General-purpose robots must infer what a new task requires and translate that understanding into appropriate physical action. In-context learning (ICL) for robots supports this process by using demonstrations and interaction to direct existing competence with neural parameters held fixed during deployment. We organize this literature review around the interfaces connecting contextual evidence to execution, distinguishing four families: context-conditioned policies, geometric demonstration transfer, world-model-based control, and skill- and agent-based execution. Comparing these interfaces clarifies their transfer assumptions and the roles of training, correspondence, and memory in making context useful. Across manipulation and navigation, we examine how these mechanisms preserve taught requirements as objects, environments, and execution conditions change. This analysis links method design to evaluation practices that distinguish responsiveness to teaching, physical transfer, and benefits from retained experience. The resulting agenda connects compositional task acquisition and faithful transfer with physical recursive self-improvement, in which experience improves the ability to learn subsequent tasks.
展开完整摘要收起摘要↓
General-purpose robots must infer what a new task requires and translate that understanding into appropriate physical action. In-context learning (ICL) for robots supports this process by using demonstrations and interaction to direct existing competence with neural parameters held fixed during deployment. We organize this literature review around the interfaces connecting contextual evidence to execution, distinguishing four families: context-conditioned policies, geometric demonstration transfer, world-model-based control, and skill- and agent-based execution. Comparing these interfaces clarifies their transfer assumptions and the roles of training, correspondence, and memory in making context useful. Across manipulation and navigation, we examine how these mechanisms preserve taught requirements as objects, environments, and execution conditions change. This analysis links method design to evaluation practices that distinguish responsiveness to teaching, physical transfer, and benefits from retained experience. The resulting agenda connects compositional task acquisition and faithful transfer with physical recursive self-improvement, in which experience improves the ability to learn subsequent tasks.
作者Yundaichuan Zhan, Weishi Wang, Wenbiao Liu, Daniel Dahlmeier, Chengwei Qin, Juncheng Li, Fredrik D. Johansson, Zhongqi Yue
We study how to build more capable general-purpose agents by extending large language models (LLMs) with native typed decision-making. We introduce Dyad, an architecture that augments a pretrained LLM with an environment-conditioned action encoder that embeds each candidate action description in parallel, then scores these embeddings against the LLM's internal state to yield a distribution over typed actions. By factorizing decision-making into representations of the evolving interaction state and environment-specific action semantics, Dyad introduces an inductive bias for learning reusable representations while keeping action scoring efficient even as the action space grows. We investigate two complementary reinforcement learning settings driven by environment interaction. With the LLM frozen, training the action encoder alone achieves consistent gains across four unseen environments, enabling modular adaptation without modifying any LLM parameters. Jointly optimizing both components outperforms conventional RL post-training across diverse interactive tasks and model scales, including a 3.80% average absolute gain on ALFWorld with a 9B model, while improving general knowledge, reasoning, and coding.
展开完整摘要收起摘要↓
We study how to build more capable general-purpose agents by extending large language models (LLMs) with native typed decision-making. We introduce Dyad, an architecture that augments a pretrained LLM with an environment-conditioned action encoder that embeds each candidate action description in parallel, then scores these embeddings against the LLM's internal state to yield a distribution over typed actions. By factorizing decision-making into representations of the evolving interaction state and environment-specific action semantics, Dyad introduces an inductive bias for learning reusable representations while keeping action scoring efficient even as the action space grows. We investigate two complementary reinforcement learning settings driven by environment interaction. With the LLM frozen, training the action encoder alone achieves consistent gains across four unseen environments, enabling modular adaptation without modifying any LLM parameters. Jointly optimizing both components outperforms conventional RL post-training across diverse interactive tasks and model scales, including a 3.80% average absolute gain on ALFWorld with a 9B model, while improving general knowledge, reasoning, and coding.
Large language models can reason in continuous space instead of decoded text, by recurring on their own hidden states or by passing those states between agents, while training supervises only the Cross-Entropy (CE) of the final decoded answer and does not constrain the thought. Theoretical and empirical analyses establish and confirm four failures of CE-only training that lead to a lower probability of the correct answer such as collapsing thoughts across distinct questions and retaining irrelevant information. We introduce REST (REpresentation-Supervised Thoughts), a training objective that turns four properties of a valid thought representation (causality, minimality, separability, and stability) into differentiable losses added to CE. We instantiate it in latent single-agent and multi-agent systems, without architectural changes or added parameters at inference. Across 7 benchmarks spanning mathematics, science, medicine, and code generation, with the same training data, compute, and latent budget, REST increases accuracy over CE-only training across agent settings and model sizes by up to 7.5 percentage points and convergence on a final answer by 30%. Furthermore, REST thoughts encode more of what is required to achieve the correct answer, and decoding them better recovers the intended output of the agent, which makes latent communication easier to interpret. Project Website: https://fard-lab.github.io/REST
展开完整摘要收起摘要↓
Large language models can reason in continuous space instead of decoded text, by recurring on their own hidden states or by passing those states between agents, while training supervises only the Cross-Entropy (CE) of the final decoded answer and does not constrain the thought. Theoretical and empirical analyses establish and confirm four failures of CE-only training that lead to a lower probability of the correct answer such as collapsing thoughts across distinct questions and retaining irrelevant information. We introduce REST (REpresentation-Supervised Thoughts), a training objective that turns four properties of a valid thought representation (causality, minimality, separability, and stability) into differentiable losses added to CE. We instantiate it in latent single-agent and multi-agent systems, without architectural changes or added parameters at inference. Across 7 benchmarks spanning mathematics, science, medicine, and code generation, with the same training data, compute, and latent budget, REST increases accuracy over CE-only training across agent settings and model sizes by up to 7.5 percentage points and convergence on a final answer by 30%. Furthermore, REST thoughts encode more of what is required to achieve the correct answer, and decoding them better recovers the intended output of the agent, which makes latent communication easier to interpret. Project Website: https://fard-lab.github.io/REST
作者Johnny Jingze Li, Abdulla Kuleib, Kalyan Basu, Gabriel A. Silva
The way a model distributes activity over each layer's attention heads offers a coarse view of how it routes information through depth; how this changes with the task is part of what a mechanistic account must explain. Holding prompt length fixed, we vary how many serial steps a task demands and measure, in every layer of 17 open-weight models, whether activity concentrates on a few heads or spreads across many as demand rises. Both occur: in most models, layers just before mid-depth concentrate activity and later layers spread it. Models differ in where and how strongly this happens. The Qwen2.5 base models from 0.5B to 7B, for example, spread less than the average model in every task and concentrate activity in parts of their second half, where Llama models from 1B to 8B and OLMo-2 spread; the contrast largely holds between Llama-3.1-70B and Qwen2.5-72B, which have the same number of layers and heads. These differences are reproducible, and post-trained models keep much of their base model's pattern. An ablation study suggests that, within a task, models whose activity is more concentrated on their top heads also depend more on those heads for the answer. Concentration and spreading across layers thus offer a new way to compare models, by how they route information through depth. Code is available at https://github.com/johnnyjli/serial-demand-heads.
展开完整摘要收起摘要↓
The way a model distributes activity over each layer's attention heads offers a coarse view of how it routes information through depth; how this changes with the task is part of what a mechanistic account must explain. Holding prompt length fixed, we vary how many serial steps a task demands and measure, in every layer of 17 open-weight models, whether activity concentrates on a few heads or spreads across many as demand rises. Both occur: in most models, layers just before mid-depth concentrate activity and later layers spread it. Models differ in where and how strongly this happens. The Qwen2.5 base models from 0.5B to 7B, for example, spread less than the average model in every task and concentrate activity in parts of their second half, where Llama models from 1B to 8B and OLMo-2 spread; the contrast largely holds between Llama-3.1-70B and Qwen2.5-72B, which have the same number of layers and heads. These differences are reproducible, and post-trained models keep much of their base model's pattern. An ablation study suggests that, within a task, models whose activity is more concentrated on their top heads also depend more on those heads for the answer. Concentration and spreading across layers thus offer a new way to compare models, by how they route information through depth. Code is available at https://github.com/johnnyjli/serial-demand-heads.
作者Neemias B. da Silva, Martin Lukk, Ali Sutani, Abhishek Moturu, Harris Yang, Daniel Silver, Matt Ratto, Thiago H. Silva
Large language models (LLMs) show considerable potential in simulating human attitudes and preferences. Prior work finds that LLM-generated responses can compress the range of attitudes found within populations and misrepresent particular subgroups in ways that vary across models and topics. We introduce Population Fidelity, an evaluation framework that distinguishes key conditions required for a set of LLM-generated responses to represent a population. It incorporates three dimensions: group-level accuracy, the amount of between-group variation, and the structure of that variation. We demonstrate the framework's utility in two ways. First, we reproduce a prior study of "machine bias" in LLM survey responses and apply the framework to its models and more recent ones, showing that poor representation reflects not only insufficient between-group variation but also variation assigned to the wrong groups. Second, we evaluate one proposed approach to improving models' population representativeness: cultural fine-tuning. We find that cultural fine-tuning can improve alignment with the survey center without improving the representation of within-population differences, a distinction that measures of aggregate agreement do not capture. We argue that representing a population requires models to reproduce several features of human attitudinal variation simultaneously. Our framework organizes these features and provides reusable code, data, and trained models for evaluating population fidelity across substantive domains and assessing proposed alignment methods.
