Video generation powers a vast array of downstream applications. However, while the de facto standard, i.e., latent diffusion models, typically employ heavily conditioned denoising networks, their decoders often remain unconditional. We observe that this architectural asymmetry leads to significant loss of detail and inconsistency relative to the input image. To address this, we argue that the decoder requires equal conditioning to preserve structural integrity. We introduce RefDecoder, a reference-conditioned video VAE decoder by injecting high-fidelity reference image signal directly into the decoding process via reference attention. Specifically, a lightweight image encoder maps the reference frame into the detail-rich high-dimensional tokens, which are co-processed with the denoised video latent tokens at each decoder up-sampling stage. We demonstrate consistent improvements across several distinct decoder backbones (e.g., Wan 2.1 and VideoVAE+), achieving up to +2.1dB PSNR over the unconditional baselines on the Inter4K, WebVid, and Large Motion reconstruction benchmarks. Notably, RefDecoder can be directly swapped into existing video generation systems without additional fine-tuning, and we report across-the-board improvements in subject consistency, background consistency, and overall quality scores on the VBench I2V benchmark. Beyond I2V, RefDecoder generalizes well to a wide range of visual generation tasks such as style transfer and video editing refinement.
展开完整摘要收起摘要↓
Video generation powers a vast array of downstream applications. However, while the de facto standard, i.e., latent diffusion models, typically employ heavily conditioned denoising networks, their decoders often remain unconditional. We observe that this architectural asymmetry leads to significant loss of detail and inconsistency relative to the input image. To address this, we argue that the decoder requires equal conditioning to preserve structural integrity. We introduce RefDecoder, a reference-conditioned video VAE decoder by injecting high-fidelity reference image signal directly into the decoding process via reference attention. Specifically, a lightweight image encoder maps the reference frame into the detail-rich high-dimensional tokens, which are co-processed with the denoised video latent tokens at each decoder up-sampling stage. We demonstrate consistent improvements across several distinct decoder backbones (e.g., Wan 2.1 and VideoVAE+), achieving up to +2.1dB PSNR over the unconditional baselines on the Inter4K, WebVid, and Large Motion reconstruction benchmarks. Notably, RefDecoder can be directly swapped into existing video generation systems without additional fine-tuning, and we report across-the-board improvements in subject consistency, background consistency, and overall quality scores on the VBench I2V benchmark. Beyond I2V, RefDecoder generalizes well to a wide range of visual generation tasks such as style transfer and video editing refinement.
作者Yidong Huang, Zun Wang, Han Lin, Dong-Ki Kim, Shayegan Omidshafiei, Jaehong Yoon, Jaemin Cho, Yue Zhang, Mohit Bansal
Generating realistic human motion is a central yet unsolved challenge in video generation. While reinforcement learning (RL)-based post-training has driven recent gains in general video quality, extending it to human motion remains bottlenecked by a reward signal that cannot reliably score motion realism. Existing video rewards primarily rely on 2D perceptual signals, without explicitly modeling the 3D body state, contact, and dynamics underlying articulated human motion, and often assign high scores to videos with floating bodies or physically implausible movements. To address this, we propose PhyMotion, a structured, fine-grained motion reward that grounds recovered 3D human trajectories in a physics simulator and evaluates motion quality along multiple dimensions of physical feasibility. Concretely, we recover SMPL body meshes from generated videos, retarget them onto a humanoid in the MuJoCo physics simulator, and evaluate the resulting motion along three axes: kinematic plausibility, contact and balance consistency, and dynamic feasibility. Each component provides a continuous and interpretable signal tied to a specific aspect of motion quality, allowing the reward to capture which aspects of motion are physically correct or violated. Experiments show that PhyMotion achieves stronger correlation with human judgments than existing reward formulations. These gains carry over to RL-based post-training, where optimizing PhyMotion leads to larger and more consistent improvements than optimizing existing rewards, improving motion realism across both autoregressive and bidirectional video generators under both automatic metrics and blind human evaluation (+68 Elo gain). Ablations show that the three axes provide complementary supervision signals, while the reward preserves overall video generation quality with only modest training overhead.
展开完整摘要收起摘要↓
Generating realistic human motion is a central yet unsolved challenge in video generation. While reinforcement learning (RL)-based post-training has driven recent gains in general video quality, extending it to human motion remains bottlenecked by a reward signal that cannot reliably score motion realism. Existing video rewards primarily rely on 2D perceptual signals, without explicitly modeling the 3D body state, contact, and dynamics underlying articulated human motion, and often assign high scores to videos with floating bodies or physically implausible movements. To address this, we propose PhyMotion, a structured, fine-grained motion reward that grounds recovered 3D human trajectories in a physics simulator and evaluates motion quality along multiple dimensions of physical feasibility. Concretely, we recover SMPL body meshes from generated videos, retarget them onto a humanoid in the MuJoCo physics simulator, and evaluate the resulting motion along three axes: kinematic plausibility, contact and balance consistency, and dynamic feasibility. Each component provides a continuous and interpretable signal tied to a specific aspect of motion quality, allowing the reward to capture which aspects of motion are physically correct or violated. Experiments show that PhyMotion achieves stronger correlation with human judgments than existing reward formulations. These gains carry over to RL-based post-training, where optimizing PhyMotion leads to larger and more consistent improvements than optimizing existing rewards, improving motion realism across both autoregressive and bidirectional video generators under both automatic metrics and blind human evaluation (+68 Elo gain). Ablations show that the three axes provide complementary supervision signals, while the reward preserves overall video generation quality with only modest training overhead.
Causal autoregressive video diffusion models support real-time streaming generation by extrapolating future chunks from previously generated content. Distilling such generators from high-fidelity bidirectional teachers yields competitive few-step models, yet a persistent gap between the history distributions encountered during training and those arising at inference constrains generation quality over long horizons. We introduce the Real-time Autoregressive Video Extrapolation Network (RAVEN), a training-time test framework that repacks each self rollout into an interleaved sequence of clean historical endpoints and noisy denoising states. This formulation aligns training attention with inference-time extrapolation and allows downstream chunk losses to supervise the history representations on which future predictions depend. We further propose Consistency-model Group Relative Policy Optimization (CM-GRPO), which reformulates a consistency sampling step as a conditional Gaussian transition and applies online Reinforcement Learning (RL) directly to this kernel, avoiding the Euler-Maruyama auxiliary process adopted in prior flow-model RL formulations. Experiments demonstrate that RAVEN surpasses recent causal video distillation baselines across quality, semantic, and dynamic degree evaluations, and that CM-GRPO provides further gains when combined with RAVEN.
展开完整摘要收起摘要↓
Causal autoregressive video diffusion models support real-time streaming generation by extrapolating future chunks from previously generated content. Distilling such generators from high-fidelity bidirectional teachers yields competitive few-step models, yet a persistent gap between the history distributions encountered during training and those arising at inference constrains generation quality over long horizons. We introduce the Real-time Autoregressive Video Extrapolation Network (RAVEN), a training-time test framework that repacks each self rollout into an interleaved sequence of clean historical endpoints and noisy denoising states. This formulation aligns training attention with inference-time extrapolation and allows downstream chunk losses to supervise the history representations on which future predictions depend. We further propose Consistency-model Group Relative Policy Optimization (CM-GRPO), which reformulates a consistency sampling step as a conditional Gaussian transition and applies online Reinforcement Learning (RL) directly to this kernel, avoiding the Euler-Maruyama auxiliary process adopted in prior flow-model RL formulations. Experiments demonstrate that RAVEN surpasses recent causal video distillation baselines across quality, semantic, and dynamic degree evaluations, and that CM-GRPO provides further gains when combined with RAVEN.
Autoregressive video diffusion models support real-time synthesis but suffer from error accumulation and context loss over long horizons. We discover that attention heads in AR video diffusion transformers serve functionally distinct roles as local heads for detail refinement, anchor heads for structural stabilization, and memory heads for long-range context aggregation, yet existing methods treat them uniformly, leading to suboptimal KV cache allocation. We propose Head Forcing, a training-free framework that assigns each head type a tailored KV cache strategy: local and anchor heads retain only essential tokens, while memory heads employ a hierarchical memory system with dynamic episodic updates for long-range consistency. A head-wise RoPE re-encoding scheme further ensures positional encodings remain within the pretrained range. Without additional training, Head Forcing extends generation from 5 seconds to minute-level duration, supports multi-prompt interactive synthesis, and consistently outperforms existing baselines. Project Page: https://jiahaotian-sjtu.github.io/headforcing.github.io/.
展开完整摘要收起摘要↓
Autoregressive video diffusion models support real-time synthesis but suffer from error accumulation and context loss over long horizons. We discover that attention heads in AR video diffusion transformers serve functionally distinct roles as local heads for detail refinement, anchor heads for structural stabilization, and memory heads for long-range context aggregation, yet existing methods treat them uniformly, leading to suboptimal KV cache allocation. We propose Head Forcing, a training-free framework that assigns each head type a tailored KV cache strategy: local and anchor heads retain only essential tokens, while memory heads employ a hierarchical memory system with dynamic episodic updates for long-range consistency. A head-wise RoPE re-encoding scheme further ensures positional encodings remain within the pretrained range. Without additional training, Head Forcing extends generation from 5 seconds to minute-level duration, supports multi-prompt interactive synthesis, and consistently outperforms existing baselines. Project Page: https://jiahaotian-sjtu.github.io/headforcing.github.io/.
