no tested compressed policy on RULER-32K, triggering full-KV fallback. Policies with empirical degradation rates below the 5% target can still fail finite-sample certification; on Llama LongBench, empirical thresholding selects uncertified policies that retain 5-10 percentage points less cache across fixed-budget methods. The proposed framework converts a deployment-level reliability requirement into a KV-memory operating point.
展开完整摘要收起摘要↓
KV-cache eviction is typically evaluated through average quality-memory trade-offs, yet a small average loss can hide requests whose utility degrades materially. We reformulate eviction as a deployment risk-control problem: a material degradation occurs when eviction lowers task utility by more than a deployment-specified tolerance relative to full-KV inference on the same request, and deployment risk is the population frequency of such events. Given a reliability contract specifying a target risk level and confidence requirement, we use a compressor-agnostic post-hoc certification procedure to select a retention policy from calibration data with a finite-sample guarantee, falling back to full KV when no compressed policy is certified. Across multiple eviction methods, Llama and Mistral models, and LongBench and RULER-32K, the same contract supports substantially different levels of eviction: on Llama, it certifies SnapKV at 75% retention on LongBench but no tested compressed policy on RULER-32K, triggering full-KV fallback. Policies with empirical degradation rates below the 5% target can still fail finite-sample certification; on Llama LongBench, empirical thresholding selects uncertified policies that retain 5-10 percentage points less cache across fixed-budget methods. The proposed framework converts a deployment-level reliability requirement into a KV-memory operating point.
作者Zirui Li, Rech Silas, Lauri Juvela, Tom Backstrom, Mikko Kurimo
Generating controllable and human-like emphasis remains an open challenge in text-to-speech, even when explicit emphasis control signals are provided in the text input, limiting the communicative accuracy of synthetic speech in real-world applications. Reinforcement learning has recently shown promise for post-training TTS systems to align with human preference, yet existing methods have not been applied to word-level prosodic control. We present EmphTTS, a non-autoregressive TTS system that applies Group Relative Policy Optimization (GRPO) to the duration predictor with an emphasis localization reward, enabling direct optimization for word-level emphasis. Evaluations show that EmphTTS achieves the best emphasis controllability and performs the best in emphasis objective evaluation. In subjective preference tests, EmphTTS is significantly preferred over synthetic groundtruth and most baselines. Ablation studies show that GRPO improves emphasis realization beyond supervised-finetuning-based duration modeling and simple speaking-rate adjustment, while alleviating the mismatch between the independently trained duration predictor and TTS model.
展开完整摘要收起摘要↓
Generating controllable and human-like emphasis remains an open challenge in text-to-speech, even when explicit emphasis control signals are provided in the text input, limiting the communicative accuracy of synthetic speech in real-world applications. Reinforcement learning has recently shown promise for post-training TTS systems to align with human preference, yet existing methods have not been applied to word-level prosodic control. We present EmphTTS, a non-autoregressive TTS system that applies Group Relative Policy Optimization (GRPO) to the duration predictor with an emphasis localization reward, enabling direct optimization for word-level emphasis. Evaluations show that EmphTTS achieves the best emphasis controllability and performs the best in emphasis objective evaluation. In subjective preference tests, EmphTTS is significantly preferred over synthetic groundtruth and most baselines. Ablation studies show that GRPO improves emphasis realization beyond supervised-finetuning-based duration modeling and simple speaking-rate adjustment, while alleviating the mismatch between the independently trained duration predictor and TTS model.
World Action Models (WAMs) jointly model action generation and environment dynamics and are mostly built on pretrained Video Diffusion Models (VDMs). In VDM-based WAMs, observations are first encoded by a VAE, and the resulting compressed latents are then processed by large video diffusion backbones to extract effective features for action generation. However, this paradigm ties WAM performance and training cost to large-scale video generation pretraining, limiting WAM efficiency and scalability. In this paper, we theoretically and empirically investigate how visual representations affect action generation in WAMs. Our results show that predictive embeddings from Joint-Embedding Predictive Architecture (JEPA) encoders better support action generation than compressed VAE latents, with I-JEPA performing best in our encoder comparison. Based on these findings, we propose LeWAM, which conditions action generation on JEPA embeddings and models environment evolution by predicting future embeddings in the same space, without relying on a video diffusion backbone. We further find that imitation learning matches demonstrated actions but does not distinguish better actions from worse ones, even though small action deviations can greatly affect task success. To address this limitation without additional environment interaction or the human oversight required for resets and safety, we introduce Demonstration-Guided DPO (DemoDPO), an offline preference refinement stage that derives preference supervision directly from demonstrations. With only 0.4B trainable parameters, LeWAM achieves an average success rate of 92.28% on RoboTwin 2.0, comparable to that of state-of-the-art VLAs and WAMs, and maintains practical effectiveness on real-world manipulation tasks.
展开完整摘要收起摘要↓
World Action Models (WAMs) jointly model action generation and environment dynamics and are mostly built on pretrained Video Diffusion Models (VDMs). In VDM-based WAMs, observations are first encoded by a VAE, and the resulting compressed latents are then processed by large video diffusion backbones to extract effective features for action generation. However, this paradigm ties WAM performance and training cost to large-scale video generation pretraining, limiting WAM efficiency and scalability. In this paper, we theoretically and empirically investigate how visual representations affect action generation in WAMs. Our results show that predictive embeddings from Joint-Embedding Predictive Architecture (JEPA) encoders better support action generation than compressed VAE latents, with I-JEPA performing best in our encoder comparison. Based on these findings, we propose LeWAM, which conditions action generation on JEPA embeddings and models environment evolution by predicting future embeddings in the same space, without relying on a video diffusion backbone. We further find that imitation learning matches demonstrated actions but does not distinguish better actions from worse ones, even though small action deviations can greatly affect task success. To address this limitation without additional environment interaction or the human oversight required for resets and safety, we introduce Demonstration-Guided DPO (DemoDPO), an offline preference refinement stage that derives preference supervision directly from demonstrations. With only 0.4B trainable parameters, LeWAM achieves an average success rate of 92.28% on RoboTwin 2.0, comparable to that of state-of-the-art VLAs and WAMs, and maintains practical effectiveness on real-world manipulation tasks.
Reinforcement-learning (RL) policies are often distributed as opaque neural checkpoints, while training logs show that a run occurred without explaining what the policy learned. We study whether independently trained policies can be represented and composed through auditable discrete behavioral rules. We define auditability as six separately testable predicates: trace integrity, lossless coding, rule coverage, behavioral agreement, composition quality, and value-model reliability. Our protocol uses a shared frozen symbolizer, passive rule extraction, an append-only hash-bound ledger, exact environment replay, and offline confidence-ranked arbitration with an explicit blind-spot fallback. The results place strict limits on this description layer. Rule-set overlap does not imply behavioral agreement: policies may share symbolic rules while choosing near-chance-matching actions on fresh states. The fused policy therefore selects among existing rules rather than generating a new skill. On a conflict-dominated task, an apparent fusion failure is traced to an induction/deployment mismatch: rules induced from sampled actions were evaluated under argmax actions, and deployment-consistent re-induction reverses the arbitration ordering. A fitted-Q generalized-policy-improvement diagnostic also fails in both environments, limiting claims that rule fusion is superior to value-based composition. One exploratory comparison favors rule fusion, but its comparator is post hoc, the task is partly saturated, and the fused policy remains below the strongest held-out actor. We contribute an evidence-bounded audit and composition protocol, not a claim of universal interpretability or autonomous skill generation. Future work must add temporally extended skills, cross-skill interfaces, composition search, and independent novelty audits.