展开完整摘要收起摘要↓
Large language models (LLMs) show considerable potential in simulating human attitudes and preferences. Prior work finds that LLM-generated responses can compress the range of attitudes found within populations and misrepresent particular subgroups in ways that vary across models and topics. We introduce Population Fidelity, an evaluation framework that distinguishes key conditions required for a set of LLM-generated responses to represent a population. It incorporates three dimensions: group-level accuracy, the amount of between-group variation, and the structure of that variation. We demonstrate the framework's utility in two ways. First, we reproduce a prior study of "machine bias" in LLM survey responses and apply the framework to its models and more recent ones, showing that poor representation reflects not only insufficient between-group variation but also variation assigned to the wrong groups. Second, we evaluate one proposed approach to improving models' population representativeness: cultural fine-tuning. We find that cultural fine-tuning can improve alignment with the survey center without improving the representation of within-population differences, a distinction that measures of aggregate agreement do not capture. We argue that representing a population requires models to reproduce several features of human attitudinal variation simultaneously. Our framework organizes these features and provides reusable code, data, and trained models for evaluating population fidelity across substantive domains and assessing proposed alignment methods.
Despite the recent success of coding agents built on large language models, it remains challenging to run them over long horizons, since every observation is appended to the context and the context grows with each one. History-based maintenance is a common remedy, which masks or summarizes old observations, or prunes what a model reads as useless, and bounds the context at little cost. However, it decides from the text of the history alone and sees nothing of how the code is connected. Since a coding agent edits code many times over a single task, and each write can change what code elsewhere means, such maintenance may keep records a write has falsified, drop ones that still hold, and miss code the agent needs next. To overcome these challenges, this paper proposes StateTape, a novel and scalable framework that rewrites a coding agent's context as the repository changes rather than as the context grows. The key idea of StateTape is to model the repository as a symbol-level code graph, whose dependencies and language rules expose which symbols a write can affect. Upon this graph, a tape marks the symbols each write changed, which turns staleness from an inference about text into an observation of the agent's writes. We propose a per-write procedure in which the tape nominates the records a write could have falsified while a small manager model settles what the write log cannot, and further provide a theoretical analysis and TraceBench, a benchmark that labels what an agent is holding against what is actually needed. Empirically, we demonstrate that StateTape can effectively clear falsified records and retrieve what is needed, and thus achieve a higher resolve rate in all experiments spanned by six coding agents and three edit-heavy benchmarks with little computational overhead.
展开完整摘要收起摘要↓
Despite the recent success of coding agents built on large language models, it remains challenging to run them over long horizons, since every observation is appended to the context and the context grows with each one. History-based maintenance is a common remedy, which masks or summarizes old observations, or prunes what a model reads as useless, and bounds the context at little cost. However, it decides from the text of the history alone and sees nothing of how the code is connected. Since a coding agent edits code many times over a single task, and each write can change what code elsewhere means, such maintenance may keep records a write has falsified, drop ones that still hold, and miss code the agent needs next. To overcome these challenges, this paper proposes StateTape, a novel and scalable framework that rewrites a coding agent's context as the repository changes rather than as the context grows. The key idea of StateTape is to model the repository as a symbol-level code graph, whose dependencies and language rules expose which symbols a write can affect. Upon this graph, a tape marks the symbols each write changed, which turns staleness from an inference about text into an observation of the agent's writes. We propose a per-write procedure in which the tape nominates the records a write could have falsified while a small manager model settles what the write log cannot, and further provide a theoretical analysis and TraceBench, a benchmark that labels what an agent is holding against what is actually needed. Empirically, we demonstrate that StateTape can effectively clear falsified records and retrieve what is needed, and thus achieve a higher resolve rate in all experiments spanned by six coding agents and three edit-heavy benchmarks with little computational overhead.
作者Roshan Reddy Upendra, Alexandre Dorais, Joe Meyer, Andrew Pouret, Anastasios Lambrianos Stappas, Dinesh Katupputhur Ramprasath, Tom Palczewski, Minghua Li
Relational in-context learning (ICL) conditions predictions on the labeled support examples and their linked tables, creating a failure mode when the support set contains target-derived features that are unavailable for the query. We formulate this problem as support-set target leakage, distinct from leakage during dataset construction, temporal splitting, or representation learning. Here, the target-derived (leaker) columns are present only in the labeled support set during relational in-context inference, while queries remain clean. We construct 14 synthetic leaker types, corresponding to 20 columns, spanning proxies with different noise levels, coverage, modalities, semantic transparency, and relational distances. We evaluate a frozen relational encoder with an ICL head on held-out RelBench databases and use Integrated Gradients (IG) to rank and remove suspicious columns. Our results show that the effect of support-set leakage varies across tasks and relational distances. Target-table leakers cause the clearest degradation, while one- and two-hop leakers are not consistently used by the model. IG ranks target-table leakers highly across datasets and partially recovers performance in settings where leakage has the largest effect.
展开完整摘要收起摘要↓
Relational in-context learning (ICL) conditions predictions on the labeled support examples and their linked tables, creating a failure mode when the support set contains target-derived features that are unavailable for the query. We formulate this problem as support-set target leakage, distinct from leakage during dataset construction, temporal splitting, or representation learning. Here, the target-derived (leaker) columns are present only in the labeled support set during relational in-context inference, while queries remain clean. We construct 14 synthetic leaker types, corresponding to 20 columns, spanning proxies with different noise levels, coverage, modalities, semantic transparency, and relational distances. We evaluate a frozen relational encoder with an ICL head on held-out RelBench databases and use Integrated Gradients (IG) to rank and remove suspicious columns. Our results show that the effect of support-set leakage varies across tasks and relational distances. Target-table leakers cause the clearest degradation, while one- and two-hop leakers are not consistently used by the model. IG ranks target-table leakers highly across datasets and partially recovers performance in settings where leakage has the largest effect.
作者Zhiheng Chen, Ye Qiao, Mohammad Abdullah Al Faruque, Sitao Huang
Physical AI models such as Vision-Language-Action (VLA) architectures enable generalist robotic policies through large-scale transformer backbones and diffusion-based action decoders. While edge GPU platforms excel at parallelizing the compute-intensive vision-transformer workloads, they exhibit fundamental limitations for the Diffusion Action Expert (DAE) module: the iterative denoising process requires repeated parameter loading from DRAM across multiple steps, resulting in memory-bound performance where the GPU's massive computational throughput remains underutilized. This mismatch between DAE's I/O-intensive characteristics and GPU's compute-centric architecture motivates a heterogeneous acceleration approach. This paper presents EdgeDAE, a heterogeneous FPGA-GPU system that strategically partitions workloads based on computational characteristics. We offload the perception-heavy vision-transformer to GPU while accelerating DAE inference on FPGA through complete on-chip parameter storage in BRAM/URAM. This architecture eliminates the memory bottleneck by co-designing quantization strategies, fixed-point arithmetic, and hardware-efficient random number generation for the FPGA fabric. Compared to an edge GPU baseline, EdgeDAE reduces end-to-end inference latency by 52.5% for Octo-Small and 38.7% for Octo-Base, with up to $2.10\times$ higher throughput; compared to a consumer GPU (RTX~4090), it achieves ${\sim}17\times$ higher energy efficiency.
展开完整摘要收起摘要↓
Physical AI models such as Vision-Language-Action (VLA) architectures enable generalist robotic policies through large-scale transformer backbones and diffusion-based action decoders. While edge GPU platforms excel at parallelizing the compute-intensive vision-transformer workloads, they exhibit fundamental limitations for the Diffusion Action Expert (DAE) module: the iterative denoising process requires repeated parameter loading from DRAM across multiple steps, resulting in memory-bound performance where the GPU's massive computational throughput remains underutilized. This mismatch between DAE's I/O-intensive characteristics and GPU's compute-centric architecture motivates a heterogeneous acceleration approach. This paper presents EdgeDAE, a heterogeneous FPGA-GPU system that strategically partitions workloads based on computational characteristics. We offload the perception-heavy vision-transformer to GPU while accelerating DAE inference on FPGA through complete on-chip parameter storage in BRAM/URAM. This architecture eliminates the memory bottleneck by co-designing quantization strategies, fixed-point arithmetic, and hardware-efficient random number generation for the FPGA fabric. Compared to an edge GPU baseline, EdgeDAE reduces end-to-end inference latency by 52.5% for Octo-Small and 38.7% for Octo-Base, with up to $2.10\times$ higher throughput; compared to a consumer GPU (RTX~4090), it achieves ${\sim}17\times$ higher energy efficiency.