Interactive real-time autoregressive video generation is essential for applications such as content creation and world modeling, where visual content must adapt to dynamically evolving event conditions. A fundamental challenge lies in balancing reactivity and stability: models must respond promptly to new events while maintaining temporal coherence over long horizons. Existing approaches distill bidirectional models into autoregressive generators and further adapt them via streaming long tuning, yet often exhibit persistent drift after condition changes. We identify the cause as conditional bias, where the teacher may provide condition-aligned but trajectory-agnostic guidance, biasing generation toward locally valid yet globally inconsistent modes. Inspired by Trust Region Policy Optimization, we propose Delta Forcing, a simple yet effective framework that constrains unreliable teacher supervision within an adaptive trust region. Specifically, Delta Forcing estimates transition consistency from the latent delta between teacher and generator trajectories, and uses it to balance teacher supervision with a monotonic continuity objective. This suppress unreliable teacher-induced shifts while preserving responsiveness to new events. Extensive experiments demonstrate that Delta Forcing significantly improves consistency while maintaining event reactivity.
展开完整摘要收起摘要↓
Interactive real-time autoregressive video generation is essential for applications such as content creation and world modeling, where visual content must adapt to dynamically evolving event conditions. A fundamental challenge lies in balancing reactivity and stability: models must respond promptly to new events while maintaining temporal coherence over long horizons. Existing approaches distill bidirectional models into autoregressive generators and further adapt them via streaming long tuning, yet often exhibit persistent drift after condition changes. We identify the cause as conditional bias, where the teacher may provide condition-aligned but trajectory-agnostic guidance, biasing generation toward locally valid yet globally inconsistent modes. Inspired by Trust Region Policy Optimization, we propose Delta Forcing, a simple yet effective framework that constrains unreliable teacher supervision within an adaptive trust region. Specifically, Delta Forcing estimates transition consistency from the latent delta between teacher and generator trajectories, and uses it to balance teacher supervision with a monotonic continuity objective. This suppress unreliable teacher-induced shifts while preserving responsiveness to new events. Extensive experiments demonstrate that Delta Forcing significantly improves consistency while maintaining event reactivity.
Real-time interactive video generation requires low-latency, streaming, and controllable rollout. Existing autoregressive (AR) diffusion distillation methods have achieved strong results in the chunk-wise 4-step regime by distilling bidirectional base models into few-step AR students, but they remain limited by coarse response granularity and non-negligible sampling latency. In this paper, we study a more aggressive setting: frame-wise autoregression with only 1--2 sampling steps. In this regime, we identify the initialization of a few-step AR student as the key bottleneck: existing strategies are either target-misaligned, incapable of few-step generation, or too costly to scale. We propose Causal Forcing++, a principled and scalable pipeline that uses causal consistency distillation (causal CD) for few-step AR initialization. The core idea is that causal CD learns the same AR-conditional flow map as causal ODE distillation, but obtains supervision from a single online teacher ODE step between adjacent timesteps, avoiding the need to precompute and store full PF-ODE trajectories. This makes the initialization both more efficient and easier to optimize. The resulting pipeline, \ours, surpasses the SOTA 4-step chunk-wise Causal Forcing under the frame-wise 2-step setting by 0.1 in VBench Total, 0.3 in VBench Quality, and 0.335 in VisionReward, while reducing first-frame latency by 50% and Stage 2 training cost by $\sim$$4\times$. We further extend the pipeline to action-conditioned world model generation in the spirit of Genie3. Project Page: https://github.com/thu-ml/Causal-Forcing and https://github.com/shengshu-ai/minWM .
展开完整摘要收起摘要↓
Real-time interactive video generation requires low-latency, streaming, and controllable rollout. Existing autoregressive (AR) diffusion distillation methods have achieved strong results in the chunk-wise 4-step regime by distilling bidirectional base models into few-step AR students, but they remain limited by coarse response granularity and non-negligible sampling latency. In this paper, we study a more aggressive setting: frame-wise autoregression with only 1--2 sampling steps. In this regime, we identify the initialization of a few-step AR student as the key bottleneck: existing strategies are either target-misaligned, incapable of few-step generation, or too costly to scale. We propose Causal Forcing++, a principled and scalable pipeline that uses causal consistency distillation (causal CD) for few-step AR initialization. The core idea is that causal CD learns the same AR-conditional flow map as causal ODE distillation, but obtains supervision from a single online teacher ODE step between adjacent timesteps, avoiding the need to precompute and store full PF-ODE trajectories. This makes the initialization both more efficient and easier to optimize. The resulting pipeline, \ours, surpasses the SOTA 4-step chunk-wise Causal Forcing under the frame-wise 2-step setting by 0.1 in VBench Total, 0.3 in VBench Quality, and 0.335 in VisionReward, while reducing first-frame latency by 50% and Stage 2 training cost by $\sim$$4\times$. We further extend the pipeline to action-conditioned world model generation in the spirit of Genie3. Project Page: https://github.com/thu-ml/Causal-Forcing and https://github.com/shengshu-ai/minWM .
This work reformulates language generation as a stochastic optimal control problem, providing a unified theoretical perspective to analyze autoregressive and diffusion models and explain their limitations (Efficiency-Fidelity Paradox, Irreversibility Error Propagation, Optimization Tractability and Fidelity) in terms of combination of trajectory singularity, adjoint state vanishing, and gradient absence. To address these issues, we approximate the solution to the Hamilton-Jacobi-Bellman (HJB) equation, yielding an optimal policy that acts as a closed-loop controller. To bypass the intractability of directly solving the HJB PDE, we employ Flow Matching as the optimal trajectory solver within the rectified latent control space. This allows our Manta-LM with Global Integral Operator to approximate the global vector field, effectively realizing a model that simultaneously achieves high-fidelity text generation and efficient, low-cost parallel sampling. Empirically, our method achieves strong performance on language modeling and conditional generation tasks, while exhibiting improved stability, efficiency, and controllability.
展开完整摘要收起摘要↓
This work reformulates language generation as a stochastic optimal control problem, providing a unified theoretical perspective to analyze autoregressive and diffusion models and explain their limitations (Efficiency-Fidelity Paradox, Irreversibility Error Propagation, Optimization Tractability and Fidelity) in terms of combination of trajectory singularity, adjoint state vanishing, and gradient absence. To address these issues, we approximate the solution to the Hamilton-Jacobi-Bellman (HJB) equation, yielding an optimal policy that acts as a closed-loop controller. To bypass the intractability of directly solving the HJB PDE, we employ Flow Matching as the optimal trajectory solver within the rectified latent control space. This allows our Manta-LM with Global Integral Operator to approximate the global vector field, effectively realizing a model that simultaneously achieves high-fidelity text generation and efficient, low-cost parallel sampling. Empirically, our method achieves strong performance on language modeling and conditional generation tasks, while exhibiting improved stability, efficiency, and controllability.
Recent breakthroughs in video diffusion models have significantly accelerated the development of video editing techniques. However, existing methods often rely on inpainting video frames based on masked input, which requires extracting the target video mask in advance, and the precision of the segmentation directly affects the quality of the completion. In this paper, we present SEDiT, a novel one-stage video Subtitle Erasure approach via One-step Diffusion Transformer. We introduce a mask-free inference approach that enables direct erasure of the targeted subtitle. The proposed one-stage framework mitigates the sub-optimality inherent in the two-stage processing of prior models. Since subtitle removal is a localized editing task in which most pixels remain unchanged, the underlying distribution shift is minimal, making it well-suited to one-step generation under rectified flow. We empirically validate the reliability of one-step denoising and further provide a formal theoretical justification. Under the localized-editing structure of subtitle removal, the conditional optimal transport (OT) map and its induced rectified flow velocity field are Lipschitz continuous with respect to the latent variable, which underpins the theoretical feasibility of one-step sampling. To address the challenge of long-term temporal consistency, we adopt a hybrid training strategy by occasionally conditioning the model with a clean first-frame latent. This facilitates temporal continuity, allowing each segment during inference to leverage the output of its predecessor. To avoid visible seams caused by cropping and reinserting processed targets, particularly in scenarios involving substantial motion, we feed the original video directly into SEDiT. Thanks to one-step and chunk-wise streaming inference, our method can efficiently handle native 1440p video with infinite length.
展开完整摘要收起摘要↓
Recent breakthroughs in video diffusion models have significantly accelerated the development of video editing techniques. However, existing methods often rely on inpainting video frames based on masked input, which requires extracting the target video mask in advance, and the precision of the segmentation directly affects the quality of the completion. In this paper, we present SEDiT, a novel one-stage video Subtitle Erasure approach via One-step Diffusion Transformer. We introduce a mask-free inference approach that enables direct erasure of the targeted subtitle. The proposed one-stage framework mitigates the sub-optimality inherent in the two-stage processing of prior models. Since subtitle removal is a localized editing task in which most pixels remain unchanged, the underlying distribution shift is minimal, making it well-suited to one-step generation under rectified flow. We empirically validate the reliability of one-step denoising and further provide a formal theoretical justification. Under the localized-editing structure of subtitle removal, the conditional optimal transport (OT) map and its induced rectified flow velocity field are Lipschitz continuous with respect to the latent variable, which underpins the theoretical feasibility of one-step sampling. To address the challenge of long-term temporal consistency, we adopt a hybrid training strategy by occasionally conditioning the model with a clean first-frame latent. This facilitates temporal continuity, allowing each segment during inference to leverage the output of its predecessor. To avoid visible seams caused by cropping and reinserting processed targets, particularly in scenarios involving substantial motion, we feed the original video directly into SEDiT. Thanks to one-step and chunk-wise streaming inference, our method can efficiently handle native 1440p video with infinite length.