展开完整摘要收起摘要↓
Reinforcement-learning (RL) policies are often distributed as opaque neural checkpoints, while training logs show that a run occurred without explaining what the policy learned. We study whether independently trained policies can be represented and composed through auditable discrete behavioral rules. We define auditability as six separately testable predicates: trace integrity, lossless coding, rule coverage, behavioral agreement, composition quality, and value-model reliability. Our protocol uses a shared frozen symbolizer, passive rule extraction, an append-only hash-bound ledger, exact environment replay, and offline confidence-ranked arbitration with an explicit blind-spot fallback. The results place strict limits on this description layer. Rule-set overlap does not imply behavioral agreement: policies may share symbolic rules while choosing near-chance-matching actions on fresh states. The fused policy therefore selects among existing rules rather than generating a new skill. On a conflict-dominated task, an apparent fusion failure is traced to an induction/deployment mismatch: rules induced from sampled actions were evaluated under argmax actions, and deployment-consistent re-induction reverses the arbitration ordering. A fitted-Q generalized-policy-improvement diagnostic also fails in both environments, limiting claims that rule fusion is superior to value-based composition. One exploratory comparison favors rule fusion, but its comparator is post hoc, the task is partly saturated, and the fused policy remains below the strongest held-out actor. We contribute an evidence-bounded audit and composition protocol, not a claim of universal interpretability or autonomous skill generation. Future work must add temporally extended skills, cross-skill interfaces, composition search, and independent novelty audits.
作者Minqiu Sun, Xin Huang, Luanzheng Guo, Nathan R. Tallent, Kento Sato, Dong Dai
Zeroth-order (ZO) optimization is an attractive option for memory-efficient LLM fine-tuning, but its fault tolerance remains underexplored. Unlike first-order training, ZO progress can be represented by lightweight seed-and-scalar step logs, yet naive log-only recovery still incurs replay cost that grows with training progress, and shortcut replay does not preserve the executed floating-point trajectory. We present ZOCheck, a fault-tolerant ZO training system that exploits this replayable structure through a CPU shadow process that continuously replays logged updates, materializes consistent recovery images off the GPU critical path, and persists them asynchronously. ZOCheck therefore combines non-blocking checkpointing during training with fast recovery from a near-current state. We also develop a cost model for choosing the snapshot policy under realistic failure rates. Experiments show that ZOCheck reduces checkpoint overhead by up to 219.7x and recovery latency by 1.55x on average compared with asynchronous full-state checkpointing, translating into up to 21.3x lower end-to-end wasted time across the evaluated failure rates, while preserving exact recovery behavior.
展开完整摘要收起摘要↓
Zeroth-order (ZO) optimization is an attractive option for memory-efficient LLM fine-tuning, but its fault tolerance remains underexplored. Unlike first-order training, ZO progress can be represented by lightweight seed-and-scalar step logs, yet naive log-only recovery still incurs replay cost that grows with training progress, and shortcut replay does not preserve the executed floating-point trajectory. We present ZOCheck, a fault-tolerant ZO training system that exploits this replayable structure through a CPU shadow process that continuously replays logged updates, materializes consistent recovery images off the GPU critical path, and persists them asynchronously. ZOCheck therefore combines non-blocking checkpointing during training with fast recovery from a near-current state. We also develop a cost model for choosing the snapshot policy under realistic failure rates. Experiments show that ZOCheck reduces checkpoint overhead by up to 219.7x and recovery latency by 1.55x on average compared with asynchronous full-state checkpointing, translating into up to 21.3x lower end-to-end wasted time across the evaluated failure rates, while preserving exact recovery behavior.
Recent one-step text-to-audio (TTA) models substantially reduce inference cost, yet their generated distributions can still be improved through post-training. We propose DriftAudio, a distributional post-training method that adapts Drifting to pretrained one-step TTA generators. Applying Drifting condition-wise is challenging under free-form text conditioning, where only one or a few real samples are typically available for a particular condition. DriftAudio instead performs Drifting on the marginal audio distribution while retaining text-conditioned generation. The drifting field is estimated in a frozen audio feature space using resampled real samples, a rolling bank of generated samples, and the current batch of generated samples. The resulting Drifting field provides a detached training target for updating only the generator, keeping the original one-step inference procedure. On AudioCaps, starting from MeanAudio, DriftAudio reduces FAD and FD by 33.9% and 17.6%, respectively, while also improving KL and CLAP. Starting from FdAudio, it further reduces FAD, FD, and KL, with trade-offs in IS and CLAP. These results demonstrate the effectiveness of marginal distributional post-training for one-step TTA generation.
展开完整摘要收起摘要↓
Recent one-step text-to-audio (TTA) models substantially reduce inference cost, yet their generated distributions can still be improved through post-training. We propose DriftAudio, a distributional post-training method that adapts Drifting to pretrained one-step TTA generators. Applying Drifting condition-wise is challenging under free-form text conditioning, where only one or a few real samples are typically available for a particular condition. DriftAudio instead performs Drifting on the marginal audio distribution while retaining text-conditioned generation. The drifting field is estimated in a frozen audio feature space using resampled real samples, a rolling bank of generated samples, and the current batch of generated samples. The resulting Drifting field provides a detached training target for updating only the generator, keeping the original one-step inference procedure. On AudioCaps, starting from MeanAudio, DriftAudio reduces FAD and FD by 33.9% and 17.6%, respectively, while also improving KL and CLAP. Starting from FdAudio, it further reduces FAD, FD, and KL, with trade-offs in IS and CLAP. These results demonstrate the effectiveness of marginal distributional post-training for one-step TTA generation.
Recent zero-shot text-to-speech (ZS-TTS) systems can reproduce a speaker's voice with high fidelity from only a few seconds of reference speech, raising concerns over unauthorized voice cloning and impersonation. Speaker identity unlearning has recently emerged as an approach to selectively suppress this capability for speakers who opt out while preserving synthesis capability for other speakers. Although existing approaches reduce speaker similarity, preventing re-identification often faces severe degradation of speech quality. Motivated by this observation, we propose GUARD, a lightweight speaker identity unlearning framework that combines a learned speaker gate with speaker-agnostic activation steering on a frozen TTS backbone. The steering vectors are optimized using group-relative reward optimization to shift outputs from forget speakers toward population-level impostor similarity while preserving intelligibility and speech naturalness. On CosyVoice2, GUARD reduces forget-speaker similarity from 0.541 to 0.103 and re-identification accuracy in a 150-speaker gallery from 73.5% to 0.5%, while preserving retain-speaker reproduction. The results demonstrate that similarity reduction alone may not fully characterize successful speaker identity unlearning and highlight re-identification as a complementary criterion for its evaluation.
展开完整摘要收起摘要↓
Recent zero-shot text-to-speech (ZS-TTS) systems can reproduce a speaker's voice with high fidelity from only a few seconds of reference speech, raising concerns over unauthorized voice cloning and impersonation. Speaker identity unlearning has recently emerged as an approach to selectively suppress this capability for speakers who opt out while preserving synthesis capability for other speakers. Although existing approaches reduce speaker similarity, preventing re-identification often faces severe degradation of speech quality. Motivated by this observation, we propose GUARD, a lightweight speaker identity unlearning framework that combines a learned speaker gate with speaker-agnostic activation steering on a frozen TTS backbone. The steering vectors are optimized using group-relative reward optimization to shift outputs from forget speakers toward population-level impostor similarity while preserving intelligibility and speech naturalness. On CosyVoice2, GUARD reduces forget-speaker similarity from 0.541 to 0.103 and re-identification accuracy in a 150-speaker gallery from 73.5% to 0.5%, while preserving retain-speaker reproduction. The results demonstrate that similarity reduction alone may not fully characterize successful speaker identity unlearning and highlight re-identification as a complementary criterion for its evaluation.