作者David Racovan, Ajay Rawat, Christopher K. May, Jeffrey A. Turkstra
The rapid proliferation of large language models (LLMs) in the context of education has introduced significant challenges in enforcement of academic integrity, especially in programming courses. We present Argus, an automated detection system for LLM-assisted student work in undergraduate C programming assignments. Argus integrates behavioral and stylistic indicators to create a holistic picture of the student's progress through an assignment and surfaces anomalies that point to potential misuse of LLM assistance. We quantify and analyze data over six years of Spring semester offerings in a large-enrollment CS2 course at Purdue University using Argus, finding that 45% of enrolled students exhibited patterns consistent with LLM-assisted code development in Spring 2026. To contextualize these results, we analyze the relationship between flagged LLM use and student performance on written, in-person proctored examinations, and find a significant negative correlation. We also explore the problem of mitigating false positives, recognizing that erroneous accusations of academic integrity carry significant consequences for students and instructors alike, particularly in the context of large enrollment courses. We argue that any automated detection system must be accompanied by a structured process for human review. We discuss the consequences for future course design, changing academic policy as these tools become more ubiquitous, and the pedagogical implications of LLM-based tools in computer science education.
展开完整摘要收起摘要↓
The rapid proliferation of large language models (LLMs) in the context of education has introduced significant challenges in enforcement of academic integrity, especially in programming courses. We present Argus, an automated detection system for LLM-assisted student work in undergraduate C programming assignments. Argus integrates behavioral and stylistic indicators to create a holistic picture of the student's progress through an assignment and surfaces anomalies that point to potential misuse of LLM assistance. We quantify and analyze data over six years of Spring semester offerings in a large-enrollment CS2 course at Purdue University using Argus, finding that 45% of enrolled students exhibited patterns consistent with LLM-assisted code development in Spring 2026. To contextualize these results, we analyze the relationship between flagged LLM use and student performance on written, in-person proctored examinations, and find a significant negative correlation. We also explore the problem of mitigating false positives, recognizing that erroneous accusations of academic integrity carry significant consequences for students and instructors alike, particularly in the context of large enrollment courses. We argue that any automated detection system must be accompanied by a structured process for human review. We discuss the consequences for future course design, changing academic policy as these tools become more ubiquitous, and the pedagogical implications of LLM-based tools in computer science education.
作者Xiangyu Zhou, Saleh Zare Zade, Rafi Ibn Sultan, Alexander Kotov, Dongxiao Zhu
Large Reasoning Models (LRMs) are commonly trained with reinforcement learning (RL) to improve their generation of chain-of-thought (CoT) reasoning before producing final answers. However, RL rewards are typically assigned based on final answers, providing little or no direct supervision over intermediate reasoning. This can lead to deceptive safety alignment, where the reasoning trace and final answer convey inconsistent safety signals. To systematically investigate this phenomenon, we introduce DSAR (Deceptive Safety Alignment Rate), a metric that jointly assesses reasoning traces and final answers to quantify their safety inconsistency. Across multiple LRMs and benchmarks, we find that deceptive safety alignment is pervasive under standard prompting conditions and is substantially amplified under prefilling attacks. We further provide a hidden representation analysis showing that models exhibit stronger safety discrimination at the final-answer stage than during intermediate reasoning. To close this gap, we propose SARA (Safety-Aware Reasoning Alignment), an RL-based method that rewards both safety-aware reasoning and safe final answers, encouraging early harmful intent recognition and enforcing reasoning-answer consistency. Experiments show that SARA significantly mitigates deceptive safety alignment under both standard and adversarial settings while preserving helpfulness and utility. Code is available at https://github.com/xzhou98/SARA.
展开完整摘要收起摘要↓
Large Reasoning Models (LRMs) are commonly trained with reinforcement learning (RL) to improve their generation of chain-of-thought (CoT) reasoning before producing final answers. However, RL rewards are typically assigned based on final answers, providing little or no direct supervision over intermediate reasoning. This can lead to deceptive safety alignment, where the reasoning trace and final answer convey inconsistent safety signals. To systematically investigate this phenomenon, we introduce DSAR (Deceptive Safety Alignment Rate), a metric that jointly assesses reasoning traces and final answers to quantify their safety inconsistency. Across multiple LRMs and benchmarks, we find that deceptive safety alignment is pervasive under standard prompting conditions and is substantially amplified under prefilling attacks. We further provide a hidden representation analysis showing that models exhibit stronger safety discrimination at the final-answer stage than during intermediate reasoning. To close this gap, we propose SARA (Safety-Aware Reasoning Alignment), an RL-based method that rewards both safety-aware reasoning and safe final answers, encouraging early harmful intent recognition and enforcing reasoning-answer consistency. Experiments show that SARA significantly mitigates deceptive safety alignment under both standard and adversarial settings while preserving helpfulness and utility. Code is available at https://github.com/xzhou98/SARA.
作者Sahand Rezaei-Shoshtari, Patryk Wozniczka, Shu Ishida, Gregg Streuber, Farnoosh Javadi, Jeffrey Landes, Angela Ju, Muhammad Azam, Bryan Lim, Johan Luttun, Indrajeet Haldar, Jonathan Shaw, Beatriz Guerra, Ivan Sosnovik, James Stoddart, Robert Giaquinto, Adam Gaier
Foundation models are powerful generators, but many engineering domains require structured representations that general-purpose systems handle poorly. We introduce FLOORA (Floor Layout Optimization with RL Alignment), a family of small domain-specific language (DSL) models for architectural layout generation. With specialized data and alignment, our 0.6B model outperforms much larger frontier models, achieving VLM judge win rates up to 92.0% on out-of-distribution real-world buildings and 96.0% on synthetic buildings. Human evaluations further corroborate these results, with FLOORA selected as the best model in 89.3% of evaluations. FLOORA combines a token-efficient DSL, custom tokenization, domain-specific pretraining, supervised fine-tuning (SFT), and reinforcement learning (RL) with learned human-preference and verifiable rewards. This pipeline improves architectural and geometric validity, supported by extensive empirical evaluation and ablation studies. Although focused on architecture, our results suggest that similar domain-specific recipes may be useful in other engineering domains with structured, verifiable outputs. Datasets, models, and inference code are available at https://github.com/AutodeskAILab/floora.
展开完整摘要收起摘要↓
Foundation models are powerful generators, but many engineering domains require structured representations that general-purpose systems handle poorly. We introduce FLOORA (Floor Layout Optimization with RL Alignment), a family of small domain-specific language (DSL) models for architectural layout generation. With specialized data and alignment, our 0.6B model outperforms much larger frontier models, achieving VLM judge win rates up to 92.0% on out-of-distribution real-world buildings and 96.0% on synthetic buildings. Human evaluations further corroborate these results, with FLOORA selected as the best model in 89.3% of evaluations. FLOORA combines a token-efficient DSL, custom tokenization, domain-specific pretraining, supervised fine-tuning (SFT), and reinforcement learning (RL) with learned human-preference and verifiable rewards. This pipeline improves architectural and geometric validity, supported by extensive empirical evaluation and ablation studies. Although focused on architecture, our results suggest that similar domain-specific recipes may be useful in other engineering domains with structured, verifiable outputs. Datasets, models, and inference code are available at https://github.com/AutodeskAILab/floora.
Group Relative Policy Optimization (GRPO) has become a promising approach for training large language model agents. However, its uniform assignment of trajectory-level advantages to all policy tokens fails to distinguish consequential decisions from less relevant ones, obscuring which intermediate decisions contributed to success. We introduce ProVer, a framework that targets potentially pivotal decisions for fine-grained credit assignment in agentic reinforcement learning. Given a rollout group, an agentic judge contrasts successful and failed trajectories to propose a segment potentially responsible for their divergent outcomes. Rather than directly trusting the judge's assessment, ProVer verifies the proposed segment by estimating its advantage from the difference in terminal success rates between current-policy continuations sampled before and after the segment. Positive estimates are then incorporated into the GRPO advantages of policy tokens within the proposed segment. By using model judgment only to select where to verify, ProVer grounds local credit in observed outcomes without exhaustively evaluating every intermediate state. Across ALFWorld, WebShop, and SearchQA, ProVer achieves the strongest average performance at both model scales, with relative improvements over GRPO of 9.91% and 7.12% for Qwen3.5-2B and Qwen3.5-4B, respectively. Further analyses demonstrate that informed segment selection improves policy training with modest additional generation overhead, even without a frontier-scale judge model, highlighting the effectiveness and efficiency of selectively targeting pivotal decisions for fine-grained credit assignment in agentic reinforcement learning.