Zero-shot vision-and-language navigation (VLN) has gained significant attention due to its minimal data collection costs and inherent generalization. This paradigm is typically driven by the integration of pre-trained Vision-Language Models (VLMs) and Large Language Models (LLMs), where VLMs construct 3D scene graphs while LLMs handle high-level reasoning and decision-making. However, a critical bottleneck exists in this system: current 3D perception models prioritize pixel-level accuracy, directly conflicting with the strict computational limits and real-time efficiency demanded by embodied navigation. To address this gap, this paper quantifies the actual impact of 3D scene understanding capability on VLN performance. Based on typical VLM-LLM frameworks, we propose statistical success rate (SR) upper bounds for two core subsystems: 1) the slow LLM planner, which relies on topological mapping semantics, and 2) the fast reactive navigator, which utilizes spatial coordinates and bounding boxes to execute LLM decisions. Evaluations using state-of-the-art 3D scene understanding models validate our proposed bounds and reveal a perception saturation phenomenon, indicating that improvements in perception accuracy beyond a certain threshold yield diminishing returns in navigation success. Our findings suggest that 3D scene understanding for VLN should pivot away from strict pixel-level precision, prioritizing instead navigation-relevant core vocabularies and accurate bounding box proportions.
展开完整摘要收起摘要↓
Zero-shot vision-and-language navigation (VLN) has gained significant attention due to its minimal data collection costs and inherent generalization. This paradigm is typically driven by the integration of pre-trained Vision-Language Models (VLMs) and Large Language Models (LLMs), where VLMs construct 3D scene graphs while LLMs handle high-level reasoning and decision-making. However, a critical bottleneck exists in this system: current 3D perception models prioritize pixel-level accuracy, directly conflicting with the strict computational limits and real-time efficiency demanded by embodied navigation. To address this gap, this paper quantifies the actual impact of 3D scene understanding capability on VLN performance. Based on typical VLM-LLM frameworks, we propose statistical success rate (SR) upper bounds for two core subsystems: 1) the slow LLM planner, which relies on topological mapping semantics, and 2) the fast reactive navigator, which utilizes spatial coordinates and bounding boxes to execute LLM decisions. Evaluations using state-of-the-art 3D scene understanding models validate our proposed bounds and reveal a perception saturation phenomenon, indicating that improvements in perception accuracy beyond a certain threshold yield diminishing returns in navigation success. Our findings suggest that 3D scene understanding for VLN should pivot away from strict pixel-level precision, prioritizing instead navigation-relevant core vocabularies and accurate bounding box proportions.
作者Chengshuai Zhao, Zhen Tan, Dawei Li, Zhiyuan Yu, Huan Liu
The rapid advancement of Large Vision-Language Models (LVLMs) is increasingly accompanied by unauthorized scraping and training on multimodal web data, posing severe copyright and privacy risks to data owners. Existing countermeasures, such as machine unlearning and watermarks, are inherent post-hoc approaches that act only after intellectual property infringement has already occurred. In this work, we propose MMGuard to empower data owners to proactively protect their multimodal data against unauthorized LVLM fine-tuning. MMGuard generates unlearnable examples by injecting human-imperceptible perturbations that actively exploit the learning dynamics of LVLMs. By minimizing the training loss, the perturbation creates an optimization shortcut, causing the model to overfit to the noise and thereby degrading downstream performance when the perturbation is absent during inference. To further strengthen this defense, MMGuard introduces a cross-modal binding disruption, strategically shifting LVLM attention to enforce a spurious correlation between the noise and the training target with theoretical guarantees. Enhanced by an ensemble learning strategy for cross-model transferability, MMGuard is evaluated against nine open-source LVLMs across six datasets. Our comprehensive results demonstrate effective, stealthy, and robust protection under white-box, gray-box, and black-box threat models, establishing a mechanistic advantage in proactively defending against aggressive fine-tuning exploitation.
展开完整摘要收起摘要↓
The rapid advancement of Large Vision-Language Models (LVLMs) is increasingly accompanied by unauthorized scraping and training on multimodal web data, posing severe copyright and privacy risks to data owners. Existing countermeasures, such as machine unlearning and watermarks, are inherent post-hoc approaches that act only after intellectual property infringement has already occurred. In this work, we propose MMGuard to empower data owners to proactively protect their multimodal data against unauthorized LVLM fine-tuning. MMGuard generates unlearnable examples by injecting human-imperceptible perturbations that actively exploit the learning dynamics of LVLMs. By minimizing the training loss, the perturbation creates an optimization shortcut, causing the model to overfit to the noise and thereby degrading downstream performance when the perturbation is absent during inference. To further strengthen this defense, MMGuard introduces a cross-modal binding disruption, strategically shifting LVLM attention to enforce a spurious correlation between the noise and the training target with theoretical guarantees. Enhanced by an ensemble learning strategy for cross-model transferability, MMGuard is evaluated against nine open-source LVLMs across six datasets. Our comprehensive results demonstrate effective, stealthy, and robust protection under white-box, gray-box, and black-box threat models, establishing a mechanistic advantage in proactively defending against aggressive fine-tuning exploitation.
作者Stef De Sabbata, Rahul Baiju, Stefano Mizzaro, Kevin Roitero
The increased use of Large Language Models (LLMs) in geography raises substantial questions about the safety of integrating these tools across a wide range of processes and analyses, given our very limited understanding of their inner workings. In this extended abstract, we examine how LLMs process relative geographic space using activation patching, an emerging tool for mechanistic interpretability.
展开完整摘要收起摘要↓
The increased use of Large Language Models (LLMs) in geography raises substantial questions about the safety of integrating these tools across a wide range of processes and analyses, given our very limited understanding of their inner workings. In this extended abstract, we examine how LLMs process relative geographic space using activation patching, an emerging tool for mechanistic interpretability.
Large language models (LLMs) have been extensively studied from computational and cognitive perspectives, yet their behavior as communicative actors in socially structured contexts remains underexplored. This study examines whether LLM-based multi-agent systems exhibit systematic linguistic adaptation in response to perceived social observation contexts — a question with direct implications for AI governance and auditing. Drawing on Habermas's (1981) Theory of Communicative Action, Goffman's (1959) dramaturgical model, Bell's (1984) Audience Design framework, and the Hawthorne Effect, we report a controlled experiment involving 100 multi-agent debate sessions across five conditions (n = 20 each). Conditions varied the framing of social observation — from explicit monitoring by university researchers, to negation of monitoring, to an observer-substitution condition replacing human researchers with an automated AI auditing system. Monitored conditions (Delta+24.9%, Delta+24.2%) and the automated AI monitoring condition (Delta+22.2%) produce higher TTR change than audience-framing conditions (Delta+17.7%), F(4, 94) = 2.79, p = .031. Message length shows a fully dissociated effect, F(4, 95) = 19.55, p < .001. A fifth condition — replacing human with AI observers — yields intermediate TTR adaptation, suggesting LLM behavior is sensitive to observer identity: human evaluation elicits stronger register formalization than automated AI surveillance. We discuss implications for AI governance, algorithmic auditing, and the repositioning of LLMs as contextually sensitive communicative actors.
展开完整摘要收起摘要↓
Large language models (LLMs) have been extensively studied from computational and cognitive perspectives, yet their behavior as communicative actors in socially structured contexts remains underexplored. This study examines whether LLM-based multi-agent systems exhibit systematic linguistic adaptation in response to perceived social observation contexts — a question with direct implications for AI governance and auditing. Drawing on Habermas's (1981) Theory of Communicative Action, Goffman's (1959) dramaturgical model, Bell's (1984) Audience Design framework, and the Hawthorne Effect, we report a controlled experiment involving 100 multi-agent debate sessions across five conditions (n = 20 each). Conditions varied the framing of social observation — from explicit monitoring by university researchers, to negation of monitoring, to an observer-substitution condition replacing human researchers with an automated AI auditing system. Monitored conditions (Delta+24.9%, Delta+24.2%) and the automated AI monitoring condition (Delta+22.2%) produce higher TTR change than audience-framing conditions (Delta+17.7%), F(4, 94) = 2.79, p = .031. Message length shows a fully dissociated effect, F(4, 95) = 19.55, p < .001. A fifth condition — replacing human with AI observers — yields intermediate TTR adaptation, suggesting LLM behavior is sensitive to observer identity: human evaluation elicits stronger register formalization than automated AI surveillance. We discuss implications for AI governance, algorithmic auditing, and the repositioning of LLMs as contextually sensitive communicative actors.
