作者Mohamed Amine Ketata, Maximilian Schambach, Stephan Günnemann
Generative models for tabular data are typically trained separately for each dataset, limiting knowledge transfer and requiring the storage of many specialized models. In this paper, we introduce CDMD, a tabular diffusion model trained jointly across heterogeneous datasets with different schemas and variable numbers of numerical and categorical features. Unlike existing cross-dataset tabular diffusion models that operate in continuous representation spaces, CDMD defines diffusion directly over the mixed-type feature space and is trained end-to-end. To accommodate heterogeneous categorical domains, we introduce a schema-restricted reverse-process parameterization for masked diffusion models, in which the output space dynamically adapts to each feature's vocabulary. We then compose numerical and categorical feature-level diffusion processes into a schema-dependent row-level process. A shared schema-aware Transformer denoiser captures dependencies between features and parameterizes the reverse process across varying schemas. On seven real-world datasets, a single jointly trained CDMD achieves the highest average generation quality among strong single-dataset and cross-dataset baselines, while using substantially fewer total parameters than the collection of separately trained models. Furthermore, pre-training on a corpus of 337 datasets improves generation on previously unseen datasets under both limited target data and limited adaptation epochs. These results demonstrate the potential of direct mixed-type diffusion for shared and transferable tabular data generation. Our code is available at https://github.com/ketatam/cdmd.
展开完整摘要收起摘要↓
Generative models for tabular data are typically trained separately for each dataset, limiting knowledge transfer and requiring the storage of many specialized models. In this paper, we introduce CDMD, a tabular diffusion model trained jointly across heterogeneous datasets with different schemas and variable numbers of numerical and categorical features. Unlike existing cross-dataset tabular diffusion models that operate in continuous representation spaces, CDMD defines diffusion directly over the mixed-type feature space and is trained end-to-end. To accommodate heterogeneous categorical domains, we introduce a schema-restricted reverse-process parameterization for masked diffusion models, in which the output space dynamically adapts to each feature's vocabulary. We then compose numerical and categorical feature-level diffusion processes into a schema-dependent row-level process. A shared schema-aware Transformer denoiser captures dependencies between features and parameterizes the reverse process across varying schemas. On seven real-world datasets, a single jointly trained CDMD achieves the highest average generation quality among strong single-dataset and cross-dataset baselines, while using substantially fewer total parameters than the collection of separately trained models. Furthermore, pre-training on a corpus of 337 datasets improves generation on previously unseen datasets under both limited target data and limited adaptation epochs. These results demonstrate the potential of direct mixed-type diffusion for shared and transferable tabular data generation. Our code is available at https://github.com/ketatam/cdmd.
Adapting a pretrained generative model to an arbitrary preference expressed as a utility function underlies reward alignment, guided design, and constraint satisfaction, enabling diverse applications. Existing fine-tuning methods trade off generality against computational cost: they either restrict the family class of supported preferences to keep optimization simple or preserve generality at the expense of efficiency. We introduce Fenchel Tilt Flow Control (FTFC), which decouples utility optimization from generative-model fitting. FTFC first optimizes for a target distribution by jointly fitting an effective reward and density-ratio weights on pretrained samples. Method combines the utility's variational structure with Fenchel duality, supporting general $f$-divergence penalties that determine how rewards are transformed into an distribution-correction weights. These weights are then frozen and used to modify a diffusion or flow model in a single stage of importance-weighted denoising or flow matching, without differentiating through sampling trajectories. We establish exact duality for concave utilities under suitable conditions and show that weighted fitting reproduces the optimal target distribution for a given utility. Across image and molecule generation benchmarks, FTFC improves over baselines on diverse preference functions, while also being up to $20\times$ more efficient. roposed method enables adaptation beyond expected-reward maximization without complex optimization, while preserving robustness for more general class of the utility functions compared to baselines.
展开完整摘要收起摘要↓
Adapting a pretrained generative model to an arbitrary preference expressed as a utility function underlies reward alignment, guided design, and constraint satisfaction, enabling diverse applications. Existing fine-tuning methods trade off generality against computational cost: they either restrict the family class of supported preferences to keep optimization simple or preserve generality at the expense of efficiency. We introduce Fenchel Tilt Flow Control (FTFC), which decouples utility optimization from generative-model fitting. FTFC first optimizes for a target distribution by jointly fitting an effective reward and density-ratio weights on pretrained samples. Method combines the utility's variational structure with Fenchel duality, supporting general $f$-divergence penalties that determine how rewards are transformed into an distribution-correction weights. These weights are then frozen and used to modify a diffusion or flow model in a single stage of importance-weighted denoising or flow matching, without differentiating through sampling trajectories. We establish exact duality for concave utilities under suitable conditions and show that weighted fitting reproduces the optimal target distribution for a given utility. Across image and molecule generation benchmarks, FTFC improves over baselines on diverse preference functions, while also being up to $20\times$ more efficient. roposed method enables adaptation beyond expected-reward maximization without complex optimization, while preserving robustness for more general class of the utility functions compared to baselines.