展开完整摘要收起摘要↓
Group Relative Policy Optimization (GRPO) has become a promising approach for training large language model agents. However, its uniform assignment of trajectory-level advantages to all policy tokens fails to distinguish consequential decisions from less relevant ones, obscuring which intermediate decisions contributed to success. We introduce ProVer, a framework that targets potentially pivotal decisions for fine-grained credit assignment in agentic reinforcement learning. Given a rollout group, an agentic judge contrasts successful and failed trajectories to propose a segment potentially responsible for their divergent outcomes. Rather than directly trusting the judge's assessment, ProVer verifies the proposed segment by estimating its advantage from the difference in terminal success rates between current-policy continuations sampled before and after the segment. Positive estimates are then incorporated into the GRPO advantages of policy tokens within the proposed segment. By using model judgment only to select where to verify, ProVer grounds local credit in observed outcomes without exhaustively evaluating every intermediate state. Across ALFWorld, WebShop, and SearchQA, ProVer achieves the strongest average performance at both model scales, with relative improvements over GRPO of 9.91% and 7.12% for Qwen3.5-2B and Qwen3.5-4B, respectively. Further analyses demonstrate that informed segment selection improves policy training with modest additional generation overhead, even without a frontier-scale judge model, highlighting the effectiveness and efficiency of selectively targeting pivotal decisions for fine-grained credit assignment in agentic reinforcement learning.
Language models are frequently employed to evaluate other language models. An LM evaluator scoring agentic behaviors across multiple criteria is valuable, provided that its decisions align with human judgment. We call the problem of evaluating this alignment Meta-Evaluation. Tackling it directly is difficult: collecting human data is expensive, absolute scoring is hard to align, and using an LM meta-evaluator recurses the question of trustworthiness. We adopt a reformulation of meta-evaluation as a preference judgment problem: rather than comparing human and LM evaluator scores of a trajectory, we ask whether their implied preferences align. Building on this, we introduce PADMÉ, a data synthesis method that generates reliable criterion-based meta-evaluation data for agentic settings. PADMÉ uses only small language models, requires no human involvement during evaluations, and operates under a low computational budget. We build a prototype of PADMÉ and synthesize a dataset of 1,000 samples across four agentic domains and three evaluation criteria. Human validation on a 150-sample subset demonstrates that PADMÉ improves agreement with human judgment from 73% to 85% over a naive baseline. Meta-evaluating 25 common models with our dataset demonstrates the correlations between evaluation performance and scoring granularity, leniency, and model size, among other factors.
展开完整摘要收起摘要↓
Language models are frequently employed to evaluate other language models. An LM evaluator scoring agentic behaviors across multiple criteria is valuable, provided that its decisions align with human judgment. We call the problem of evaluating this alignment Meta-Evaluation. Tackling it directly is difficult: collecting human data is expensive, absolute scoring is hard to align, and using an LM meta-evaluator recurses the question of trustworthiness. We adopt a reformulation of meta-evaluation as a preference judgment problem: rather than comparing human and LM evaluator scores of a trajectory, we ask whether their implied preferences align. Building on this, we introduce PADMÉ, a data synthesis method that generates reliable criterion-based meta-evaluation data for agentic settings. PADMÉ uses only small language models, requires no human involvement during evaluations, and operates under a low computational budget. We build a prototype of PADMÉ and synthesize a dataset of 1,000 samples across four agentic domains and three evaluation criteria. Human validation on a 150-sample subset demonstrates that PADMÉ improves agreement with human judgment from 73% to 85% over a naive baseline. Meta-evaluating 25 common models with our dataset demonstrates the correlations between evaluation performance and scoring granularity, leniency, and model size, among other factors.
Generative and representation learning remain asymmetrically connected: semantic representations are used to improve diffusion generation, whereas the models' own representations are often treated as a by-product of synthesis. We ask whether diffusion models can instead be trained to learn substantially stronger semantic representations without sacrificing generation quality. SelfFlow takes a step in this direction by introducing self-supervised patch alignment into flow matching, but its main gains remain in faster convergence and improved generation. Inspired by DINO and iBOT, we extend this framework with cross-view class-token alignment to further strengthen semantic representations. Specifically, we form two independently noised, dual-timestep observations of each image and align each student class-token representation with the stop-gradient EMA-teacher target from the other observation. This objective is optimized jointly with the inherited flow-matching and local patch objectives. Notably, although the additional objective acts only on the class token, it strengthens both class-token and patch representations. Compared with a matched two-view baseline, ImageNet linear-probing accuracy improves by 9.4% using the class token and 10.1% using mean-pooled patch tokens, while frozen-backbone VOC2012 segmentation improves by 3.6 mIoU. These representation gains are achieved while maintaining comparable ImageNet generation FID. In text-to-image training, the same objective also improves generation FID, reducing it from 2.52 to 2.37 at matched checkpoints. Our results show that representation need not remain a by-product of generation or merely a tool for improving it: it can be directly optimized as a first-class capability of diffusion pretraining alongside generation.
展开完整摘要收起摘要↓
Generative and representation learning remain asymmetrically connected: semantic representations are used to improve diffusion generation, whereas the models' own representations are often treated as a by-product of synthesis. We ask whether diffusion models can instead be trained to learn substantially stronger semantic representations without sacrificing generation quality. SelfFlow takes a step in this direction by introducing self-supervised patch alignment into flow matching, but its main gains remain in faster convergence and improved generation. Inspired by DINO and iBOT, we extend this framework with cross-view class-token alignment to further strengthen semantic representations. Specifically, we form two independently noised, dual-timestep observations of each image and align each student class-token representation with the stop-gradient EMA-teacher target from the other observation. This objective is optimized jointly with the inherited flow-matching and local patch objectives. Notably, although the additional objective acts only on the class token, it strengthens both class-token and patch representations. Compared with a matched two-view baseline, ImageNet linear-probing accuracy improves by 9.4% using the class token and 10.1% using mean-pooled patch tokens, while frozen-backbone VOC2012 segmentation improves by 3.6 mIoU. These representation gains are achieved while maintaining comparable ImageNet generation FID. In text-to-image training, the same objective also improves generation FID, reducing it from 2.52 to 2.37 at matched checkpoints. Our results show that representation need not remain a by-product of generation or merely a tool for improving it: it can be directly optimized as a first-class capability of diffusion pretraining alongside generation.
D-peptides combine protease resistance with high target specificity, but computational design of D-peptide binders remains immature. Mirror-Peptidizer introduced an in silico mirror-image screening pipeline using target reflection, backbone generation, and ProteinMPNN sequence design, but its raw ProteinMPNN negative log-likelihood (NLL) ranking was not validated against measured affinities, and only 4 of 9 tested MDM2 designs bound detectably. We introduce Mirror-Score, a calibrated, inference-only scoring framework for heterochiral D-peptide/L-protein complexes, and a public benchmark of 31 crystal complexes across four target families, including 18 with literature-verified affinities. Raw ProteinMPNN NLL is not a valid affinity ranker: its pooled Spearman correlation with affinity is 0.19, and correlations reverse between MDM2/CHIP (+0.62) and gp41 (-0.70). We therefore evaluate Boltz-2 mirror-space cofolding confidence. For the complete viral-entry family (7 structures representing 3 peptides), interface predicted local distance difference test (pLDDT) achieves structure-level leave-one-out Spearman rho = 0.90 (p = 0.006) and correctly orders all three peptides by affinity, whereas NLL fails (structure-level rho = 0.18). Because only three independent chemotypes are represented, this result indicates directional consistency rather than a statistically validated predictor. Cross-family calibration does not transfer at current sample sizes, supporting family-matched calibration as the practical deployment mode. We also specify a prospective design protocol for the antimicrobial-resistance targets LasR and LecB from Pseudomonas aeruginosa, including mirrored structures, ligand-derived hotspot maps, diffusion-model-ready inputs, and Mirror-Score ranking. Code, benchmark data, structures, and analysis scripts are openly available at https://github.com/Jiadalee/Mirror-Score.