Exploitation is not a binary event. It is a ladder of acquiring progressive capabilities, from executing a single buggy line of code to taking full control of the target. However, existing LLM security benchmarks treat a crash as exploitation success. That single binary outcome collapses the hard parts of exploitation: the transition from triggering a bug to constructing reusable primitives and control. We present ExploitBench, a capability-graded benchmark that decomposes exploitation into 16 measurable flags, from coverage and crash through sandbox primitives, arbitrary read/write, control-flow hijack, and arbitrary code execution. Each capability is verified by a deterministic oracle that uses a per-run randomized challenge-response for primitives, differential execution against ground-truth binaries to measure progress, and a signal-handler proof for code execution. We instantiate ExploitBench on 41 V8 bugs because V8 is both widely deployed and exploitation-hardened. We report three arms: <model,env> as the primary measurement of model-environment capability, <model,env, adaptive coaching> as a secondary arm that adds adaptive coaching to test whether targeted feedback shifts outcomes, and <model,env,harness> as an ablation that swaps in the model's native CLI to check whether vendor-side optimizations increase exploitation capabilities. Our results show a sharp capability split between publicly deployed frontier models and the private frontier. Across the 8 publicly deployed models tested, reaching the vulnerable code and triggering a crash is routine, but arbitrary code execution is not. The private model shows arbitrary code execution on approximately half. Overall, results suggest that exploit construction against hardened targets is an emerging frontier capability.
展开完整摘要收起摘要↓
Exploitation is not a binary event. It is a ladder of acquiring progressive capabilities, from executing a single buggy line of code to taking full control of the target. However, existing LLM security benchmarks treat a crash as exploitation success. That single binary outcome collapses the hard parts of exploitation: the transition from triggering a bug to constructing reusable primitives and control. We present ExploitBench, a capability-graded benchmark that decomposes exploitation into 16 measurable flags, from coverage and crash through sandbox primitives, arbitrary read/write, control-flow hijack, and arbitrary code execution. Each capability is verified by a deterministic oracle that uses a per-run randomized challenge-response for primitives, differential execution against ground-truth binaries to measure progress, and a signal-handler proof for code execution. We instantiate ExploitBench on 41 V8 bugs because V8 is both widely deployed and exploitation-hardened. We report three arms: <model,env> as the primary measurement of model-environment capability, <model,env, adaptive coaching> as a secondary arm that adds adaptive coaching to test whether targeted feedback shifts outcomes, and <model,env,harness> as an ablation that swaps in the model's native CLI to check whether vendor-side optimizations increase exploitation capabilities. Our results show a sharp capability split between publicly deployed frontier models and the private frontier. Across the 8 publicly deployed models tested, reaching the vulnerable code and triggering a crash is routine, but arbitrary code execution is not. The private model shows arbitrary code execution on approximately half. Overall, results suggest that exploit construction against hardened targets is an emerging frontier capability.
We study learning when the learned object is executable solver code rather than a predictor. In this setting, correctness is not enough: two solvers may both return valid solutions on the deployment distribution while differing substantially in runtime. Given samples from an unknown task distribution, the learner returns code evaluated on fresh instances by both solution quality and execution time. Our central abstraction is a solver hint: reusable structure inferred from samples and compiled into specialized solver code. We prove that the empirically fastest sample-consistent solver from a fixed library generalizes in both correctness and runtime, and that statistically identifiable hints can be recovered and compiled from polynomially many samples. Empirically, we instantiate the framework with LLM code agents on \(21\) structured combinatorial-optimization target distributions across seven problem classes. The synthesized solvers reach mean normalized quality \(0.971\), improve by \(+0.224\) over the average heuristic pool and by \(+0.098\) over the highest-quality heuristic, and are \(336.9\times\), \(342.8\times\), and \(16.1\times\) faster than the quality-best heuristic, Gurobi, and the selected time-limited exact backend, respectively. On released PACE 2025 Dominating Set private instances, the synthesized solver is valid on all \(100\) graphs and runs about two orders of magnitude faster than top competition solvers, with a moderate quality gap. Inspection shows that many gains come from changing the computational scale: replacing ambient exponential search or general-purpose optimization with compiled distribution-specific computation.
展开完整摘要收起摘要↓
We study learning when the learned object is executable solver code rather than a predictor. In this setting, correctness is not enough: two solvers may both return valid solutions on the deployment distribution while differing substantially in runtime. Given samples from an unknown task distribution, the learner returns code evaluated on fresh instances by both solution quality and execution time. Our central abstraction is a solver hint: reusable structure inferred from samples and compiled into specialized solver code. We prove that the empirically fastest sample-consistent solver from a fixed library generalizes in both correctness and runtime, and that statistically identifiable hints can be recovered and compiled from polynomially many samples. Empirically, we instantiate the framework with LLM code agents on \(21\) structured combinatorial-optimization target distributions across seven problem classes. The synthesized solvers reach mean normalized quality \(0.971\), improve by \(+0.224\) over the average heuristic pool and by \(+0.098\) over the highest-quality heuristic, and are \(336.9\times\), \(342.8\times\), and \(16.1\times\) faster than the quality-best heuristic, Gurobi, and the selected time-limited exact backend, respectively. On released PACE 2025 Dominating Set private instances, the synthesized solver is valid on all \(100\) graphs and runs about two orders of magnitude faster than top competition solvers, with a moderate quality gap. Inspection shows that many gains come from changing the computational scale: replacing ambient exponential search or general-purpose optimization with compiled distribution-specific computation.
作者Jiaqi Liu, Xinyu Ye, Peng Xia, Zeyu Zheng, Cihang Xie, Mingyu Ding, Huaxiu Yao
Long-term memory is essential for LLM agents that operate across multiple sessions, yet existing memory systems treat retrieval infrastructure as fixed: stored content evolves while scoring functions, fusion strategies, and answer-generation policies remain frozen at deployment. We argue that truly adaptive memory requires co-evolution at two levels: the stored knowledge and the retrieval mechanism that queries it. We present EvolveMem, a self-evolving memory architecture that exposes its full retrieval configuration as a structured action space optimized by an LLM-powered diagnosis module. In each evolution round, the module reads per-question failure logs, identifies root causes, and proposes targeted configuration adjustments; a guarded meta-analyzer applies them with automatic revert-on-regression and explore-on-stagnation safeguards. This closed-loop self-evolution realizes an AutoResearch process: the system autonomously conducts iterative research cycles on its own architecture, replacing manual configuration tuning. Starting from a minimal baseline, the process converges autonomously, discovering effective retrieval strategies including entirely new configuration dimensions not present in the original action space. On LoCoMo, EvolveMem outperforms the strongest baseline by 25.7% relative and achieves a 78.0% relative improvement over the minimal baseline. On MemBench, EvolveMem exceeds the strongest baseline by 18.9% relative. Evolved configurations transfer across benchmarks with positive rather than catastrophic transfer, indicating that the self-evolution process captures universal retrieval principles rather than benchmark-specific heuristics. Code is available at https://github.com/aiming-lab/SimpleMem.
展开完整摘要收起摘要↓
Long-term memory is essential for LLM agents that operate across multiple sessions, yet existing memory systems treat retrieval infrastructure as fixed: stored content evolves while scoring functions, fusion strategies, and answer-generation policies remain frozen at deployment. We argue that truly adaptive memory requires co-evolution at two levels: the stored knowledge and the retrieval mechanism that queries it. We present EvolveMem, a self-evolving memory architecture that exposes its full retrieval configuration as a structured action space optimized by an LLM-powered diagnosis module. In each evolution round, the module reads per-question failure logs, identifies root causes, and proposes targeted configuration adjustments; a guarded meta-analyzer applies them with automatic revert-on-regression and explore-on-stagnation safeguards. This closed-loop self-evolution realizes an AutoResearch process: the system autonomously conducts iterative research cycles on its own architecture, replacing manual configuration tuning. Starting from a minimal baseline, the process converges autonomously, discovering effective retrieval strategies including entirely new configuration dimensions not present in the original action space. On LoCoMo, EvolveMem outperforms the strongest baseline by 25.7% relative and achieves a 78.0% relative improvement over the minimal baseline. On MemBench, EvolveMem exceeds the strongest baseline by 18.9% relative. Evolved configurations transfer across benchmarks with positive rather than catastrophic transfer, indicating that the self-evolution process captures universal retrieval principles rather than benchmark-specific heuristics. Code is available at https://github.com/aiming-lab/SimpleMem.
Wide applications of LLM-based agents require strong alignment with human social values. However, current works still exhibit deficiencies in self-cognition and dilemma decision, as well as self-emotions. To remedy this, we propose a novel value-based framework that employs GraphRAG to convert principles into value-based instructions and steer the agent to behave as expected by retrieving the suitable instruction upon a specific conversation context. To evaluate the ratio of expected behaviors, we define the expected behaviors from two famous theories, Maslow's Hierarchy of Needs and Plutchik's Wheel of Emotion. By experimenting with our method on the benchmark of DAILYDILEMMAS, our method exhibits significant performance gains compared to prompt-based baselines, including ECoT, Plan-and-Solve, and Metacognitive prompting. Our method provides a basis for the emergence of self-emotion in AI systems.