作者Stipe Frković, Metod Jazbec, Christian A. Naesseth
Masked diffusion language models (dLLMs) have shown strong potential for faster inference through parallel token generation when combined with confidence-based samplers. However, recent work has shown that such methods can defer unmasking high-entropy fork positions at which multiple plausible continuations exist. This results in reduced generation diversity, as shown by worse pass@k scaling, and limits gains obtainable from RL post-training. To avoid this flexibility trap, prior work advocated for autoregressive (AR) sampling. Here, we show that discarding confidence-based sampling is unnecessary and, once inference cost is taken into account, wasteful. We first propose Fork-dLLM, a simple hybrid sampler that uses AR-style ordering only at uncertain fallback steps while retaining parallel generation otherwise. We then extend the same principle to post-training with ForkGRPO, which uses Fork-dLLM rollouts and applies the GRPO objective only at fallback steps, preserving exact policy-likelihood ratios while substantially reducing rollout and optimization cost. In our experiments, Fork-dLLM matches the strong pass@k scaling of AR sampling while being 2-3x more efficient, and ForkGRPO achieves downstream performance comparable to or better than AR-based GRPO baselines at a substantially lower training cost.
展开完整摘要收起摘要↓
Masked diffusion language models (dLLMs) have shown strong potential for faster inference through parallel token generation when combined with confidence-based samplers. However, recent work has shown that such methods can defer unmasking high-entropy fork positions at which multiple plausible continuations exist. This results in reduced generation diversity, as shown by worse pass@k scaling, and limits gains obtainable from RL post-training. To avoid this flexibility trap, prior work advocated for autoregressive (AR) sampling. Here, we show that discarding confidence-based sampling is unnecessary and, once inference cost is taken into account, wasteful. We first propose Fork-dLLM, a simple hybrid sampler that uses AR-style ordering only at uncertain fallback steps while retaining parallel generation otherwise. We then extend the same principle to post-training with ForkGRPO, which uses Fork-dLLM rollouts and applies the GRPO objective only at fallback steps, preserving exact policy-likelihood ratios while substantially reducing rollout and optimization cost. In our experiments, Fork-dLLM matches the strong pass@k scaling of AR sampling while being 2-3x more efficient, and ForkGRPO achieves downstream performance comparable to or better than AR-based GRPO baselines at a substantially lower training cost.
Diffusion distillation is widely adopted to accelerate sampling, and the resulting few-step models are broadly believed to match or even surpass their multi-step teachers in generation. However, standard evaluations such as GenEval2 typically draw only one sample per prompt, so improved scores may fail to reveal losses in distribution coverage. We therefore revisit whether distilled models truly match their teachers beyond single-draw performance using \textbf{pass@$\mathbf{k}$}, which measures the probability that at least one of $k$ independent samples satisfies a quality criterion. At $k{=}1$, pass@$k$ reduces to standard single-draw evaluation. As $k$ grows, the curve reveals whether additional draws find genuinely different successes or merely revisit the same modes, directly exposing how broadly a model covers the space of valid outputs. We first show that classifier-free guidance (CFG), whose quality--coverage tradeoff is well established, is the clearest case: higher guidance improves pass@$1$, but its advantage shrinks and reverses at larger $k$. Applying pass@$k$ to few-step distilled models, we find the same tradeoff splits along training objectives: distribution-matching objectives concentrate the student's output distribution, boosting early-hit rates while eroding large-budget coverage, whereas consistency and trajectory-based objectives better preserve the teacher's coverage even at large $k$. We further show that this tradeoff extends to few-step causal video generation. Our findings reveal a previously overlooked cost of diffusion distillation: across both image and video generation, the choice of training objective fundamentally determines whether a few-step model inherits its teacher's distribution coverage or trades it away for single-draw quality.