展开完整摘要收起摘要↓
D-peptides combine protease resistance with high target specificity, but computational design of D-peptide binders remains immature. Mirror-Peptidizer introduced an in silico mirror-image screening pipeline using target reflection, backbone generation, and ProteinMPNN sequence design, but its raw ProteinMPNN negative log-likelihood (NLL) ranking was not validated against measured affinities, and only 4 of 9 tested MDM2 designs bound detectably. We introduce Mirror-Score, a calibrated, inference-only scoring framework for heterochiral D-peptide/L-protein complexes, and a public benchmark of 31 crystal complexes across four target families, including 18 with literature-verified affinities. Raw ProteinMPNN NLL is not a valid affinity ranker: its pooled Spearman correlation with affinity is 0.19, and correlations reverse between MDM2/CHIP (+0.62) and gp41 (-0.70). We therefore evaluate Boltz-2 mirror-space cofolding confidence. For the complete viral-entry family (7 structures representing 3 peptides), interface predicted local distance difference test (pLDDT) achieves structure-level leave-one-out Spearman rho = 0.90 (p = 0.006) and correctly orders all three peptides by affinity, whereas NLL fails (structure-level rho = 0.18). Because only three independent chemotypes are represented, this result indicates directional consistency rather than a statistically validated predictor. Cross-family calibration does not transfer at current sample sizes, supporting family-matched calibration as the practical deployment mode. We also specify a prospective design protocol for the antimicrobial-resistance targets LasR and LecB from Pseudomonas aeruginosa, including mirrored structures, ligand-derived hotspot maps, diffusion-model-ready inputs, and Mirror-Score ranking. Code, benchmark data, structures, and analysis scripts are openly available at https://github.com/Jiadalee/Mirror-Score.
Does representation learning stop when the training loss stops improving? We study this question for matrix Muon, whose polar-normalised updates have a step length set by the gradient's rank rather than its norm. Near the edge of stability, full-batch Muon on teacher-student problems enters approximately period-2 loss oscillations that persist for thousands of steps: the cycle-mean loss stays flat or rises, yet the weights keep moving and the learned features continue to align with the teacher subspace. For linear teacher-student learning toys, we derive explicit cycle and alignment formulas and conditional plateau and decay bounds. For a population mean-field ReLU model, we prove that, under stated dimension, initialisation and small-head conditions, the leading eigenspace of the average gradient outer product (AGOP) recovers the teacher subspace exactly during a loss plateau, before the loss later drops. In all 33 ReLU, GELU and SiLU teacher configurations we study, direction-only alignment metrics show the student AGOP aligned with, or still aligning to, the teacher subspace during the period-2 oscillations; projected head refitting on selected configurations shows that the learned directions are useful for prediction, and further measurements distinguish AGOP alignment from weight-mass concentration. In deep residual ReLU students, freezing the downstream layers while the first layer trains with full-batch exact polar updates recreates a nearly flat cycle-mean loss with improving input-AGOP alignment; freezing and unfreezing switch between this plateau and loss decrease, and the effect is sensitive to momentum and to the choice of orthogonaliser.
展开完整摘要收起摘要↓
Does representation learning stop when the training loss stops improving? We study this question for matrix Muon, whose polar-normalised updates have a step length set by the gradient's rank rather than its norm. Near the edge of stability, full-batch Muon on teacher-student problems enters approximately period-2 loss oscillations that persist for thousands of steps: the cycle-mean loss stays flat or rises, yet the weights keep moving and the learned features continue to align with the teacher subspace. For linear teacher-student learning toys, we derive explicit cycle and alignment formulas and conditional plateau and decay bounds. For a population mean-field ReLU model, we prove that, under stated dimension, initialisation and small-head conditions, the leading eigenspace of the average gradient outer product (AGOP) recovers the teacher subspace exactly during a loss plateau, before the loss later drops. In all 33 ReLU, GELU and SiLU teacher configurations we study, direction-only alignment metrics show the student AGOP aligned with, or still aligning to, the teacher subspace during the period-2 oscillations; projected head refitting on selected configurations shows that the learned directions are useful for prediction, and further measurements distinguish AGOP alignment from weight-mass concentration. In deep residual ReLU students, freezing the downstream layers while the first layer trains with full-batch exact polar updates recreates a nearly flat cycle-mean loss with improving input-AGOP alignment; freezing and unfreezing switch between this plateau and loss decrease, and the effect is sensitive to momentum and to the choice of orthogonaliser.
Next-latent prediction fits a map from the current embedding to the next one. LeNEPA carries this objective to time series, replacing the stop-gradient of next-embedding prediction with the isotropy penalty of LeJEPA. A world model is a transition kernel that can be rolled out. The one-step regression identifies a conditional mean, and a mean is a kernel only in special cases. For a linear-Gaussian Markov latent, the mean transition and the innovation covariance are fixed by the one-step problem, and the open-loop squared error at horizon $K$ equals the trace of the sum of the pushed-forward innovation covariances. That error grows with $K$ after the one-step fit is exact. If the conditional mean is nonlinear, composing it is not the multi-step conditional mean. If the observation is a non-injective function of a Markov state, a memoryless one-step map does not determine future observations, while a short window can. An isotropy penalty is a function of the embedding marginal, so its partial derivative in the transition weights is zero. On a scalar autoregression with coefficient $0.9$, the one-step mean squared error is $0.998$ and the $16$-step open-loop error is $5.10$. On a hidden rotation, an eight-step window reaches $16$-step error $0.056$, while the current scalar alone reaches $0.778$. Raising the isotropy weight from $0.1$ to $10$ leaves eight-step latent error inside $[0.78,0.85]$ on three seeds.
展开完整摘要收起摘要↓
Next-latent prediction fits a map from the current embedding to the next one. LeNEPA carries this objective to time series, replacing the stop-gradient of next-embedding prediction with the isotropy penalty of LeJEPA. A world model is a transition kernel that can be rolled out. The one-step regression identifies a conditional mean, and a mean is a kernel only in special cases. For a linear-Gaussian Markov latent, the mean transition and the innovation covariance are fixed by the one-step problem, and the open-loop squared error at horizon $K$ equals the trace of the sum of the pushed-forward innovation covariances. That error grows with $K$ after the one-step fit is exact. If the conditional mean is nonlinear, composing it is not the multi-step conditional mean. If the observation is a non-injective function of a Markov state, a memoryless one-step map does not determine future observations, while a short window can. An isotropy penalty is a function of the embedding marginal, so its partial derivative in the transition weights is zero. On a scalar autoregression with coefficient $0.9$, the one-step mean squared error is $0.998$ and the $16$-step open-loop error is $5.10$. On a hidden rotation, an eight-step window reaches $16$-step error $0.056$, while the current scalar alone reaches $0.778$. Raising the isotropy weight from $0.1$ to $10$ leaves eight-step latent error inside $[0.78,0.85]$ on three seeds.
We present OLIVE (OnLine InterVEntion). At each iteration, the evolving student policy generates a new prefix, the teacher continues it autoregressively, and the student is updated using cross-entropy computed on the teacher-generated tokens. Each design choice targets a corresponding limitation of existing distillation methods: (1) sequential covariate shift in offline supervised fine-tuning (SFT) on fixed teacher trajectories, (2) fragmented supervision under prefix failure in token-level on-policy distillation (OPD), and (3) the need for access to teacher token probabilities in distribution-matching distillation. OLIVE achieves higher reasoning performance than OPD (with a top-16 KL approximation) at comparable GPU-hour cost. Our asynchronous implementation further reduces OLIVE's total training time by 23.8%. We evaluate OLIVE on both hard reasoning tasks and agentic tasks which reflects modern post-training scenarios, and it consistently outperforms existing distillation methods under the same training budget. By regenerating prefixes from the evolving student, OLIVE continues improving after offline distillation plateaus while better preserving the general capabilities and plasticity of the student. Using only text from GPT-5.4-mini, continuously training with OLIVE outperforms offline SFT from the same teacher by 13% on ScienceWorld. These results support OLIVE as an effective and efficient approach to online language-model distillation.
展开完整摘要收起摘要↓
We present OLIVE (OnLine InterVEntion). At each iteration, the evolving student policy generates a new prefix, the teacher continues it autoregressively, and the student is updated using cross-entropy computed on the teacher-generated tokens. Each design choice targets a corresponding limitation of existing distillation methods: (1) sequential covariate shift in offline supervised fine-tuning (SFT) on fixed teacher trajectories, (2) fragmented supervision under prefix failure in token-level on-policy distillation (OPD), and (3) the need for access to teacher token probabilities in distribution-matching distillation. OLIVE achieves higher reasoning performance than OPD (with a top-16 KL approximation) at comparable GPU-hour cost. Our asynchronous implementation further reduces OLIVE's total training time by 23.8%. We evaluate OLIVE on both hard reasoning tasks and agentic tasks which reflects modern post-training scenarios, and it consistently outperforms existing distillation methods under the same training budget. By regenerating prefixes from the evolving student, OLIVE continues improving after offline distillation plateaus while better preserving the general capabilities and plasticity of the student. Using only text from GPT-5.4-mini, continuously training with OLIVE outperforms offline SFT from the same teacher by 13% on ScienceWorld. These results support OLIVE as an effective and efficient approach to online language-model distillation.