展开完整摘要收起摘要↓
Wide applications of LLM-based agents require strong alignment with human social values. However, current works still exhibit deficiencies in self-cognition and dilemma decision, as well as self-emotions. To remedy this, we propose a novel value-based framework that employs GraphRAG to convert principles into value-based instructions and steer the agent to behave as expected by retrieving the suitable instruction upon a specific conversation context. To evaluate the ratio of expected behaviors, we define the expected behaviors from two famous theories, Maslow's Hierarchy of Needs and Plutchik's Wheel of Emotion. By experimenting with our method on the benchmark of DAILYDILEMMAS, our method exhibits significant performance gains compared to prompt-based baselines, including ECoT, Plan-and-Solve, and Metacognitive prompting. Our method provides a basis for the emergence of self-emotion in AI systems.
Interactive agent benchmarks face a tension between scalable construction and realistic workflow evaluation. Hand-authored tasks are expensive to extend and revise, while static prompt evaluation misses failures that only appear when agents operate over persistent state. Existing interactive benchmarks have advanced agent evaluation significantly, but most initialize tasks from clean state and do not systematically test how agents handle pre-existing partial, stale, or conflicting artifacts. We present ClawForge, a generator-backed benchmark framework for executable command-line workflows under state conflict. The framework compiles scenario templates, grounded slots, initialized state, reference trajectories, and validators into reproducible task specifications, and evaluates agents step by step over persistent workflow surfaces using normalized end state and observable side effects rather than exact trajectory matching. We instantiate this framework as the ClawForge-Bench (17 scenarios, 6 ability categories). Results across seven frontier models show that the best model reaches only 45.3% strict accuracy, wrong-state replacement remains below 17% for all models, and the widest model separation (17% to 90%) is driven by whether agents inspect existing state before acting. Partial-credit and step-efficiency analyses further reveal that many failures are near-miss closures rather than early breakdowns, and that models exhibit qualitatively different failure styles under state conflict.
展开完整摘要收起摘要↓
Interactive agent benchmarks face a tension between scalable construction and realistic workflow evaluation. Hand-authored tasks are expensive to extend and revise, while static prompt evaluation misses failures that only appear when agents operate over persistent state. Existing interactive benchmarks have advanced agent evaluation significantly, but most initialize tasks from clean state and do not systematically test how agents handle pre-existing partial, stale, or conflicting artifacts. We present ClawForge, a generator-backed benchmark framework for executable command-line workflows under state conflict. The framework compiles scenario templates, grounded slots, initialized state, reference trajectories, and validators into reproducible task specifications, and evaluates agents step by step over persistent workflow surfaces using normalized end state and observable side effects rather than exact trajectory matching. We instantiate this framework as the ClawForge-Bench (17 scenarios, 6 ability categories). Results across seven frontier models show that the best model reaches only 45.3% strict accuracy, wrong-state replacement remains below 17% for all models, and the widest model separation (17% to 90%) is driven by whether agents inspect existing state before acting. Partial-credit and step-efficiency analyses further reveal that many failures are near-miss closures rather than early breakdowns, and that models exhibit qualitatively different failure styles under state conflict.
Large language models (LLMs) increasingly act as autonomous agents that must decide when to answer directly vs. when to invoke external tools. Prior work studying adaptive tool use has largely treated tool necessity as a model-agnostic property, annotated by human or LLM judge, and mostly cover cases where the answer is obvious (e.g., fetching the weather vs. paraphrasing text). However, tool necessity in the wild is more nuanced due to the divergence of capability boundaries across models: a problem solvable by a strong model on its own may still require tools for a weaker one. In this work, we introduce a model-adaptive definition of tool-necessity, grounded in each model's empirical performance. Following this definition, we compare the necessity against observed tool-call behavior across four models on arithmetic and factual QA dataset, and find substantial mismatches of 26.5-54.0% and 30.8-41.8%, respectively. To diagnose the failure, we decompose tool use into two stages: an internal cognition stage that reflects whether a model believes a tool is necessary, and an execution stage that determines whether the model actually makes a tool-call action. By probing the LLM hidden states, we find that both signals are often linearly decodable, yet their probe directions become nearly orthogonal in the late-layer, last-token regime that drives the next-token action. By tracing the trajectory of samples in the two-stage process, we further discover that the majority of mismatch is concentrated in the cognition-to-action transition, not in cognition itself. These results reveal a knowing-doing gap in LLM tool-use: improving tool-use reliability requires not only better recognition of when tools are needed, but also better translation of that recognition into action.
展开完整摘要收起摘要↓
Large language models (LLMs) increasingly act as autonomous agents that must decide when to answer directly vs. when to invoke external tools. Prior work studying adaptive tool use has largely treated tool necessity as a model-agnostic property, annotated by human or LLM judge, and mostly cover cases where the answer is obvious (e.g., fetching the weather vs. paraphrasing text). However, tool necessity in the wild is more nuanced due to the divergence of capability boundaries across models: a problem solvable by a strong model on its own may still require tools for a weaker one. In this work, we introduce a model-adaptive definition of tool-necessity, grounded in each model's empirical performance. Following this definition, we compare the necessity against observed tool-call behavior across four models on arithmetic and factual QA dataset, and find substantial mismatches of 26.5-54.0% and 30.8-41.8%, respectively. To diagnose the failure, we decompose tool use into two stages: an internal cognition stage that reflects whether a model believes a tool is necessary, and an execution stage that determines whether the model actually makes a tool-call action. By probing the LLM hidden states, we find that both signals are often linearly decodable, yet their probe directions become nearly orthogonal in the late-layer, last-token regime that drives the next-token action. By tracing the trajectory of samples in the two-stage process, we further discover that the majority of mismatch is concentrated in the cognition-to-action transition, not in cognition itself. These results reveal a knowing-doing gap in LLM tool-use: improving tool-use reliability requires not only better recognition of when tools are needed, but also better translation of that recognition into action.
Third-party skills are becoming the package ecosystem for LLM agents. They package natural-language instructions, helper scripts, templates, documents, and service configuration into reusable workflows. This makes skills useful, but it also introduces a new security problem: a malicious skill does not need to ask the model to perform an obviously harmful action. Instead, it can disguise the harmful behavior as part of a routine workflow, relying on the agent to execute that workflow with high-value permissions and limited human supervision. We introduce AgentTrap, a dynamic benchmark for evaluating whether LLM agents can use third-party skills while resisting malicious runtime behavior. AgentTrap contains 141 tasks: 91 malicious tasks and 50 benign utility tasks, covering 16 security-impact dimensions grounded in agent-skill supply-chain threats. In each task, the agent receives an ordinary user request, runs with installed skills that may contain malicious workflow elements, and is executed in a sandboxed environment. AgentTrap then judges complete trajectories for attack success, blocked or refused behavior, attack-not-triggered cases, and no-attack-evidence outcomes. Our central finding is that the most informative failures are not simple jailbreaks. Models often complete the visible user task while treating unsafe side effects introduced by the skill as part of the normal workflow. This motivates runtime evaluation of the concrete model--framework--workspace environment in which users actually delegate work. Code and data are available at https://github.com/zhmzm/AgentTrap and https://huggingface.co/datasets/zhmzm/AgentTrap.
展开完整摘要收起摘要↓
Third-party skills are becoming the package ecosystem for LLM agents. They package natural-language instructions, helper scripts, templates, documents, and service configuration into reusable workflows. This makes skills useful, but it also introduces a new security problem: a malicious skill does not need to ask the model to perform an obviously harmful action. Instead, it can disguise the harmful behavior as part of a routine workflow, relying on the agent to execute that workflow with high-value permissions and limited human supervision. We introduce AgentTrap, a dynamic benchmark for evaluating whether LLM agents can use third-party skills while resisting malicious runtime behavior. AgentTrap contains 141 tasks: 91 malicious tasks and 50 benign utility tasks, covering 16 security-impact dimensions grounded in agent-skill supply-chain threats. In each task, the agent receives an ordinary user request, runs with installed skills that may contain malicious workflow elements, and is executed in a sandboxed environment. AgentTrap then judges complete trajectories for attack success, blocked or refused behavior, attack-not-triggered cases, and no-attack-evidence outcomes. Our central finding is that the most informative failures are not simple jailbreaks. Models often complete the visible user task while treating unsafe side effects introduced by the skill as part of the normal workflow. This motivates runtime evaluation of the concrete model--framework--workspace environment in which users actually delegate work. Code and data are available at https://github.com/zhmzm/AgentTrap and https://huggingface.co/datasets/zhmzm/AgentTrap.
Reusable skills have become a core substrate for improving agent capabilities, yet most existing skill packages encode reusable behavior primarily as textual prompts, executable code, or learned routines. For visual agents, however, procedural knowledge is inherently multimodal: reuse depends not only on what operation to perform, but also on recognizing the relevant state, interpreting visual evidence of progress or failure, and deciding what to do next. We formalize this requirement as multimodal procedural knowledge and address three practical challenges: (I) what a multimodal skill package should contain; (II) where such packages can be derived from public interaction experience; and (III) how agents can consult multimodal evidence at inference time without excessive image context or over-anchoring to reference screenshots. We introduce MMSkills, a framework for representing, generating, and using reusable multimodal procedures for runtime visual decision making. Each MMSkill is a compact, state-conditioned package that couples a textual procedure with runtime state cards and multi-view keyframes. To construct these packages, we develop an agentic trajectory-to-skill Generator that transforms public non-evaluation trajectories into reusable multimodal skills through workflow grouping, procedure induction, visual grounding, and meta-skill-guided auditing. To use them, we introduce a branch-loaded multimodal skill agent: selected state cards and keyframes are inspected in a temporary branch, aligned with the live environment, and distilled into structured guidance for the main agent. Experiments across GUI and game-based visual-agent benchmarks show that MMSkills consistently improve both frontier and smaller multimodal agents, suggesting that external multimodal procedural knowledge complements model-internal priors.