作者Sandra Arcos-Holzinger, Debashish Chakraborty, Rohita Mocharla, Will Walden, Andrew Yates, Reno Kriz, Sarah M. Erfani, James Bailey, Vishal M. Patel, Sanjeev Khudanpur
We propose LAYERSCOPE, a label-free, layerwise framework that aims to characterize a model's learned representations in video and multimodal settings. Evaluating downstream performance using representations from final or intermediate layers typically requires large amounts of labeled data, repeated task-specific evaluations, and substantial computation. To address these limitations, LAYERSCOPE uses local, global, distributional, and correspondence-based geometric metrics to compare layerwise representation structure within and across models without requiring task-specific labels. We evaluate seven architecturally diverse models across video and multimodal classification, clustering, and text-to-video retrieval tasks from MVEB/MVEB+. We find that intermediate-layer representations can outperform final-layer and model-default outputs. We also find that no single geometric metric consistently predicts downstream performance, but note that distinct layerwise geometric signatures emerge across model families. LID shows task-dependent relationships with performance, while RankMe provides the strongest measure for classification and clustering, but is not a universal layer selector. We also find that pairing-aware metrics explain retrieval better than distributional distances alone. LAYERSCOPE therefore offers a framework for comparing representations across models and layers, enabling a more systematic evaluation in video and multimodal settings.
展开完整摘要收起摘要↓
We propose LAYERSCOPE, a label-free, layerwise framework that aims to characterize a model's learned representations in video and multimodal settings. Evaluating downstream performance using representations from final or intermediate layers typically requires large amounts of labeled data, repeated task-specific evaluations, and substantial computation. To address these limitations, LAYERSCOPE uses local, global, distributional, and correspondence-based geometric metrics to compare layerwise representation structure within and across models without requiring task-specific labels. We evaluate seven architecturally diverse models across video and multimodal classification, clustering, and text-to-video retrieval tasks from MVEB/MVEB+. We find that intermediate-layer representations can outperform final-layer and model-default outputs. We also find that no single geometric metric consistently predicts downstream performance, but note that distinct layerwise geometric signatures emerge across model families. LID shows task-dependent relationships with performance, while RankMe provides the strongest measure for classification and clustering, but is not a universal layer selector. We also find that pairing-aware metrics explain retrieval better than distributional distances alone. LAYERSCOPE therefore offers a framework for comparing representations across models and layers, enabling a more systematic evaluation in video and multimodal settings.
Text-to-audio (TTA) generation aims to synthesize realistic audio that faithfully reflects natural-language descriptions. Most TTA systems adopt a two-stage latent paradigm: an audio tokenizer is optimized for reconstruction and then frozen, after which a generative model is trained in the resulting latent space. However, reconstruction-oriented representations may be suboptimal for generation, motivating joint representation and generative learning. To this end, we introduce Unite-Audio, to our knowledge, is the first to jointly learn continuous audio representations and latent flow matching for TTA. By coupling reconstruction with self-supervised generative prediction, Unite-Audio allows the generative objective to directly shape the latent space rather than treating it as a fixed intermediate representation. We further employ Flow-GRPO post-training to improve text-conditioned generation. Experiments show competitive TTA performance with a compact latent flow model, while ablation studies confirm the benefit of jointly learning the audio representation and generative model. Audio samples are available at https://runwushi.github.io/Unite-Audio.
展开完整摘要收起摘要↓
Text-to-audio (TTA) generation aims to synthesize realistic audio that faithfully reflects natural-language descriptions. Most TTA systems adopt a two-stage latent paradigm: an audio tokenizer is optimized for reconstruction and then frozen, after which a generative model is trained in the resulting latent space. However, reconstruction-oriented representations may be suboptimal for generation, motivating joint representation and generative learning. To this end, we introduce Unite-Audio, to our knowledge, is the first to jointly learn continuous audio representations and latent flow matching for TTA. By coupling reconstruction with self-supervised generative prediction, Unite-Audio allows the generative objective to directly shape the latent space rather than treating it as a fixed intermediate representation. We further employ Flow-GRPO post-training to improve text-conditioned generation. Experiments show competitive TTA performance with a compact latent flow model, while ablation studies confirm the benefit of jointly learning the audio representation and generative model. Audio samples are available at https://runwushi.github.io/Unite-Audio.
作者Shreya Deshmukh, Imen Mahdi, Nick Heppert, Abhinav Valada
Advances in generative modeling have recently been extensively employed in robotics for policy learning. In particular, Conditional Flow Matching (CFM) trained with expert demonstrations has been shown to outperform existing methods on robot manipulation benchmarks. While prior work has mainly focused on single-task settings, we study the problem from a multi-task perspective, as training independent models for each task is computationally expensive. Multi-Task policy learning comes with its own set of challenges, as naively training on a concatenated dataset of demonstrations would either require increased model capacity to accommodate the added complexity or result in drops in performance. We propose to distill knowledge from single-task CFM experts into a shared multi-task policy by transferring their learned velocity fields. We combine this distillation signal with the original CFM objective to retain fidelity to the demonstrations. Experiments on RLBench show that our approach improves multi-task policy performance over naive training while maintaining a fixed model size.
展开完整摘要收起摘要↓
Advances in generative modeling have recently been extensively employed in robotics for policy learning. In particular, Conditional Flow Matching (CFM) trained with expert demonstrations has been shown to outperform existing methods on robot manipulation benchmarks. While prior work has mainly focused on single-task settings, we study the problem from a multi-task perspective, as training independent models for each task is computationally expensive. Multi-Task policy learning comes with its own set of challenges, as naively training on a concatenated dataset of demonstrations would either require increased model capacity to accommodate the added complexity or result in drops in performance. We propose to distill knowledge from single-task CFM experts into a shared multi-task policy by transferring their learned velocity fields. We combine this distillation signal with the original CFM objective to retain fidelity to the demonstrations. Experiments on RLBench show that our approach improves multi-task policy performance over naive training while maintaining a fixed model size.
Diffusion models have shown strong potential for kilometer-scale precipitation downscaling, but their performance in geographically unseen regions and event regimes remains insufficiently understood. Building on the wavelet diffusion model (WDM) framework, this study evaluates cross-region and cross-event generalization. Six 3 x 3 deg U.S. regions represent convective, winter, tropical, and atmospheric-river precipitation regimes. Low-resolution inputs are generated by block averaging NOAA Multi-Radar/Multi-Sensor (MRMS) composite reflectivity fields. A WDM trained only on Oklahoma (OK) samples and a WDM trained on all six regions are compared with nearest-neighbor and Bicubic interpolation. Model performance is evaluated using three metric families that measure image-domain reconstruction, spectral and distributional fidelity, and bin-wise precipitation detection. The OK-trained WDM remains competitive outside OK. Although the all-region WDM delivers the best and most consistent overall image-domain and detection performance, its gains are uneven across precipitation intensities. Bin-wise critical success index (CSI) over 5-dBZ reflectivity bins shows that WDM improvements concentrate in localized higher-reflectivity structures, which image-domain metrics partly obscure. In addition, the performance differences among samples are strongly associated with the spatial organization of the precipitation field, quantified by Moran's I as the spatial autocorrelation of each reflectivity bin. The sample-level Moran's I-CSI correlation stratified by sample intensity reaches 0.901 in all six regions, including regions unseen during training. Overall, these findings support future efforts to transfer downscaling models to regions with limited local training data and to generate globally consistent, high-resolution precipitation products.