作者Yuxiao Yang, Shangzhe Li, Tianrun Yu, Kaixiang Zhao, Taylor W. Killian, Weitong Zhang
On-policy distillation (OPD) trains language models using dense token-level teacher supervision on student-generated trajectories. However, its reliance on frequently refreshed student rollouts often incurs substantial generation cost. We introduce REVO, an off-policy distillation framework that improves rollout efficiency by reusing each student rollout for multi-step learner updates. REVO addresses prefix-level and current-token policy mismatch through stabilized prefix weighting and one-step resampling from the current student, which enables repeated updates without regenerating full trajectories. To prioritize informative token positions within reused rollouts, REVO uses the variance of the student-teacher log-probability ratio to quantify the remaining token-level learning signal and guide repeated optimization. Across multiple student-teacher scales, REVO with only 50 rollout iterations matches or exceeds OPD baselines trained for 200 iterations on both in-domain and cross-domain reasoning benchmarks.
展开完整摘要收起摘要↓
On-policy distillation (OPD) trains language models using dense token-level teacher supervision on student-generated trajectories. However, its reliance on frequently refreshed student rollouts often incurs substantial generation cost. We introduce REVO, an off-policy distillation framework that improves rollout efficiency by reusing each student rollout for multi-step learner updates. REVO addresses prefix-level and current-token policy mismatch through stabilized prefix weighting and one-step resampling from the current student, which enables repeated updates without regenerating full trajectories. To prioritize informative token positions within reused rollouts, REVO uses the variance of the student-teacher log-probability ratio to quantify the remaining token-level learning signal and guide repeated optimization. Across multiple student-teacher scales, REVO with only 50 rollout iterations matches or exceeds OPD baselines trained for 200 iterations on both in-domain and cross-domain reasoning benchmarks.
Flow-matching text-to-speech (TTS) models achieve high synthesis quality but require many neural function evaluations (NFEs) to integrate their generative trajectories. Recent few-step flow-map distillation approaches for TTS construct targets from numerically integrated teacher trajectories, creating a trade-off between target accuracy and training cost. We propose Local Flow-Map Distillation (LFMD), which adapts Eulerian Map Distillation to conditional TTS and avoids teacher trajectory integration during target construction. For inference, we derive a sampling schedule (TD-DP) from teacher dynamics and consistency of the learned maps, with a single cost graph supporting multiple NFE budgets without external audio-metric evaluation. Because scheduling offers no flexibility at one NFE, we refine this regime with alignment-aware temporal self-distillation using soft-DTW. Across Seed-TTS and LibriSpeech-PC, LFMD improves low-NFE synthesis over a matched integral-distillation baseline. On Seed-TTS, the refined student reaches 1.80% WER with 1-NFE, compared with 1.76% for its 32-NFE teacher.
展开完整摘要收起摘要↓
Flow-matching text-to-speech (TTS) models achieve high synthesis quality but require many neural function evaluations (NFEs) to integrate their generative trajectories. Recent few-step flow-map distillation approaches for TTS construct targets from numerically integrated teacher trajectories, creating a trade-off between target accuracy and training cost. We propose Local Flow-Map Distillation (LFMD), which adapts Eulerian Map Distillation to conditional TTS and avoids teacher trajectory integration during target construction. For inference, we derive a sampling schedule (TD-DP) from teacher dynamics and consistency of the learned maps, with a single cost graph supporting multiple NFE budgets without external audio-metric evaluation. Because scheduling offers no flexibility at one NFE, we refine this regime with alignment-aware temporal self-distillation using soft-DTW. Across Seed-TTS and LibriSpeech-PC, LFMD improves low-NFE synthesis over a matched integral-distillation baseline. On Seed-TTS, the refined student reaches 1.80% WER with 1-NFE, compared with 1.76% for its 32-NFE teacher.
作者Yizheng Huang, Wensheng Lin, Lixin Li, Qinghe Du, Wenchi Cheng, Zhu Han
As autonomous systems and embodied intelligence enter the dynamic physical world, multi-agent collaboration calls for a paradigm shift in communication design. However, existing communication paradigms overlook that agents form action understanding from their own states, environmental observations, and collaboration relations through a process that evolves as a task unfolds. Consequently, reliable bit delivery, general semantic recovery, or single-task utility optimization alone cannot ensure that heterogeneous agents form coordinated actions compatible with their own conditions from shared information during task execution. To address this gap, this paper proposes embodied semantic communication (ESC) as a paradigm that transforms information transmission into action-oriented semantic interaction. Specifically, ESC characterizes how an explicit communication link can encapsulate multimodal perceptual states, intrinsic hardware capabilities, and collaborative intents into unified actionable semantic representations, thereby enabling heterogeneous receiving agents to parse, align, and ground them in local motor control. This paper clarifies the conceptual boundary, system characteristics, and environment-constrained technical pathways of ESC. It maps the underlying mathematical tools, including semantic information theory, world models, and multi-agent decision theory. Finally, this paper summarizes key open challenges, including measurable semantic reliability, ambiguity-triggered interaction under dynamic environments and tasks, and bandwidth-adaptive semantic transmission, outlining a roadmap for collective embodied networks.
展开完整摘要收起摘要↓
As autonomous systems and embodied intelligence enter the dynamic physical world, multi-agent collaboration calls for a paradigm shift in communication design. However, existing communication paradigms overlook that agents form action understanding from their own states, environmental observations, and collaboration relations through a process that evolves as a task unfolds. Consequently, reliable bit delivery, general semantic recovery, or single-task utility optimization alone cannot ensure that heterogeneous agents form coordinated actions compatible with their own conditions from shared information during task execution. To address this gap, this paper proposes embodied semantic communication (ESC) as a paradigm that transforms information transmission into action-oriented semantic interaction. Specifically, ESC characterizes how an explicit communication link can encapsulate multimodal perceptual states, intrinsic hardware capabilities, and collaborative intents into unified actionable semantic representations, thereby enabling heterogeneous receiving agents to parse, align, and ground them in local motor control. This paper clarifies the conceptual boundary, system characteristics, and environment-constrained technical pathways of ESC. It maps the underlying mathematical tools, including semantic information theory, world models, and multi-agent decision theory. Finally, this paper summarizes key open challenges, including measurable semantic reliability, ambiguity-triggered interaction under dynamic environments and tasks, and bandwidth-adaptive semantic transmission, outlining a roadmap for collective embodied networks.
作者Jie Zhang, Andrei Baroian, Jan N. van Rijn, Avital Shafran, Florian Tramèr
Large language models are vulnerable to prompt injection attacks, where third-party adversarial content can hijack the model's behavior. In this paper, we study the role played by the adversarial data's input modality, and identify a systematic asymmetry: multimodal LLMs are more likely to follow adversarial instruction when they appear as text than when the same instruction is delivered through a non-textual channel (e.g., as an image). We hypothesize that this modality gap arises from text-centric instruction tuning, which teaches models to obey textual instructions while treating other modalities mainly as content to parse or describe. We then demonstrate how this gap can be turned into a training-free defense, by rendering all untrusted payloads as typographic images (or audio) before they reach the model. Across ten models and two prompt injection benchmarks (DirectInject and AgentDojo) we show that our defense Pictionary consistently reduces attack success rates even against the strongest adaptive attacks and human red teamers, while largely preserving benign utility. We further show that benign fine-tuning on image-rendered instructions erodes the modality gap, tracing it to the text-centric instruction-tuning distribution.
展开完整摘要收起摘要↓
Large language models are vulnerable to prompt injection attacks, where third-party adversarial content can hijack the model's behavior. In this paper, we study the role played by the adversarial data's input modality, and identify a systematic asymmetry: multimodal LLMs are more likely to follow adversarial instruction when they appear as text than when the same instruction is delivered through a non-textual channel (e.g., as an image). We hypothesize that this modality gap arises from text-centric instruction tuning, which teaches models to obey textual instructions while treating other modalities mainly as content to parse or describe. We then demonstrate how this gap can be turned into a training-free defense, by rendering all untrusted payloads as typographic images (or audio) before they reach the model. Across ten models and two prompt injection benchmarks (DirectInject and AgentDojo) we show that our defense Pictionary consistently reduces attack success rates even against the strongest adaptive attacks and human red teamers, while largely preserving benign utility. We further show that benign fine-tuning on image-rendered instructions erodes the modality gap, tracing it to the text-centric instruction-tuning distribution.