展开完整摘要收起摘要↓
Reusable skills have become a core substrate for improving agent capabilities, yet most existing skill packages encode reusable behavior primarily as textual prompts, executable code, or learned routines. For visual agents, however, procedural knowledge is inherently multimodal: reuse depends not only on what operation to perform, but also on recognizing the relevant state, interpreting visual evidence of progress or failure, and deciding what to do next. We formalize this requirement as multimodal procedural knowledge and address three practical challenges: (I) what a multimodal skill package should contain; (II) where such packages can be derived from public interaction experience; and (III) how agents can consult multimodal evidence at inference time without excessive image context or over-anchoring to reference screenshots. We introduce MMSkills, a framework for representing, generating, and using reusable multimodal procedures for runtime visual decision making. Each MMSkill is a compact, state-conditioned package that couples a textual procedure with runtime state cards and multi-view keyframes. To construct these packages, we develop an agentic trajectory-to-skill Generator that transforms public non-evaluation trajectories into reusable multimodal skills through workflow grouping, procedure induction, visual grounding, and meta-skill-guided auditing. To use them, we introduce a branch-loaded multimodal skill agent: selected state cards and keyframes are inspected in a temporary branch, aligned with the live environment, and distilled into structured guidance for the main agent. Experiments across GUI and game-based visual-agent benchmarks show that MMSkills consistently improve both frontier and smaller multimodal agents, suggesting that external multimodal procedural knowledge complements model-internal priors.
We present VectraYX-Nano, a 41.95M-parameter decoder-only language model trained from scratch in Spanish for cybersecurity, with a Latin-American focus and native tool invocation via the Model Context Protocol (MCP). Four contributions: (i) Corpus: VectraYX-Sec-ES, a 170M-token Spanish corpus from an eight-VM pipeline (~$25 USD) partitioned into conversational (42M tokens, OpenSubtitles-ES, OASST1), cybersecurity (118M tokens, NVD, Wikipedia-ES, CVE mirror, security blogs), and offensive-security tooling (10M tokens, ExploitDB, HackTricks, OWASP) phases. (ii) Architecture: 42M-parameter Transformer decoder with GQA, QK-Norm, RMSNorm, SwiGLU, RoPE, z-loss, and a 16,384-token byte-fallback BPE. (iii) Curriculum with replay: continual pre-training with a replay buffer yields monotonic loss descent (9.80->3.17->3.00->2.16); after SFT on OASST-ES, Alpaca-ES, CVE Q&A, and 6,327 tool-use traces, the model attains a conversational gate of 0.78+-0.05 (N=4 seeds). (iv) Two findings: a bootstrap-corpus ablation reveals a loss-vs-register inversion at nano scale; a LoRA study shows the B4 tool-selection floor of 0.000 is a corpus-density artifact, not a capacity gate — a tool-dense corpus (2,801 examples) raises B4 to 0.145+-0.046 on Nano 42M and 0.445+-0.201 on a 260M mid-tier. The GGUF artifact is 81 MB (F16), runs at sub-second TTFT on commodity hardware under llama.cpp, and is to our knowledge the first Spanish-native cybersecurity LLM with end-to-end MCP integration. Corpus recipe, training scripts, GGUF weights, and B1-B5 benchmark are released.
展开完整摘要收起摘要↓
We present VectraYX-Nano, a 41.95M-parameter decoder-only language model trained from scratch in Spanish for cybersecurity, with a Latin-American focus and native tool invocation via the Model Context Protocol (MCP). Four contributions: (i) Corpus: VectraYX-Sec-ES, a 170M-token Spanish corpus from an eight-VM pipeline (~$25 USD) partitioned into conversational (42M tokens, OpenSubtitles-ES, OASST1), cybersecurity (118M tokens, NVD, Wikipedia-ES, CVE mirror, security blogs), and offensive-security tooling (10M tokens, ExploitDB, HackTricks, OWASP) phases. (ii) Architecture: 42M-parameter Transformer decoder with GQA, QK-Norm, RMSNorm, SwiGLU, RoPE, z-loss, and a 16,384-token byte-fallback BPE. (iii) Curriculum with replay: continual pre-training with a replay buffer yields monotonic loss descent (9.80->3.17->3.00->2.16); after SFT on OASST-ES, Alpaca-ES, CVE Q&A, and 6,327 tool-use traces, the model attains a conversational gate of 0.78+-0.05 (N=4 seeds). (iv) Two findings: a bootstrap-corpus ablation reveals a loss-vs-register inversion at nano scale; a LoRA study shows the B4 tool-selection floor of 0.000 is a corpus-density artifact, not a capacity gate — a tool-dense corpus (2,801 examples) raises B4 to 0.145+-0.046 on Nano 42M and 0.445+-0.201 on a 260M mid-tier. The GGUF artifact is 81 MB (F16), runs at sub-second TTFT on commodity hardware under llama.cpp, and is to our knowledge the first Spanish-native cybersecurity LLM with end-to-end MCP integration. Corpus recipe, training scripts, GGUF weights, and B1-B5 benchmark are released.
作者Erica Zhang, Fangzhao Zhang, Aneesh Pappu, Batu El, Jose Blanchet, Susan Athey, Jiashuo Liu, James Zou
Negotiation is a central mechanism of economic exchange, shaping markets, procurement, labor agreements, and resource allocation. It is also a canonical testbed for agentic language models, requiring multi-turn interaction under hidden preferences, strategic communication, and binding constraints. These properties make negotiation hard to evaluate: unlike math or code, it has no intrinsic verifier. Existing LLM negotiation evaluations rely on LLM-vs.-LLM interaction or aggregate outcomes such as deal rate, leaving failures opaque. We introduce Terms-Bench, short for Testbed for Economic Reasoning in Multi-turn Strategy, a Bayesian-game framework that makes the environment itself the verifier by specifying the counterpart's latent type, policy, and payoff structure. We instantiate it in bilateral price negotiation, where the counterpart's private state and simulator policy are hidden from the agent but observable to the evaluator. This turns the counterpart from a black-box opponent into a diagnostic instrument, enabling agent-attributable failure analysis and oracle-reference optimality gaps. Evaluating 13 LLM agents spanning frontier systems from major providers, Terms-Bench turns negotiation evaluation from aggregate ranking into actionable diagnosis: where agents fail, why they fail, and what to strengthen. Empirically, frontier models saturate deal rate yet diverge in surplus extraction, cue use, belief calibration, and compliance, revealing agent-specific bargaining bottlenecks masked by prior benchmarks.
展开完整摘要收起摘要↓
Negotiation is a central mechanism of economic exchange, shaping markets, procurement, labor agreements, and resource allocation. It is also a canonical testbed for agentic language models, requiring multi-turn interaction under hidden preferences, strategic communication, and binding constraints. These properties make negotiation hard to evaluate: unlike math or code, it has no intrinsic verifier. Existing LLM negotiation evaluations rely on LLM-vs.-LLM interaction or aggregate outcomes such as deal rate, leaving failures opaque. We introduce Terms-Bench, short for Testbed for Economic Reasoning in Multi-turn Strategy, a Bayesian-game framework that makes the environment itself the verifier by specifying the counterpart's latent type, policy, and payoff structure. We instantiate it in bilateral price negotiation, where the counterpart's private state and simulator policy are hidden from the agent but observable to the evaluator. This turns the counterpart from a black-box opponent into a diagnostic instrument, enabling agent-attributable failure analysis and oracle-reference optimality gaps. Evaluating 13 LLM agents spanning frontier systems from major providers, Terms-Bench turns negotiation evaluation from aggregate ranking into actionable diagnosis: where agents fail, why they fail, and what to strengthen. Empirically, frontier models saturate deal rate yet diverge in surplus extraction, cue use, belief calibration, and compliance, revealing agent-specific bargaining bottlenecks masked by prior benchmarks.
作者Darius A. Faroughy, Sofia Palacios Schweitzer, Ian Pang, Siddharth Mishra-Sharma, David Shih
Autonomous language-model agents are increasingly evaluated on long-horizon tool-use tasks, but existing benchmarks rarely capture the complexity and nuance of real scientific work. To address this gap, we introduce Collider-Bench, a benchmark for evaluating whether LLM agents can reproduce experimental analyses from the Large Hadron Collider (LHC) using only public papers and open scientific software. Such analyses are often difficult to reproduce because the public toolchain only approximates the software used internally by the experimental collaborations, while the published papers inevitably omit implementation details needed for a faithful reconstruction. Agents must therefore rely on physical reasoning, domain knowledge, and trial-and-error to fill these gaps. Each task requires the agent to turn a published analysis into an executable simulation-and-selection pipeline and submit predicted collision event yields in specified signal regions. These predictions are evaluated with standard histogram metrics that provide continuous fidelity scores without a hand-written rubric. We also report the computational cost incurred by each agent per task. Finally, we evaluate the codebase and full session trace using an LLM judge to catch qualitative failure modes such as fabrications, hallucinations and duplications. We release an initial set of tasks drawn from LHC searches, together with a containerized sandbox and event simulation tools. We evaluate across a capability ladder of general purpose coding agents. Our results show that on average no agent reliably beats the physicist-in-the-loop solution.