展开完整摘要收起摘要↓
Diffusion models have shown strong potential for kilometer-scale precipitation downscaling, but their performance in geographically unseen regions and event regimes remains insufficiently understood. Building on the wavelet diffusion model (WDM) framework, this study evaluates cross-region and cross-event generalization. Six 3 x 3 deg U.S. regions represent convective, winter, tropical, and atmospheric-river precipitation regimes. Low-resolution inputs are generated by block averaging NOAA Multi-Radar/Multi-Sensor (MRMS) composite reflectivity fields. A WDM trained only on Oklahoma (OK) samples and a WDM trained on all six regions are compared with nearest-neighbor and Bicubic interpolation. Model performance is evaluated using three metric families that measure image-domain reconstruction, spectral and distributional fidelity, and bin-wise precipitation detection. The OK-trained WDM remains competitive outside OK. Although the all-region WDM delivers the best and most consistent overall image-domain and detection performance, its gains are uneven across precipitation intensities. Bin-wise critical success index (CSI) over 5-dBZ reflectivity bins shows that WDM improvements concentrate in localized higher-reflectivity structures, which image-domain metrics partly obscure. In addition, the performance differences among samples are strongly associated with the spatial organization of the precipitation field, quantified by Moran's I as the spatial autocorrelation of each reflectivity bin. The sample-level Moran's I-CSI correlation stratified by sample intensity reaches 0.901 in all six regions, including regions unseen during training. Overall, these findings support future efforts to transfer downscaling models to regions with limited local training data and to generate globally consistent, high-resolution precipitation products.
作者Ming-Zhi Jiang, An-Tzu Teng, Jun-En Liu, Po-An Chen, Yung-Ming Li
Online discussion of political and gender-related issues is often heated, and when opinions in a network draw closer, the convergence is readily taken as genuine consensus. Whether it carries a cost is a question existing methods cannot answer: coevolutionary opinion formation games measure the Price of Anarchy (PoA) of agents that update by numerical rules, while simulations with large language model (LLM) agents report only descriptive indices. We introduce the Hybrid Coevolutionary Opinion Game (H-COG), in which analytical and LLM-driven agents share one network, choose their neighbors by opinion similarity in every round, and hold stances drawn from real Reddit comments on gun control and abortion. To our knowledge, H-COG is the first framework to place Friedkin-Johnsen best-response agents and LLM agents in one coevolutionary game and to measure the social cost and PoA of LLM-driven populations. We prove that on any fixed network, given the LLM agents' opinions, the analytical agents' opinion stage has a unique equilibrium and the social optimum has a closed form, and that the convergence guarantee of Chen et al. for optimistic gradient ascent carries over to H-COG. All runs converge structurally. LLM-driven populations are less polarized yet have about five times the PoA of analytical ones; half of the gap comes from agents being pulled away from their own prior positions, a distance we prove must carry a cost whenever expressed opinions are more concentrated than intrinsic ones. Echo chambers form under every composition and grow out of the rewiring rule rather than the initial topology. Opinions in these populations draw closer largely because agents give up their own positions.
展开完整摘要收起摘要↓
Online discussion of political and gender-related issues is often heated, and when opinions in a network draw closer, the convergence is readily taken as genuine consensus. Whether it carries a cost is a question existing methods cannot answer: coevolutionary opinion formation games measure the Price of Anarchy (PoA) of agents that update by numerical rules, while simulations with large language model (LLM) agents report only descriptive indices. We introduce the Hybrid Coevolutionary Opinion Game (H-COG), in which analytical and LLM-driven agents share one network, choose their neighbors by opinion similarity in every round, and hold stances drawn from real Reddit comments on gun control and abortion. To our knowledge, H-COG is the first framework to place Friedkin-Johnsen best-response agents and LLM agents in one coevolutionary game and to measure the social cost and PoA of LLM-driven populations. We prove that on any fixed network, given the LLM agents' opinions, the analytical agents' opinion stage has a unique equilibrium and the social optimum has a closed form, and that the convergence guarantee of Chen et al. for optimistic gradient ascent carries over to H-COG. All runs converge structurally. LLM-driven populations are less polarized yet have about five times the PoA of analytical ones; half of the gap comes from agents being pulled away from their own prior positions, a distance we prove must carry a cost whenever expressed opinions are more concentrated than intrinsic ones. Echo chambers form under every composition and grow out of the rewiring rule rather than the initial topology. Opinions in these populations draw closer largely because agents give up their own positions.
作者Mengran Li, Bo Li, Chengyang Zhang, Yang Yan, Jinfeng Xu, Zhenchao Tang
AI virtual cells aim to predict cellular responses to specified interventions, yet held-out predictive performance alone does not establish use of the supplied perturbation information. This prediction-claim gap matters in agentic model discovery, where language-model agents generate and revise predictors using score-based feedback. We introduce CELLAUDIT, which audits input-use claims by asking whether an input can enter the cited computation, whether fitted predictions depend on it, and whether that dependence improves prediction of observed response. On a paired morphology-transcriptomics perturbation benchmark (BBBC047), an agent-selected predictor attains a mean held-out Global Pearson correlation coefficient (PCC) of 0.3153 but remains invariant to compound replacement; a control-profile-only predictor reaches 0.3142. Source inspection identifies a compound-query pathway blocked by singleton key-value attention, and the invariance persists after refitting with disjoint control wells. In a stratified audit of 48 candidates across two linked tasks, 47 change predictions under compound replacement on both held-out folds, but only 20 show target-loss gains with intervals above zero on both folds. On BBBC047, falsification-guided revisions recover positive mean compound contributions while retaining gains over the control-profile-only baseline. In matched sci-Plex searches, audit-enriched feedback yields higher held-out performance and larger mean compound and dose contributions across five trajectories, although paired intervals span zero. Refitting fixed designs on an independently acquired cohort shows predictive generalization need not imply generalization of input-use claims: dose contribution persists, whereas support for compound identity does not. CELLAUDIT adds a falsification layer to agentic model discovery, moving from generate-score-revise toward discover-falsify-revise.
展开完整摘要收起摘要↓
AI virtual cells aim to predict cellular responses to specified interventions, yet held-out predictive performance alone does not establish use of the supplied perturbation information. This prediction-claim gap matters in agentic model discovery, where language-model agents generate and revise predictors using score-based feedback. We introduce CELLAUDIT, which audits input-use claims by asking whether an input can enter the cited computation, whether fitted predictions depend on it, and whether that dependence improves prediction of observed response. On a paired morphology-transcriptomics perturbation benchmark (BBBC047), an agent-selected predictor attains a mean held-out Global Pearson correlation coefficient (PCC) of 0.3153 but remains invariant to compound replacement; a control-profile-only predictor reaches 0.3142. Source inspection identifies a compound-query pathway blocked by singleton key-value attention, and the invariance persists after refitting with disjoint control wells. In a stratified audit of 48 candidates across two linked tasks, 47 change predictions under compound replacement on both held-out folds, but only 20 show target-loss gains with intervals above zero on both folds. On BBBC047, falsification-guided revisions recover positive mean compound contributions while retaining gains over the control-profile-only baseline. In matched sci-Plex searches, audit-enriched feedback yields higher held-out performance and larger mean compound and dose contributions across five trajectories, although paired intervals span zero. Refitting fixed designs on an independently acquired cohort shows predictive generalization need not imply generalization of input-use claims: dose contribution persists, whereas support for compound identity does not. CELLAUDIT adds a falsification layer to agentic model discovery, moving from generate-score-revise toward discover-falsify-revise.
作者Nathalia Gomez, Haig Shamlian, Omar Khan, Tiffany D. Do
As embodied agents take on increasingly social and relational roles in VR, visual realism and embodiment alone may be insufficient; users must also perceive these agents as emotionally attuned, supportive, and humanlike. Prior work suggests that verbal attunement and nonverbal mimicry can each improve users' social evaluations of embodied agents. However, behavioral mimicry has largely been studied outside of real-time, conversational AI interactions, leaving limited understanding of how users respond when an agent simultaneously generates contextually responsive dialogue and adapts its nonverbal behavior during an immersive conversation. To address this gap, we developed an embodied AI counselor that combines conversational AI with real-time facial-expression and posture mimicry, while producing either verbally attuned or neutral responses. We evaluated the system in a 2 X 2 within-subjects study with 20 participants, manipulating verbal attunement and behavioral mimicry. Results showed that verbal attunement was the most reliable driver of perceived empathy. Behavioral mimicry showed a marginal relationship with perceived humanness, while greater mimicry exposure showed preliminary, exploratory positive associations with empathy, positivity, and humanness, particularly among female participants. Together, these findings show that multimodal synchrony is not a simple additive strategy for designing empathic conversational agents in VR and underscore the need to consider how verbal and nonverbal behaviors are combined during real-time interaction.