Open-world aerial object-goal search is a foundational yet challenging task, requiring aerial agents to autonomously explore large-scale, unstructured three-dimensional environments and reach target objects specified by semantic descriptions or reference images, rather than following route-specific instructions. However, research in this task remains at a nascent stage and relies on small, environment-specific benchmarks with heterogeneous action spaces and data formats. These limitations hinder large-scale training and cross-benchmark evaluation, constraining the scalability and generalizability of aerial agents. To address this problem, we propose AerialDojo-200K, a large-scale benchmark suite for open-world aerial object-goal search, with 3 times as many scenes and 18.7 times as many task instances as the largest existing benchmark for this task. Specifically, we construct 42 simulation scenes spanning four scene families and 21 scene types, including 18 urban, 12 natural, six infrastructure, and six disaster scenes. To ensure data quality, 12 annotators spent two months manually annotating 109 landmarks, 2099 target objects, and 2099 object anchors across these scenes. We further construct 205,732 task instances, comprising over 100K semantic-goal and over 100K image-goal instances across Base, Standard, and Long-Horizon settings. Each task instance includes a collision-free reference trajectory and corresponding multi-view video recordings. We also develop a unified evaluation framework with a scene partition comprising 21 in-distribution scenes and 21 out-of-distribution scenes. Finally, our evaluation of five open-source and four closed-source multimodal large language models reveals that there is still a long way to go toward achieving general-purpose aerial agents. All can be found at https://fengtt42.github.io/AerialDojo/.
展开完整摘要收起摘要↓
Open-world aerial object-goal search is a foundational yet challenging task, requiring aerial agents to autonomously explore large-scale, unstructured three-dimensional environments and reach target objects specified by semantic descriptions or reference images, rather than following route-specific instructions. However, research in this task remains at a nascent stage and relies on small, environment-specific benchmarks with heterogeneous action spaces and data formats. These limitations hinder large-scale training and cross-benchmark evaluation, constraining the scalability and generalizability of aerial agents. To address this problem, we propose AerialDojo-200K, a large-scale benchmark suite for open-world aerial object-goal search, with 3 times as many scenes and 18.7 times as many task instances as the largest existing benchmark for this task. Specifically, we construct 42 simulation scenes spanning four scene families and 21 scene types, including 18 urban, 12 natural, six infrastructure, and six disaster scenes. To ensure data quality, 12 annotators spent two months manually annotating 109 landmarks, 2099 target objects, and 2099 object anchors across these scenes. We further construct 205,732 task instances, comprising over 100K semantic-goal and over 100K image-goal instances across Base, Standard, and Long-Horizon settings. Each task instance includes a collision-free reference trajectory and corresponding multi-view video recordings. We also develop a unified evaluation framework with a scene partition comprising 21 in-distribution scenes and 21 out-of-distribution scenes. Finally, our evaluation of five open-source and four closed-source multimodal large language models reveals that there is still a long way to go toward achieving general-purpose aerial agents. All can be found at https://fengtt42.github.io/AerialDojo/.
作者Xiaoyu Wu, Weihang Guo, Yifei Wang, Xinze Feng, Lydia E. Kavraki, Zhiwei Steven Wu
Standard video generators do not natively compact historical context into reusable memory tokens. As generation continues, the growing history makes it increasingly difficult to retain information from earlier frames due to long-context degradation. Key-frame-based approaches address this challenge by retaining selected past frames, but can discard information needed for future generation. Rather than relying on frame selection alone, we study whether a frozen video generator can supply the supervision needed to learn a compact representation of the history. We propose Prediction-Aligned Context Compaction (PACC), which uses a learned compressor to aggregate information across past frames into compact memory tokens. We train the compressor through on-policy distillation, using the same frozen generator both as a student when conditioned on compressed memory and as a teacher when conditioned on the full history. The student generates continuations, while the teacher provides targets for the same noisy inputs at each denoising step. Only the compressor is updated to align the student's predictions with these targets. We evaluate PACC on MBench, which jointly measures memory-event coverage and consistency. PACC outperforms the strongest baseline by 6.63 points on Causal-rCM and 3.19 points on Causal Forcing. Evaluation on VBench-Long using MovieGen prompts further shows that PACC produces minute-long videos with generation quality competitive with baselines. Together, these results show that learning to compact historical context can improve long-video memory without modifying the underlying generator.
展开完整摘要收起摘要↓
Standard video generators do not natively compact historical context into reusable memory tokens. As generation continues, the growing history makes it increasingly difficult to retain information from earlier frames due to long-context degradation. Key-frame-based approaches address this challenge by retaining selected past frames, but can discard information needed for future generation. Rather than relying on frame selection alone, we study whether a frozen video generator can supply the supervision needed to learn a compact representation of the history. We propose Prediction-Aligned Context Compaction (PACC), which uses a learned compressor to aggregate information across past frames into compact memory tokens. We train the compressor through on-policy distillation, using the same frozen generator both as a student when conditioned on compressed memory and as a teacher when conditioned on the full history. The student generates continuations, while the teacher provides targets for the same noisy inputs at each denoising step. Only the compressor is updated to align the student's predictions with these targets. We evaluate PACC on MBench, which jointly measures memory-event coverage and consistency. PACC outperforms the strongest baseline by 6.63 points on Causal-rCM and 3.19 points on Causal Forcing. Evaluation on VBench-Long using MovieGen prompts further shows that PACC produces minute-long videos with generation quality competitive with baselines. Together, these results show that learning to compact historical context can improve long-video memory without modifying the underlying generator.
Most videos capture only a narrow field of view and provide no spatial audio, limiting the sense of immersion they can provide. Recent video generation models can expand perspective videos into panoramic ones, but do not provide the corresponding spatial soundscape. Without spatially consistent audio, these expanded visual worlds remain incomplete. This paper presents OmniDream, a training-free framework that transforms a silent monocular video into an immersive audiovisual experience, where viewers can freely look around while sounds remain spatially aligned with the visual scene. At the core of OmniDream is an object-centric audio representation that disentangles each sound source's intrinsic audio content from its scene-dependent acoustic effects, enabling independent audio generation, physics-based simulation of propagation effects, and flexible spatial audio rendering. Experiments show improved audio-visual alignment, spatial correctness, and perceptual immersiveness over baselines. Examples are available on https://huggingface.co/spaces/CuriousAlien000/spatial-audio-360-demo
展开完整摘要收起摘要↓
Most videos capture only a narrow field of view and provide no spatial audio, limiting the sense of immersion they can provide. Recent video generation models can expand perspective videos into panoramic ones, but do not provide the corresponding spatial soundscape. Without spatially consistent audio, these expanded visual worlds remain incomplete. This paper presents OmniDream, a training-free framework that transforms a silent monocular video into an immersive audiovisual experience, where viewers can freely look around while sounds remain spatially aligned with the visual scene. At the core of OmniDream is an object-centric audio representation that disentangles each sound source's intrinsic audio content from its scene-dependent acoustic effects, enabling independent audio generation, physics-based simulation of propagation effects, and flexible spatial audio rendering. Experiments show improved audio-visual alignment, spatial correctness, and perceptual immersiveness over baselines. Examples are available on https://huggingface.co/spaces/CuriousAlien000/spatial-audio-360-demo
作者Antonio Pariente, Ignacio Boero, Nikolai Matni, Alejandro Ribeiro
World models jointly learn latent representations and dynamics that predict how high-dimensional observations evolve under actions. In this work, we propose a JEPA-style world model in which, rather than learning arbitrary latent dynamics, we restrict them to follow a bilinear parameterization. This structure enables efficient planning and control while shifting the modeling burden onto the encoder, encouraging richer representations that expose the controllable geometry of the system. In particular, this structured parameterization allows us to structurally enforce action recoverability, thereby preventing representation collapse by construction. Although prescribing a bilinear parametrization may appear restrictive, we show that a broad class of nonlinear dynamical systems admits a transformation under which the dynamics become bilinear. Empirically, we show across standard 2D and 3D control tasks that representations with bilinear-parameterized dynamics can be learned directly from high-dimensional observations, reducing planning time by nearly three orders of magnitude while retaining or even improving control accuracy. We also propose more demanding regimes of longer-horizon planning and real-time control, and demonstrate that our method succeeds in both, moving JEPA-style world models beyond short-horizon offline planning.
展开完整摘要收起摘要↓
World models jointly learn latent representations and dynamics that predict how high-dimensional observations evolve under actions. In this work, we propose a JEPA-style world model in which, rather than learning arbitrary latent dynamics, we restrict them to follow a bilinear parameterization. This structure enables efficient planning and control while shifting the modeling burden onto the encoder, encouraging richer representations that expose the controllable geometry of the system. In particular, this structured parameterization allows us to structurally enforce action recoverability, thereby preventing representation collapse by construction. Although prescribing a bilinear parametrization may appear restrictive, we show that a broad class of nonlinear dynamical systems admits a transformation under which the dynamics become bilinear. Empirically, we show across standard 2D and 3D control tasks that representations with bilinear-parameterized dynamics can be learned directly from high-dimensional observations, reducing planning time by nearly three orders of magnitude while retaining or even improving control accuracy. We also propose more demanding regimes of longer-horizon planning and real-time control, and demonstrate that our method succeeds in both, moving JEPA-style world models beyond short-horizon offline planning.