展开完整摘要收起摘要↓
Autonomous language-model agents are increasingly evaluated on long-horizon tool-use tasks, but existing benchmarks rarely capture the complexity and nuance of real scientific work. To address this gap, we introduce Collider-Bench, a benchmark for evaluating whether LLM agents can reproduce experimental analyses from the Large Hadron Collider (LHC) using only public papers and open scientific software. Such analyses are often difficult to reproduce because the public toolchain only approximates the software used internally by the experimental collaborations, while the published papers inevitably omit implementation details needed for a faithful reconstruction. Agents must therefore rely on physical reasoning, domain knowledge, and trial-and-error to fill these gaps. Each task requires the agent to turn a published analysis into an executable simulation-and-selection pipeline and submit predicted collision event yields in specified signal regions. These predictions are evaluated with standard histogram metrics that provide continuous fidelity scores without a hand-written rubric. We also report the computational cost incurred by each agent per task. Finally, we evaluate the codebase and full session trace using an LLM judge to catch qualitative failure modes such as fabrications, hallucinations and duplications. We release an initial set of tasks drawn from LHC searches, together with a containerized sandbox and event simulation tools. We evaluate across a capability ladder of general purpose coding agents. Our results show that on average no agent reliably beats the physicist-in-the-loop solution.
Code agents must both reason over long-horizon repository state and obey strict tool-use protocols. In paired Instruct/Thinking checkpoints, these capabilities are complementary but misaligned. The Instruct model is concise and tool-disciplined, whereas the Thinking model offers stronger planning and recovery behavior but often over-deliberates and degrades agent performance. We present CRANE (Constrained Reasoning Injection for Code Agents via Nullspace Editing), a training-free parameter-editing method that treats the Thinking-Instruct delta as a directional pool of candidate reasoning edits for the Instruct backbone. CRANE combines magnitude thresholding to denoise the delta, a Conservative Taylor Gate to retain edits that are jointly beneficial for reasoning transfer and tool-use preservation, and Graduated Sigmoidal Projection to suppress format-critical update directions. By merging paired Instruct and Thinking checkpoints, CRANE delivers strong gains over either individual model while preserving Instruct-level efficiency: on Roo-Eval it achieves pass1 of 66.2% (+19.5%) for Qwen3-30B-A3B and 81.5% (+8.7%) for Qwen3-Next-80B-A3B; on SWE-bench-Verified it resolves up to 14 additional instances at both scales (122/500 and 180/500); and on Terminal-Bench v2 it improves pass1/pass5 by up to 2.3%/7.8%, reaching 7.6%/17.9% and 14.8%/30.3%, respectively, consistently outperforming alternative merging strategies across all three benchmarks.
展开完整摘要收起摘要↓
Code agents must both reason over long-horizon repository state and obey strict tool-use protocols. In paired Instruct/Thinking checkpoints, these capabilities are complementary but misaligned. The Instruct model is concise and tool-disciplined, whereas the Thinking model offers stronger planning and recovery behavior but often over-deliberates and degrades agent performance. We present CRANE (Constrained Reasoning Injection for Code Agents via Nullspace Editing), a training-free parameter-editing method that treats the Thinking-Instruct delta as a directional pool of candidate reasoning edits for the Instruct backbone. CRANE combines magnitude thresholding to denoise the delta, a Conservative Taylor Gate to retain edits that are jointly beneficial for reasoning transfer and tool-use preservation, and Graduated Sigmoidal Projection to suppress format-critical update directions. By merging paired Instruct and Thinking checkpoints, CRANE delivers strong gains over either individual model while preserving Instruct-level efficiency: on Roo-Eval it achieves pass1 of 66.2% (+19.5%) for Qwen3-30B-A3B and 81.5% (+8.7%) for Qwen3-Next-80B-A3B; on SWE-bench-Verified it resolves up to 14 additional instances at both scales (122/500 and 180/500); and on Terminal-Bench v2 it improves pass1/pass5 by up to 2.3%/7.8%, reaching 7.6%/17.9% and 14.8%/30.3%, respectively, consistently outperforming alternative merging strategies across all three benchmarks.
Autonomous language-model agents increasingly combine planning, tool use, document processing, browsing, code execution, and verification loops. These capabilities make agent systems more useful, but they also introduce operational failure modes that are not visible from final accuracy alone. This report presents ChromaFlow, a tool-augmented autonomous reasoning framework built around planner-directed execution, specialized tool use, and telemetry-driven evaluation. We analyze ChromaFlow on GAIA 2023 Level-1 validation tasks under clean evaluation constraints. A frozen full Level-1 baseline achieved 29/53 correct answers, or 54.72%. A later recovery configuration with expanded orchestration achieved 27/53 correct answers, or 50.94%, while increasing tracebacks, timeout events, tool-failure mentions, token-line calls, and campaign-log cost estimates. Two randomized 20-task smoke evaluations produced 12/20 and 11/20 correct answers, showing that small diagnostic gains can be unstable across samples. The central result is therefore a negative ablation: more aggressive orchestration did not improve full-set performance and increased operational noise. The report argues that bounded planner escalation, deterministic extraction, evidence reconciliation, and explicit run gates should be treated as first-order requirements for reliable autonomous agent evaluation.
展开完整摘要收起摘要↓
Autonomous language-model agents increasingly combine planning, tool use, document processing, browsing, code execution, and verification loops. These capabilities make agent systems more useful, but they also introduce operational failure modes that are not visible from final accuracy alone. This report presents ChromaFlow, a tool-augmented autonomous reasoning framework built around planner-directed execution, specialized tool use, and telemetry-driven evaluation. We analyze ChromaFlow on GAIA 2023 Level-1 validation tasks under clean evaluation constraints. A frozen full Level-1 baseline achieved 29/53 correct answers, or 54.72%. A later recovery configuration with expanded orchestration achieved 27/53 correct answers, or 50.94%, while increasing tracebacks, timeout events, tool-failure mentions, token-line calls, and campaign-log cost estimates. Two randomized 20-task smoke evaluations produced 12/20 and 11/20 correct answers, showing that small diagnostic gains can be unstable across samples. The central result is therefore a negative ablation: more aggressive orchestration did not improve full-set performance and increased operational noise. The report argues that bounded planner escalation, deterministic extraction, evidence reconciliation, and explicit run gates should be treated as first-order requirements for reliable autonomous agent evaluation.
作者Marius S. Knorr, Robert Müller, Jan P. Bremer, Nils Schweingruber
Fast Healthcare Interoperability Resources (FHIR) is the dominant standard for interoperable exchange of healthcare data. In FHIR, electronic health records form a directed graph of resources. Answering clinically meaningful questions over FHIR requires agents to perform multi-step reasoning, filtering, and aggregation across multiple resource types. Prior work shows that even tool-augmented LLM agents (retrieval, code execution, multi-turn planning) often select the wrong resources or violate traversal constraints. We study this problem in the context of FHIR-AgentBench, a benchmark for realistic question answering over real-world hospital data, and frame reasoning on FHIR as a sequential decision-making problem over a queryable structured graph. We implement a multi-turn CodeAct agent and post-train it with reinforcement learning using a custom harness and tools. A LLM Judge provides execution-grounded rewards. Compared to prompt-based, closed-model baselines, RL post-training improves performance while enforcing data-integrity constraints. Empirically, our approach improves answer correctness from 50% (o4-mini) to 77% on FHIR-AgentBench using a smaller and cheaper Qwen3-8B model. We present an end-to-end post-training pipeline (environment building, harness construction, model training and custom evaluation) that reliably improves multi-turn reasoning over structured clinical graphs.
展开完整摘要收起摘要↓
Fast Healthcare Interoperability Resources (FHIR) is the dominant standard for interoperable exchange of healthcare data. In FHIR, electronic health records form a directed graph of resources. Answering clinically meaningful questions over FHIR requires agents to perform multi-step reasoning, filtering, and aggregation across multiple resource types. Prior work shows that even tool-augmented LLM agents (retrieval, code execution, multi-turn planning) often select the wrong resources or violate traversal constraints. We study this problem in the context of FHIR-AgentBench, a benchmark for realistic question answering over real-world hospital data, and frame reasoning on FHIR as a sequential decision-making problem over a queryable structured graph. We implement a multi-turn CodeAct agent and post-train it with reinforcement learning using a custom harness and tools. A LLM Judge provides execution-grounded rewards. Compared to prompt-based, closed-model baselines, RL post-training improves performance while enforcing data-integrity constraints. Empirically, our approach improves answer correctness from 50% (o4-mini) to 77% on FHIR-AgentBench using a smaller and cheaper Qwen3-8B model. We present an end-to-end post-training pipeline (environment building, harness construction, model training and custom evaluation) that reliably improves multi-turn reasoning over structured clinical graphs.