展开完整摘要收起摘要↓
As embodied agents take on increasingly social and relational roles in VR, visual realism and embodiment alone may be insufficient; users must also perceive these agents as emotionally attuned, supportive, and humanlike. Prior work suggests that verbal attunement and nonverbal mimicry can each improve users' social evaluations of embodied agents. However, behavioral mimicry has largely been studied outside of real-time, conversational AI interactions, leaving limited understanding of how users respond when an agent simultaneously generates contextually responsive dialogue and adapts its nonverbal behavior during an immersive conversation. To address this gap, we developed an embodied AI counselor that combines conversational AI with real-time facial-expression and posture mimicry, while producing either verbally attuned or neutral responses. We evaluated the system in a 2 X 2 within-subjects study with 20 participants, manipulating verbal attunement and behavioral mimicry. Results showed that verbal attunement was the most reliable driver of perceived empathy. Behavioral mimicry showed a marginal relationship with perceived humanness, while greater mimicry exposure showed preliminary, exploratory positive associations with empathy, positivity, and humanness, particularly among female participants. Together, these findings show that multimodal synchrony is not a simple additive strategy for designing empathic conversational agents in VR and underscore the need to consider how verbal and nonverbal behaviors are combined during real-time interaction.
作者Jasem Khelifi, Issam Oukhay, Ali Ouni, Mohammed Sayagh, Mohamed Aymen Saied
Agentic workflows shift software development from prompting AI agents for individual tasks to defining recurring work that agents execute automatically. GitHub Agentic Workflows (gh-aw) enables this approach through Markdown files that combine natural-language instructions with configuration and compile into executable GitHub Actions workflows. Unlike conventional workflows that primarily prescribe scripted operations, these files delegate tasks requiring interpretation to AI agents. They also couple agent instructions with execution triggers, making those instructions operational specifications for repeated repository activities. However, how developers structure and maintain these specifications, and which execution requirements and safeguards they express, remains insufficiently understood. In this paper, we examine the structure, evolution, and instruction content of gh-aw Markdown files to inform how practitioners define and maintain agent-run work. We analyze 1,248 files from 276 repositories, 20,841 commit-file events, and 288 resolved instruction-label sets from a sample of 294 files. Our results show that workflow instructions extend beyond short prompts, with a median of 556.5 words and code blocks in 62.1% of files. Among files with at least 120 days of observed activity, 78.2% still receive updates in month 4, while size-normalized churn decreases after the first month. Tasks, outputs, constraints, and process instructions each appear in over 93% of labeled workflows, yet only 9.4% explicitly address prompt-injection defense. LLM classification achieves an F1 score of 0.818 and Cohen's Kappa of 0.715 against the resolved human labels. These findings suggest that developers should account for the evolution of workflows copied or referenced across repositories and consider adding prompt-injection defenses, resource budgets, and evidencecredibility checks where applicable.
展开完整摘要收起摘要↓
Agentic workflows shift software development from prompting AI agents for individual tasks to defining recurring work that agents execute automatically. GitHub Agentic Workflows (gh-aw) enables this approach through Markdown files that combine natural-language instructions with configuration and compile into executable GitHub Actions workflows. Unlike conventional workflows that primarily prescribe scripted operations, these files delegate tasks requiring interpretation to AI agents. They also couple agent instructions with execution triggers, making those instructions operational specifications for repeated repository activities. However, how developers structure and maintain these specifications, and which execution requirements and safeguards they express, remains insufficiently understood. In this paper, we examine the structure, evolution, and instruction content of gh-aw Markdown files to inform how practitioners define and maintain agent-run work. We analyze 1,248 files from 276 repositories, 20,841 commit-file events, and 288 resolved instruction-label sets from a sample of 294 files. Our results show that workflow instructions extend beyond short prompts, with a median of 556.5 words and code blocks in 62.1% of files. Among files with at least 120 days of observed activity, 78.2% still receive updates in month 4, while size-normalized churn decreases after the first month. Tasks, outputs, constraints, and process instructions each appear in over 93% of labeled workflows, yet only 9.4% explicitly address prompt-injection defense. LLM classification achieves an F1 score of 0.818 and Cohen's Kappa of 0.715 against the resolved human labels. These findings suggest that developers should account for the evolution of workflows copied or referenced across repositories and consider adding prompt-injection defenses, resource budgets, and evidencecredibility checks where applicable.
We present Hunyuan-A13B, an open-source large language model based on a Mixture-of-Experts architecture. It contains 80 billion total parameters but activates only 13 billion during inference, balancing model capability, computational efficiency, and deployment cost. The model is pretrained on a rigorously filtered 20T-token corpus with enhanced STEM data curation, improving factual reliability and reasoning ability. High-quality supervised fine-tuning and large-scale reinforcement learning further enhance its overall performance. Hunyuan-A13B also introduces a dual-mode Chain-of-Thought framework that adapts reasoning depth to task complexity: fast thinking for routine queries and slow thinking for complex, multi-step problems. Evaluations show competitive performance across mathematics, science, programming, general language understanding, and agent tasks, often approaching that of much larger models. Its high inference throughput makes it suitable for latency-sensitive applications. We release Hunyuan-A13B to support open research and practical LLM deployment.
展开完整摘要收起摘要↓
We present Hunyuan-A13B, an open-source large language model based on a Mixture-of-Experts architecture. It contains 80 billion total parameters but activates only 13 billion during inference, balancing model capability, computational efficiency, and deployment cost. The model is pretrained on a rigorously filtered 20T-token corpus with enhanced STEM data curation, improving factual reliability and reasoning ability. High-quality supervised fine-tuning and large-scale reinforcement learning further enhance its overall performance. Hunyuan-A13B also introduces a dual-mode Chain-of-Thought framework that adapts reasoning depth to task complexity: fast thinking for routine queries and slow thinking for complex, multi-step problems. Evaluations show competitive performance across mathematics, science, programming, general language understanding, and agent tasks, often approaching that of much larger models. Its high inference throughput makes it suitable for latency-sensitive applications. We release Hunyuan-A13B to support open research and practical LLM deployment.
We study coding agents for long-horizon, dexterous robotics and ask whether their solutions can provide scalable supervision for learning general robot policies. To test this, we develop EMBODIEDSWE-BENCH, a simulation benchmark for coding agents spanning contact-rich manipulation, deformable objects, and long-horizon tasks requiring up to half an hour of continuous interaction. We find that frontier coding agents can solve complex long-horizon tasks and transfer prior solutions across both tasks and embodiments. We also design supporting tools that help agents more effectively solve these tasks. However, the resulting solutions require substantial iterative interaction and are typically specialized to individual task instances. We therefore introduce EMBODIEDSWE-GEN, which expands a single solution from coding agent into large diverse trajectories for training a VLA. VLA performance improves with more generated demonstrations, and agent-aided diversification improves generalization to held-out task variations. We also show that a VLA finetuned solely on coding-agent-generated simulation demonstrations completes a long-horizon task on real robot. Together, our framework uses coding agents to solve complex robotics tasks and turn verified solutions into scalable supervision for robot policies.