Unified multimodal models (UMMs) combine image generation and visual understanding in a shared backbone. Since generation and understanding are inverse tasks, recent studies self-train UMMs by letting the two branches cooperatively supervise each other. We introduce MATE (Mutually Adversarial self-Training with Evolving data), a reinforcement-learning-based post-training framework in which the two branches instead challenge each other, and the challenges evolve as the model trains. MATE lets generation and understanding take turns to be challenger and solver. Given an image, the understanding branch proposes several candidate descriptions that the generation branch must turn back into similar images, and vice versa. The candidates are screened for consistency with the image or prompt they were proposed from, and the solver is trained on the candidate it handles worst. The adversary thus comes from the model's own outputs, and no separate adversary is trained. Moreover, the candidates that defeat one branch become the sources of the next challenges to the other in the next epoch, which keeps the challenges evolving with the model and turns the training into self-play in data space. On Janus-Pro-1B, MATE improves GenEval by 2.4 points, DPG-Bench by 1.7 points, and the average over nine understanding benchmarks by 0.7 points, while strengthening consistency across repeated image-text cycles.
展开完整摘要收起摘要↓
Unified multimodal models (UMMs) combine image generation and visual understanding in a shared backbone. Since generation and understanding are inverse tasks, recent studies self-train UMMs by letting the two branches cooperatively supervise each other. We introduce MATE (Mutually Adversarial self-Training with Evolving data), a reinforcement-learning-based post-training framework in which the two branches instead challenge each other, and the challenges evolve as the model trains. MATE lets generation and understanding take turns to be challenger and solver. Given an image, the understanding branch proposes several candidate descriptions that the generation branch must turn back into similar images, and vice versa. The candidates are screened for consistency with the image or prompt they were proposed from, and the solver is trained on the candidate it handles worst. The adversary thus comes from the model's own outputs, and no separate adversary is trained. Moreover, the candidates that defeat one branch become the sources of the next challenges to the other in the next epoch, which keeps the challenges evolving with the model and turns the training into self-play in data space. On Janus-Pro-1B, MATE improves GenEval by 2.4 points, DPG-Bench by 1.7 points, and the average over nine understanding benchmarks by 0.7 points, while strengthening consistency across repeated image-text cycles.
Instruction-guided image editing should change what the instruction names and leave the rest of the image untouched. In dual classifier-free guidance (CFG), an editor combines two directions at every denoising step, one that pushes toward the instructed edit and one that pulls back toward the source image, using global weights. We introduce Attention-Scoped Guidance (ASG), a sampler wrapper that makes these weights spatial. It reads a soft support map from the instruction attention that the editor already computes, then weakens text guidance where support is low and strengthens image anchoring where support is high. The wrapper requires no training, no external mask, and no additional network evaluation. On the full MagicBrush and PIE-Bench++ splits, ASG improves preservation-oriented metrics, leading three of four MagicBrush metrics and PIE-Bench++ background PSNR. A dose-matched control that removes the spatial placement loses up to 0.73 CLIP on PIE-Bench++, confirming that the spatial allocation itself carries the gain.
展开完整摘要收起摘要↓
Instruction-guided image editing should change what the instruction names and leave the rest of the image untouched. In dual classifier-free guidance (CFG), an editor combines two directions at every denoising step, one that pushes toward the instructed edit and one that pulls back toward the source image, using global weights. We introduce Attention-Scoped Guidance (ASG), a sampler wrapper that makes these weights spatial. It reads a soft support map from the instruction attention that the editor already computes, then weakens text guidance where support is low and strengthens image anchoring where support is high. The wrapper requires no training, no external mask, and no additional network evaluation. On the full MagicBrush and PIE-Bench++ splits, ASG improves preservation-oriented metrics, leading three of four MagicBrush metrics and PIE-Bench++ background PSNR. A dose-matched control that removes the spatial placement loses up to 0.73 CLIP on PIE-Bench++, confirming that the spatial allocation itself carries the gain.
Pixel-space diffusion Transformers (DiTs) directly operate on high-dimensional visual data, yet their hidden representations typically undergo uniform refinement across depth. Natural images, however, are inherently organized at different levels of granularity. Global structure can often be represented compactly, whereas local textures and fine details require richer representations. Motivated by this, we introduce heterogeneous refinement in pixel-space DiTs, assigning different feature groups distinct refinement budgets across depth. Consequently, an ordered feature specialization emerges: sparsely refined features predominantly encode global visual structure, whereas more frequently refined features increasingly specialize toward localized, high-frequency details. We refer to these two groups as persistent and active features, respectively. Building on this emergent specialization, we introduce Persistence Forcing (PerF), which explicitly exploits this persistent--active feature organization for pixel-space image generation. This enables persistent features to continuously condition actively refined features, allowing stable global information to guide the ongoing refinement of finer visual details. During generative sampling, this interaction further induces a meaningful guidance direction that promotes coherent global structure and naturally complements classifier-free guidance. On ImageNet $256\times256$, PerF-L achieves FID of $1.91$, approaching $1.86$ of JiT-H with only half the parameters, while PerF-H further achieves FID of $1.63$ and $1.76$ on ImageNet $256\times256$ and $512\times512$, respectively.
展开完整摘要收起摘要↓
Pixel-space diffusion Transformers (DiTs) directly operate on high-dimensional visual data, yet their hidden representations typically undergo uniform refinement across depth. Natural images, however, are inherently organized at different levels of granularity. Global structure can often be represented compactly, whereas local textures and fine details require richer representations. Motivated by this, we introduce heterogeneous refinement in pixel-space DiTs, assigning different feature groups distinct refinement budgets across depth. Consequently, an ordered feature specialization emerges: sparsely refined features predominantly encode global visual structure, whereas more frequently refined features increasingly specialize toward localized, high-frequency details. We refer to these two groups as persistent and active features, respectively. Building on this emergent specialization, we introduce Persistence Forcing (PerF), which explicitly exploits this persistent--active feature organization for pixel-space image generation. This enables persistent features to continuously condition actively refined features, allowing stable global information to guide the ongoing refinement of finer visual details. During generative sampling, this interaction further induces a meaningful guidance direction that promotes coherent global structure and naturally complements classifier-free guidance. On ImageNet $256\times256$, PerF-L achieves FID of $1.91$, approaching $1.86$ of JiT-H with only half the parameters, while PerF-H further achieves FID of $1.63$ and $1.76$ on ImageNet $256\times256$ and $512\times512$, respectively.
作者Hua, Xu, Dongxin Li, Gwen Yidou-Weng, Guy Van den Broeck, Wei Wang, Anji Liu
Discrete diffusion models generate sequences by iteratively resolving multiple tokens in parallel, offering a flexible alternative to left-to-right generation. However, guiding this process with a sequence-level objective is difficult because the value of one unresolved token depends on the other tokens with which it can form a high-reward sequence. Enumerating all such completions makes the whole guidance computation grow exponentially with the number of unresolved positions. We introduce COFFEE, a plug-and-play framework that avoids this enumeration by separating sequence dependence from the objective. At each diffusion step, a target-free carrier absorbs the marginal token distributions predicted by the denoiser to construct a joint model over the unresolved tokens, while a compiled finite-state model records how their combinations affect the sequence-level preference. Pairing their states allows COFFEE to transfer global preferences to unresolved positions and sample a clean reconstruction without retraining the diffusion model. The same framework supports explicit hard constraints and learned soft objectives. We evaluate COFFEE across multiple symbolic, language, and biological benchmarks, where it achieves strong control results with task-dependent quality and diversity trade-offs. By making objectives available to inference rather than only evaluation, COFFEE brings joint conditioning, completion-weighted guidance, and optimization-based constraints into pretrained neural generation, showing the potential of neural-symbolic methods in diffusion guidance.
展开完整摘要收起摘要↓
Discrete diffusion models generate sequences by iteratively resolving multiple tokens in parallel, offering a flexible alternative to left-to-right generation. However, guiding this process with a sequence-level objective is difficult because the value of one unresolved token depends on the other tokens with which it can form a high-reward sequence. Enumerating all such completions makes the whole guidance computation grow exponentially with the number of unresolved positions. We introduce COFFEE, a plug-and-play framework that avoids this enumeration by separating sequence dependence from the objective. At each diffusion step, a target-free carrier absorbs the marginal token distributions predicted by the denoiser to construct a joint model over the unresolved tokens, while a compiled finite-state model records how their combinations affect the sequence-level preference. Pairing their states allows COFFEE to transfer global preferences to unresolved positions and sample a clean reconstruction without retraining the diffusion model. The same framework supports explicit hard constraints and learned soft objectives. We evaluate COFFEE across multiple symbolic, language, and biological benchmarks, where it achieves strong control results with task-dependent quality and diversity trade-offs. By making objectives available to inference rather than only evaluation, COFFEE brings joint conditioning, completion-weighted guidance, and optimization-based constraints into pretrained neural generation, showing the potential of neural-symbolic methods in diffusion guidance.