The rapid evolution of Embodied AI has enabled Vision-Language-Action (VLA) models to excel in multimodal perception and task execution. However, applying Reinforcement Learning (RL) to these massive models in large-scale distributed environments faces severe systemic bottlenecks, primarily due to the resource conflict between high-fidelity physical simulation and the intensive VRAM/bandwidth demands of deep learning. This conflict often leaves overall throughput constrained by execution-phase inefficiencies. To address these challenges, we propose D-VLA, a high-concurrency, low-latency distributed RL framework for large-scale embodied foundation models. D-VLA introduces "Plane Decoupling," physically isolating high-frequency training data from low-frequency weight control to eliminate interference between simulation and optimization. We further design a four-thread asynchronous "Swimlane" pipeline, enabling full parallel overlap of sampling, inference, gradient computation, and parameter distribution. Additionally, a dual-pool VRAM management model and topology-aware replication resolve memory fragmentation and optimize communication efficiency. Experiments on benchmarks like LIBERO show that D-VLA significantly outperforms mainstream RL frameworks in throughput and sampling efficiency for billion-parameter VLA models. In trillion-parameter scalability tests, our framework maintains exceptional stability and linear speedup, providing a robust system for high-performance general-purpose embodied agents.
展开完整摘要收起摘要↓
The rapid evolution of Embodied AI has enabled Vision-Language-Action (VLA) models to excel in multimodal perception and task execution. However, applying Reinforcement Learning (RL) to these massive models in large-scale distributed environments faces severe systemic bottlenecks, primarily due to the resource conflict between high-fidelity physical simulation and the intensive VRAM/bandwidth demands of deep learning. This conflict often leaves overall throughput constrained by execution-phase inefficiencies. To address these challenges, we propose D-VLA, a high-concurrency, low-latency distributed RL framework for large-scale embodied foundation models. D-VLA introduces "Plane Decoupling," physically isolating high-frequency training data from low-frequency weight control to eliminate interference between simulation and optimization. We further design a four-thread asynchronous "Swimlane" pipeline, enabling full parallel overlap of sampling, inference, gradient computation, and parameter distribution. Additionally, a dual-pool VRAM management model and topology-aware replication resolve memory fragmentation and optimize communication efficiency. Experiments on benchmarks like LIBERO show that D-VLA significantly outperforms mainstream RL frameworks in throughput and sampling efficiency for billion-parameter VLA models. In trillion-parameter scalability tests, our framework maintains exceptional stability and linear speedup, providing a robust system for high-performance general-purpose embodied agents.
Distilling reasoning traces from strong large language models into smaller ones is a promising route to improve intelligence in resource-constrained settings. Existing approaches face a fundamental trade-off: offline distillation from teacher-generated traces provides high-quality, sample-efficient supervision but suffers from distributional drift: during training, the student model conditions on teacher-generated prefixes, whereas during inference the student autoregresses on self-generated prefixes, leading to compounding errors over long reasoning trajectories. Meanwhile, on-policy or self-distillation methods better match the student's inference-time distribution, but require costly online sampling and often produce low-quality traces in early training. We propose a principled offline reasoning distillation framework that preserves the efficiency and supervision quality of offline teacher-generated data while correcting teacher-student distribution drift. It adaptively emphasizes teacher supervision that is better aligned with the student's on-policy distribution. Evaluations on mathematical reasoning benchmarks of GSM8K, MATH, MATH500, and harder held-out competition-style tasks, including AMC, AIME, and OlympiadBench, show that our method improves reasoning accuracy over prior offline distillation algorithms and yields more stable reasoning traces while preserving instruction-following capabilities. Our work shows that lightweight, distribution-correction-aware training can substantially strengthen offline reasoning distillation without online rollouts.
展开完整摘要收起摘要↓
Distilling reasoning traces from strong large language models into smaller ones is a promising route to improve intelligence in resource-constrained settings. Existing approaches face a fundamental trade-off: offline distillation from teacher-generated traces provides high-quality, sample-efficient supervision but suffers from distributional drift: during training, the student model conditions on teacher-generated prefixes, whereas during inference the student autoregresses on self-generated prefixes, leading to compounding errors over long reasoning trajectories. Meanwhile, on-policy or self-distillation methods better match the student's inference-time distribution, but require costly online sampling and often produce low-quality traces in early training. We propose a principled offline reasoning distillation framework that preserves the efficiency and supervision quality of offline teacher-generated data while correcting teacher-student distribution drift. It adaptively emphasizes teacher supervision that is better aligned with the student's on-policy distribution. Evaluations on mathematical reasoning benchmarks of GSM8K, MATH, MATH500, and harder held-out competition-style tasks, including AMC, AIME, and OlympiadBench, show that our method improves reasoning accuracy over prior offline distillation algorithms and yields more stable reasoning traces while preserving instruction-following capabilities. Our work shows that lightweight, distribution-correction-aware training can substantially strengthen offline reasoning distillation without online rollouts.
The rise in deployment of large language models has driven a surge in GPU demand and datacenter scaling, raising concerns about electricity use, grid stress, and the impacts of modern AI workloads. Distillation is often promoted as one of the most effective paths to obtain cheaper, more efficient models, yet these claims rarely account for the full end-to-end energy and resource costs, including crucial teacher-side workloads such as data generation, logit caching, and evaluation. We present a comprehensive energy accounting framework that measures the complete computational cost of distillation pipelines via detailed stage-wise tracking of GPU device power consumption. In our experiments, we separate and log empirical energy use across distinct phases and systematically measure the energy and emissions of two common distillation methods: the classic logit-based knowledge distillation and synthetic-data supervised fine-tuning, constructing energy-quality Pareto frontiers that expose the previously ignored costs. From these measurements and analyses, we derive practical design rules for selecting distillation methods and hyperparameters under energy and budget constraints, and release an open-source measurement harness and accounting protocol to provide a standardized foundation for comparable, reproducible distillation research, explicitly accountable for complete pipeline energy impact.
展开完整摘要收起摘要↓
The rise in deployment of large language models has driven a surge in GPU demand and datacenter scaling, raising concerns about electricity use, grid stress, and the impacts of modern AI workloads. Distillation is often promoted as one of the most effective paths to obtain cheaper, more efficient models, yet these claims rarely account for the full end-to-end energy and resource costs, including crucial teacher-side workloads such as data generation, logit caching, and evaluation. We present a comprehensive energy accounting framework that measures the complete computational cost of distillation pipelines via detailed stage-wise tracking of GPU device power consumption. In our experiments, we separate and log empirical energy use across distinct phases and systematically measure the energy and emissions of two common distillation methods: the classic logit-based knowledge distillation and synthetic-data supervised fine-tuning, constructing energy-quality Pareto frontiers that expose the previously ignored costs. From these measurements and analyses, we derive practical design rules for selecting distillation methods and hyperparameters under energy and budget constraints, and release an open-source measurement harness and accounting protocol to provide a standardized foundation for comparable, reproducible distillation research, explicitly accountable for complete pipeline energy impact.
作者Zhuojin Li, Hsin-Pai Cheng, Hong Cai, Shizhong Han, Fatih Porikli
Diffusion Transformers (DiTs) deliver remarkable image and video generation quality but incur high computational cost, limiting scalability and on-device deployment. We introduce CoReDiT, a structured token pruning framework for DiTs across vision tasks. CoReDiT uses a linear-time spatial coherence score to estimate local redundancy in the latent token lattice and skips high coherence (redundant) tokens in self-attention. To maintain a dense representation and avoid visual discontinuities, we reconstruct skipped attention outputs via coherence-guided aggregation of spatially neighboring retained tokens. We further introduce a progressive, block-adaptive pruning schedule that increases pruning gradually and allocates larger budgets to blocks and denoising steps with higher redundancy. Across state-of-the-art diffusion backbones including PixArt-α and MagicDrive-V2, CoReDiT achieves up to 55% self-attention FLOPs reduction and inference speedups of 1.33x on cloud GPUs and 1.72x on mobile NPUs, while maintaining high visual quality. Notably, CoReDiT also increases on-device memory head-room, enabling higher-resolution generation.
展开完整摘要收起摘要↓
Diffusion Transformers (DiTs) deliver remarkable image and video generation quality but incur high computational cost, limiting scalability and on-device deployment. We introduce CoReDiT, a structured token pruning framework for DiTs across vision tasks. CoReDiT uses a linear-time spatial coherence score to estimate local redundancy in the latent token lattice and skips high coherence (redundant) tokens in self-attention. To maintain a dense representation and avoid visual discontinuities, we reconstruct skipped attention outputs via coherence-guided aggregation of spatially neighboring retained tokens. We further introduce a progressive, block-adaptive pruning schedule that increases pruning gradually and allocates larger budgets to blocks and denoising steps with higher redundancy. Across state-of-the-art diffusion backbones including PixArt-α and MagicDrive-V2, CoReDiT achieves up to 55% self-attention FLOPs reduction and inference speedups of 1.33x on cloud GPUs and 1.72x on mobile NPUs, while maintaining high visual quality. Notably, CoReDiT also increases on-device memory head-room, enabling higher-resolution generation.