展开完整摘要收起摘要↓
We study coding agents for long-horizon, dexterous robotics and ask whether their solutions can provide scalable supervision for learning general robot policies. To test this, we develop EMBODIEDSWE-BENCH, a simulation benchmark for coding agents spanning contact-rich manipulation, deformable objects, and long-horizon tasks requiring up to half an hour of continuous interaction. We find that frontier coding agents can solve complex long-horizon tasks and transfer prior solutions across both tasks and embodiments. We also design supporting tools that help agents more effectively solve these tasks. However, the resulting solutions require substantial iterative interaction and are typically specialized to individual task instances. We therefore introduce EMBODIEDSWE-GEN, which expands a single solution from coding agent into large diverse trajectories for training a VLA. VLA performance improves with more generated demonstrations, and agent-aided diversification improves generalization to held-out task variations. We also show that a VLA finetuned solely on coding-agent-generated simulation demonstrations completes a long-horizon task on real robot. Together, our framework uses coding agents to solve complex robotics tasks and turn verified solutions into scalable supervision for robot policies.
Time-series foundation models (TSFMs) provide forecasts for operational decisions, but accuracy alone does not determine their value. Evaluating agents that use these models requires measuring decision quality and forecast cost. FWBench evaluates this capability on 1,251 electricity and cycle-hire cases using fixed forecast tools and simulated capacity contracts. Agents select models, histories and horizons, then submit capacities to minimize a stated loss-cost objective. We evaluated two hosted and eight local configurations, including small language models, and tested local models with and without TSFMs. GPT-6 Astra bought inexpensive short-horizon forecasts selectively, using 2.5% of the budget, and outperformed fixed policies when the saved decisions were scored with three loss-cost weightings. FWBench enables reproducible evaluation of how language models select and use time-series forecasts to make decisions under cost constraints.
展开完整摘要收起摘要↓
Time-series foundation models (TSFMs) provide forecasts for operational decisions, but accuracy alone does not determine their value. Evaluating agents that use these models requires measuring decision quality and forecast cost. FWBench evaluates this capability on 1,251 electricity and cycle-hire cases using fixed forecast tools and simulated capacity contracts. Agents select models, histories and horizons, then submit capacities to minimize a stated loss-cost objective. We evaluated two hosted and eight local configurations, including small language models, and tested local models with and without TSFMs. GPT-6 Astra bought inexpensive short-horizon forecasts selectively, using 2.5% of the budget, and outperformed fixed policies when the saved decisions were scored with three loss-cost weightings. FWBench enables reproducible evaluation of how language models select and use time-series forecasts to make decisions under cost constraints.
AI research agents must predict the effects of computational changes after budgeted experiments. WhatWorkedBench evaluates this experimental understanding through a delivered response surface of configuration scores. Agents inspect workflow code and buy measurements; exhaustive CPU references score conditional component effects, mean pair interactions, configuration choice, and delivery. The catalog spans 36 task conditions, 30 sources, 8 workflow families, and 1248 indexed configuration records, with 4,206 numerical controls. At eight purchased measurements plus two free anchors, pair-effect ridge selects an exact optimum on 15 of 22 sources; 13 of these cases have at least one conditional-effect error exceeding 10% of the task utility range. Shared-estimator comparisons measure acquisition and reconstruction on common observations. In a prospective typed study on 12 four-factor sources, DeepSeek Flash submits 12/12 direct tables and gains 0.147 recovery over a Gaussian process (GP) fitted to the same observations. Pro delivers 11/12 artifacts, with an all-attempt GP difference of -0.001 and a delivered-only difference of +0.063. On six prespecified new agent-evaluation sources, Flash and Pro gains are 0.149 and 0.074. In eight typed six-factor episodes, seven final tables satisfy verified code equivalences. Four fresh agents pass all six registered rules through named estimators and deliver consistent tables at 0.682 recovery versus 0.710 for separate direct-table runs. WhatWorkedBench links acquisition, inference, program structure, and delivery.
展开完整摘要收起摘要↓
AI research agents must predict the effects of computational changes after budgeted experiments. WhatWorkedBench evaluates this experimental understanding through a delivered response surface of configuration scores. Agents inspect workflow code and buy measurements; exhaustive CPU references score conditional component effects, mean pair interactions, configuration choice, and delivery. The catalog spans 36 task conditions, 30 sources, 8 workflow families, and 1248 indexed configuration records, with 4,206 numerical controls. At eight purchased measurements plus two free anchors, pair-effect ridge selects an exact optimum on 15 of 22 sources; 13 of these cases have at least one conditional-effect error exceeding 10% of the task utility range. Shared-estimator comparisons measure acquisition and reconstruction on common observations. In a prospective typed study on 12 four-factor sources, DeepSeek Flash submits 12/12 direct tables and gains 0.147 recovery over a Gaussian process (GP) fitted to the same observations. Pro delivers 11/12 artifacts, with an all-attempt GP difference of -0.001 and a delivered-only difference of +0.063. On six prespecified new agent-evaluation sources, Flash and Pro gains are 0.149 and 0.074. In eight typed six-factor episodes, seven final tables satisfy verified code equivalences. Four fresh agents pass all six registered rules through named estimators and deliver consistent tables at 0.682 recovery versus 0.710 for separate direct-table runs. WhatWorkedBench links acquisition, inference, program structure, and delivery.
作者Shenhua Gu, Hongqiang Zhu, Fan Zhang, Jinming Zhang, Hao Chen
LLM coding agents increasingly operate inside game-engine repositories, where application logic is inseparable from serialized assets: a single gameplay change may span C\# scripts, prefabs, scenes, and ScriptableObjects wired together by Unity GUIDs. The retrieval tools agents carry today---shell utilities and code-only indexes---cannot answer basic cross-file questions, because these relationships live in .meta files and YAML assets rather than in code. We present Unity Insight, to our knowledge the first persistent, LLM-facing, agent-integrated cross-file code--asset index for Unity projects, shipping in production with Tuanjie Codely, the agent CLI of Tuanjie Engine, since its public launch on 2026-07-28. In a paired experiment---28 project-specific questions on two Unity games, same model and harness, one run per arm per question---the index-backed agent spent 53% fewer tokens and 52% less wall-clock time than a general-purpose exploration agent (exact paired sign tests, $p{<}0.004$), using only its typed index-query tools.
展开完整摘要收起摘要↓
LLM coding agents increasingly operate inside game-engine repositories, where application logic is inseparable from serialized assets: a single gameplay change may span C\# scripts, prefabs, scenes, and ScriptableObjects wired together by Unity GUIDs. The retrieval tools agents carry today---shell utilities and code-only indexes---cannot answer basic cross-file questions, because these relationships live in .meta files and YAML assets rather than in code. We present Unity Insight, to our knowledge the first persistent, LLM-facing, agent-integrated cross-file code--asset index for Unity projects, shipping in production with Tuanjie Codely, the agent CLI of Tuanjie Engine, since its public launch on 2026-07-28. In a paired experiment---28 project-specific questions on two Unity games, same model and harness, one run per arm per question---the index-backed agent spent 53% fewer tokens and 52% less wall-clock time than a general-purpose exploration agent (exact paired sign tests, $p{<}0.004$), using only its typed index-query tools.
作者Zhilong Ge, Yuting Shao, Yutao Yang, Yuxuan Cai, Jie Zhou, Kai Chen, Bo Zhang, Qin Chen, Liang He
Human-written agent skills encode rich workflows for real-world problem solving, but are typically used as external inference-time instructions rather than internalized as reusable model capabilities. We introduce SkillGym, a framework that transforms these skills into executable, verifiable training environments for large language model agents. Its skill-to-task pipeline instantiates concrete tasks, verifies outcomes with code-based checkers, and assesses empirical skill dependence through contrastive executions. We construct and release 2,756 environments across 12 categories and collect 8,364 successful trajectories from multiple models and harnesses, averaging 49 tool calls and over 60k logged text tokens. These resources support supervised fine-tuning on verified workflows and reinforcement learning with outcome-based rewards. Under Claude Code, supervised fine-tuning improves Qwen3.5-35B-A3B by 199 Elo on GDPval-AA v2, 19.10 percentage points on Terminal-Bench 2.1, and 28.13 and 12.38 points on SkillsBench v1.1 with and without skills, respectively. Our 35B SkillGym-Agent reaches 51.47% on skill-assisted SkillsBench, exceeding reported scores for Claude Sonnet 4.6, GPT-5.4 Mini, and DeepSeek V4 Pro. Without skills, it also surpasses skill-assisted bases under Codex and Claude Code, suggesting reusable procedural competence.
展开完整摘要收起摘要↓
Human-written agent skills encode rich workflows for real-world problem solving, but are typically used as external inference-time instructions rather than internalized as reusable model capabilities. We introduce SkillGym, a framework that transforms these skills into executable, verifiable training environments for large language model agents. Its skill-to-task pipeline instantiates concrete tasks, verifies outcomes with code-based checkers, and assesses empirical skill dependence through contrastive executions. We construct and release 2,756 environments across 12 categories and collect 8,364 successful trajectories from multiple models and harnesses, averaging 49 tool calls and over 60k logged text tokens. These resources support supervised fine-tuning on verified workflows and reinforcement learning with outcome-based rewards. Under Claude Code, supervised fine-tuning improves Qwen3.5-35B-A3B by 199 Elo on GDPval-AA v2, 19.10 percentage points on Terminal-Bench 2.1, and 28.13 and 12.38 points on SkillsBench v1.1 with and without skills, respectively. Our 35B SkillGym-Agent reaches 51.47% on skill-assisted SkillsBench, exceeding reported scores for Claude Sonnet 4.6, GPT-5.4 Mini, and DeepSeek V4 Pro. Without skills, it also surpasses skill-assisted bases under Codex and Claude Code, suggesting reusable procedural competence.
Consumers increasingly delegate purchasing decisions to Large Language Models (LLMs) acting as surrogate consumers. Using "Tool-Lab," an adaptation of information-board process tracing that places product attributes behind costly tool calls, we examine how marketing pricing cues (i.e., just-below pricing and promotional framing) influence AI shopping agents. Across eight commercially deployed LLMs from three providers, we trace pre-choice information acquisition. Under zero cost, pricing cues rarely mislead. Imposing acquisition costs under a vague goal prompt leads LLMs to omit diagnostic attributes required to compute unit price and choose suboptimal choices resembling human heuristics. Relative to a specific goal prompt that mainly preserves diagnostic search and choice optimality, a vague goal prompt under constraints creates a search-mediated vulnerability. This research demonstrates that marketing heuristics in delegated AI shopping are governed by storefront information architecture, not necessarily immutable LLM flaws.
展开完整摘要收起摘要↓
Consumers increasingly delegate purchasing decisions to Large Language Models (LLMs) acting as surrogate consumers. Using "Tool-Lab," an adaptation of information-board process tracing that places product attributes behind costly tool calls, we examine how marketing pricing cues (i.e., just-below pricing and promotional framing) influence AI shopping agents. Across eight commercially deployed LLMs from three providers, we trace pre-choice information acquisition. Under zero cost, pricing cues rarely mislead. Imposing acquisition costs under a vague goal prompt leads LLMs to omit diagnostic attributes required to compute unit price and choose suboptimal choices resembling human heuristics. Relative to a specific goal prompt that mainly preserves diagnostic search and choice optimality, a vague goal prompt under constraints creates a search-mediated vulnerability. This research demonstrates that marketing heuristics in delegated AI shopping are governed by storefront information architecture, not necessarily immutable LLM flaws.
作者Chuyi Wang, Xiaohui Xie, Tongze Wang, Fangchen Luo, Yong Cui
LLMs increasingly operate through coding-agent harnesses that inspect repositories, invoke tools, and modify files. Substituting the model behind such an agent can therefore change security-relevant decisions, including whether it verifies changes or recovers safely from failures. Existing LLM fingerprints largely infer identity from direct text or token distributions. In coding agents, these signals are mediated by system instructions, controller logic, tools, and execution feedback, limiting their transfer. We present LIDAR (LLM Identification from Decisions and Actions at Runtime), an active black-box fingerprinting method for coding-agent execution. Three coding probe pairs expose post-edit verification, transient-failure recovery, and specification--test conflict resolution under controlled changes. LIDAR represents the resulting trajectories with complementary instance-level and distribution-level features and compares them with clean references using a lightweight probabilistic identifier. It requires no access to model weights, logits, or provider internals. Across 36 models from seven families and two agent harnesses, LIDAR achieves high Top-1 accuracy and MRR and outperforms four existing fingerprinting and API-auditing baselines. Ablations confirm that the two feature levels, all probe pairs, and their controlled variants contribute. These results show that agent execution behavior provides model-identity evidence beyond final outputs.
展开完整摘要收起摘要↓
LLMs increasingly operate through coding-agent harnesses that inspect repositories, invoke tools, and modify files. Substituting the model behind such an agent can therefore change security-relevant decisions, including whether it verifies changes or recovers safely from failures. Existing LLM fingerprints largely infer identity from direct text or token distributions. In coding agents, these signals are mediated by system instructions, controller logic, tools, and execution feedback, limiting their transfer. We present LIDAR (LLM Identification from Decisions and Actions at Runtime), an active black-box fingerprinting method for coding-agent execution. Three coding probe pairs expose post-edit verification, transient-failure recovery, and specification--test conflict resolution under controlled changes. LIDAR represents the resulting trajectories with complementary instance-level and distribution-level features and compares them with clean references using a lightweight probabilistic identifier. It requires no access to model weights, logits, or provider internals. Across 36 models from seven families and two agent harnesses, LIDAR achieves high Top-1 accuracy and MRR and outperforms four existing fingerprinting and API-auditing baselines. Ablations confirm that the two feature levels, all probe pairs, and their controlled variants contribute. These results show that agent execution behavior provides model-identity evidence beyond final outputs.
作者Zheng Zhang, Liu Liu, Qi Chai, Deheng Ye, Peilin Zhao, Mao Zheng, Hao Wang
Role-playing agents based on large language models have been widely applied in areas such as personalized assistance and social simulation. Recent RL methods typically train on a fixed scenario pool collected before learning begins. This creates a distributional bottleneck: as the agent improves, the scenarios where it performs poorly also change, while the training distribution remains static. Therefore, we propose AdvRole, an adversarial context rewriting framework that turns role-playing RL into a closed-loop curriculum. AdvRole alternates between an Actor that learns to role-play and a Rewriter that edits character profiles and dialogue contexts into actor-specific hard scenarios. The Rewriter is trained with a performance-gap reward, which favors rewrites that reduce the current Actor's score relative to the original scenario. As a result, the scenario pool evolves with the Actor and continuously targets under-mastered regions of the character-context space. Experiments on three role-playing benchmarks covering English and Chinese, as well as a new multilingual benchmark we release, show that AdvRole consistently outperforms baselines.
展开完整摘要收起摘要↓
Role-playing agents based on large language models have been widely applied in areas such as personalized assistance and social simulation. Recent RL methods typically train on a fixed scenario pool collected before learning begins. This creates a distributional bottleneck: as the agent improves, the scenarios where it performs poorly also change, while the training distribution remains static. Therefore, we propose AdvRole, an adversarial context rewriting framework that turns role-playing RL into a closed-loop curriculum. AdvRole alternates between an Actor that learns to role-play and a Rewriter that edits character profiles and dialogue contexts into actor-specific hard scenarios. The Rewriter is trained with a performance-gap reward, which favors rewrites that reduce the current Actor's score relative to the original scenario. As a result, the scenario pool evolves with the Actor and continuously targets under-mastered regions of the character-context space. Experiments on three role-playing benchmarks covering English and Chinese, as well as a new multilingual benchmark we release, show that AdvRole consistently outperforms baselines.
作者Yue Huang, Zhangchen Xu, Yuchen Ma, Wenjie Wang, Zheyuan Liu, Ziwei Xu, Pin-Yu Chen, Michel Galley, Zinan Lin, Stefan Feuerriegel, Radha Poovendran, Misha Sra, Alex Pentland, Xiangliang Zhang, Zichen Chen
Autonomous research agents can design experiments, evaluate results, and write reports, giving them control over both a scientific result and the evidence used to support it. This creates a risk of reward hacking: meeting the reward criteria without achieving the intended goal. We study (1) how often models reward-hack without instructions to do so, (2) how effective and detectable their methods are when hacking is allowed, and (3) how they adapt when an LLM review panel returns its decision and reasons. Across 17 language models and 38 tasks, the spontaneous reward-hacking rate is 30.5% on open-ended research-pipeline tasks and 2.9% on task-specific kernels. When hacking is allowed on tasks whose pass thresholds exceed our best compliant baselines, 505/677 attempts (74.6%) are confirmed reward hacks: they both clear the threshold and receive mechanism-verification panel confirmation of an evaluation exploit. An LLM panel reviewing only submitted code and reported scores misses 33/505 confirmed hacks (6.5%). Direct methods that achieve the highest scores are often easy to detect, while less direct methods evade more often. In a five-round loop, the number of model-task pairs with an evasion rises from 7 to 56. Among 79 pairs evaluated under two feedback conditions, cumulative evasion reaches 40.5% with detailed feedback and 20.3% with generic rejection. The detailed condition includes the review decision, reasons, and attempt history, so this comparison does not isolate the effect of explanations. These findings highlight the need for stronger defenses, including metrics kept outside the agent's control and independent recomputation on data chosen to expose likely exploits.
展开完整摘要收起摘要↓
Autonomous research agents can design experiments, evaluate results, and write reports, giving them control over both a scientific result and the evidence used to support it. This creates a risk of reward hacking: meeting the reward criteria without achieving the intended goal. We study (1) how often models reward-hack without instructions to do so, (2) how effective and detectable their methods are when hacking is allowed, and (3) how they adapt when an LLM review panel returns its decision and reasons. Across 17 language models and 38 tasks, the spontaneous reward-hacking rate is 30.5% on open-ended research-pipeline tasks and 2.9% on task-specific kernels. When hacking is allowed on tasks whose pass thresholds exceed our best compliant baselines, 505/677 attempts (74.6%) are confirmed reward hacks: they both clear the threshold and receive mechanism-verification panel confirmation of an evaluation exploit. An LLM panel reviewing only submitted code and reported scores misses 33/505 confirmed hacks (6.5%). Direct methods that achieve the highest scores are often easy to detect, while less direct methods evade more often. In a five-round loop, the number of model-task pairs with an evasion rises from 7 to 56. Among 79 pairs evaluated under two feedback conditions, cumulative evasion reaches 40.5% with detailed feedback and 20.3% with generic rejection. The detailed condition includes the review decision, reasons, and attempt history, so this comparison does not isolate the effect of explanations. These findings highlight the need for stronger defenses, including metrics kept outside the agent's control and independent recomputation on data chosen to expose likely exploits.
作者Ewelina Gajewska, Katarzyna Budzynska, Jaroslaw Chudziak
Large Language Models (LLMs) are increasingly deployed as argumentative agents in persuasive dialogues, necessitating rigorous evaluation of their debating competence relative to human interlocutors. In this study, we focus on character attacks (ad hominem arguments), traditionally dismissed as fallacies, which play a pivotal role in political persuasive dialogues where ethos often rivals propositional content. Specifically, we investigate whether modern LLMs can replicate human competence to strategically use and respond to such attacks. We analyse a corpus of natural language political dialogues to identify defensive strategies human interlocutors naturally employ in ethos-centred debates and structure them into a dialogue game. Empirically, we benchmark LLM-generated dialogues against the ElecDeb60to16-fallacy corpus of U.S. presidential debates, contrasting human debaters' repertoire of defensive strategies with those of artificial agents. Results reveal a substantial difference: most LLMs rigidly prioritise logical defences, failing to exploit ethotic counterattacks as valid moves in political discourse. We argue that current safety fine-tuning constraints the strategic action space of these LLMs, making them unable to fully engage in naturalistic interactions within domains where character contestation is a normative expectation rather than a mere fallacy.
展开完整摘要收起摘要↓
Large Language Models (LLMs) are increasingly deployed as argumentative agents in persuasive dialogues, necessitating rigorous evaluation of their debating competence relative to human interlocutors. In this study, we focus on character attacks (ad hominem arguments), traditionally dismissed as fallacies, which play a pivotal role in political persuasive dialogues where ethos often rivals propositional content. Specifically, we investigate whether modern LLMs can replicate human competence to strategically use and respond to such attacks. We analyse a corpus of natural language political dialogues to identify defensive strategies human interlocutors naturally employ in ethos-centred debates and structure them into a dialogue game. Empirically, we benchmark LLM-generated dialogues against the ElecDeb60to16-fallacy corpus of U.S. presidential debates, contrasting human debaters' repertoire of defensive strategies with those of artificial agents. Results reveal a substantial difference: most LLMs rigidly prioritise logical defences, failing to exploit ethotic counterattacks as valid moves in political discourse. We argue that current safety fine-tuning constraints the strategic action space of these LLMs, making them unable to fully engage in naturalistic interactions within domains where character contestation is a normative expectation rather than a mere fallacy.
作者Yakup Koray Budanaz, Pratyai Mazumder, Alexandru Calotoiu, Torsten Hoefler
Imperative code fixes an execution order the computation does not require, and a parallelizing compiler must prove which parts of that order it can remove. We introduce the Canonical Parallel Form (CPF), a device-neutral program state from which every ordering constraint our analyses prove unnecessary has been removed. CPF is reached by output-preserving normalization, by lifting semantic operations such as tensor contractions, and by deriving parallelism in three levels ordered by decidability: syntactic subscript tests, exact affine dependence tests over integer sets, and SMT queries over non-linear integer arithmetic, which also admit parallelism guarded behind a runtime check. Heuristics can then specialize the canonical form for each architecture. Across 248 loop-level reasoning kernels on an AMD MI300A, CPF reaches 4.4x over Numba on its 24 Zen 4 cores and 25.4x on its CDNA 3 GPU. Against the other auto-parallelizing optimizers, CPF is 2.9x faster on the CPU and 8.7x on the GPU than DaCe's own auto-parallelizer, 1.9x faster than Pluto, and 1.5x faster than PPCG on the affine subset of the kernels. Because the pipeline is deterministic, CPF also serves as an agent's starting source, cutting the token cost per kernel by up to a factor of 2.72x while leaving the achieved speed-up unchanged, since the coding agents reason less about parallelism.
展开完整摘要收起摘要↓
Imperative code fixes an execution order the computation does not require, and a parallelizing compiler must prove which parts of that order it can remove. We introduce the Canonical Parallel Form (CPF), a device-neutral program state from which every ordering constraint our analyses prove unnecessary has been removed. CPF is reached by output-preserving normalization, by lifting semantic operations such as tensor contractions, and by deriving parallelism in three levels ordered by decidability: syntactic subscript tests, exact affine dependence tests over integer sets, and SMT queries over non-linear integer arithmetic, which also admit parallelism guarded behind a runtime check. Heuristics can then specialize the canonical form for each architecture. Across 248 loop-level reasoning kernels on an AMD MI300A, CPF reaches 4.4x over Numba on its 24 Zen 4 cores and 25.4x on its CDNA 3 GPU. Against the other auto-parallelizing optimizers, CPF is 2.9x faster on the CPU and 8.7x on the GPU than DaCe's own auto-parallelizer, 1.9x faster than Pluto, and 1.5x faster than PPCG on the affine subset of the kernels. Because the pipeline is deterministic, CPF also serves as an agent's starting source, cutting the token cost per kernel by up to a factor of 2.72x while leaving the achieved speed-up unchanged, since the coding agents reason less about parallelism.