展开完整摘要收起摘要↓
Diffusion distillation is widely adopted to accelerate sampling, and the resulting few-step models are broadly believed to match or even surpass their multi-step teachers in generation. However, standard evaluations such as GenEval2 typically draw only one sample per prompt, so improved scores may fail to reveal losses in distribution coverage. We therefore revisit whether distilled models truly match their teachers beyond single-draw performance using \textbf{pass@$\mathbf{k}$}, which measures the probability that at least one of $k$ independent samples satisfies a quality criterion. At $k{=}1$, pass@$k$ reduces to standard single-draw evaluation. As $k$ grows, the curve reveals whether additional draws find genuinely different successes or merely revisit the same modes, directly exposing how broadly a model covers the space of valid outputs. We first show that classifier-free guidance (CFG), whose quality--coverage tradeoff is well established, is the clearest case: higher guidance improves pass@$1$, but its advantage shrinks and reverses at larger $k$. Applying pass@$k$ to few-step distilled models, we find the same tradeoff splits along training objectives: distribution-matching objectives concentrate the student's output distribution, boosting early-hit rates while eroding large-budget coverage, whereas consistency and trajectory-based objectives better preserve the teacher's coverage even at large $k$. We further show that this tradeoff extends to few-step causal video generation. Our findings reveal a previously overlooked cost of diffusion distillation: across both image and video generation, the choice of training objective fundamentally determines whether a few-step model inherits its teacher's distribution coverage or trades it away for single-draw quality.
作者Xu Huang, Ye Huang, Zijun Liao, Yuwei Niu, Xiaojie Li, Menghan Zhou, De Wen Soh, Xiaotong Li, Daquan Zhou
High-compression tokenizers are essential for scaling latent image generative models. However, aggressive compression creates a fundamental tradeoff between reconstruction fidelity and generation efficiency: high compression image encoder always increases the learning difficulty of diffusion training, resulting in slow model convergence. Recent representation autoencoders speed up the diffusion training by improving the latent feature's expressive capability by replacing VAE encoders with pretrained semantic encoders, yet they are typically limited to moderate compression and lose pixel-level details necessary for faithful reconstruction. To achieve both high compression and fast diffusion training, we propose DC-SAE, a Decoupled Compact Semantic Autoencoder designed for high-compression image generation with accelerated diffusion model convergence. DC-SAE consists of two key components: (1) a macro-level architecture design that leverages semantic encoders to enable higher compression ratios, and (2) a pixel-level encoder that preserves low-level details, ensuring high-fidelity image reconstruction. We empirically demonstrate that DC-SAE performs strongly on image generation tasks, achieving both compact latent representations and efficient training dynamics. Specifically, on the ImageNet dataset with $512 \times 512$ resolution, DC-SAE achieves $32\times$ spatial compression, with 29.79 PSNR and 3.37 gFID, substantially outperforming the previous state-of-the-art high-compression tokenizer baselines DC-AE by 13.5% and 54.9% on PSNR and gFID, respectively, maintaining comparable throughput and faster diffusion model training convergence. Beyond class-conditional generation, a $1.6$B-parameter DiT using DC-SAE achieves 0.84 on GenEval and 86.007 on DPG-Bench for text-to-image generation at $1024\times1024$ resolution.
展开完整摘要收起摘要↓
High-compression tokenizers are essential for scaling latent image generative models. However, aggressive compression creates a fundamental tradeoff between reconstruction fidelity and generation efficiency: high compression image encoder always increases the learning difficulty of diffusion training, resulting in slow model convergence. Recent representation autoencoders speed up the diffusion training by improving the latent feature's expressive capability by replacing VAE encoders with pretrained semantic encoders, yet they are typically limited to moderate compression and lose pixel-level details necessary for faithful reconstruction. To achieve both high compression and fast diffusion training, we propose DC-SAE, a Decoupled Compact Semantic Autoencoder designed for high-compression image generation with accelerated diffusion model convergence. DC-SAE consists of two key components: (1) a macro-level architecture design that leverages semantic encoders to enable higher compression ratios, and (2) a pixel-level encoder that preserves low-level details, ensuring high-fidelity image reconstruction. We empirically demonstrate that DC-SAE performs strongly on image generation tasks, achieving both compact latent representations and efficient training dynamics. Specifically, on the ImageNet dataset with $512 \times 512$ resolution, DC-SAE achieves $32\times$ spatial compression, with 29.79 PSNR and 3.37 gFID, substantially outperforming the previous state-of-the-art high-compression tokenizer baselines DC-AE by 13.5% and 54.9% on PSNR and gFID, respectively, maintaining comparable throughput and faster diffusion model training convergence. Beyond class-conditional generation, a $1.6$B-parameter DiT using DC-SAE achieves 0.84 on GenEval and 86.007 on DPG-Bench for text-to-image generation at $1024\times1024$ resolution.
Offline multi-objective optimization (MOO) seeks solutions with better objective trade-offs using only a fixed dataset, without querying the objectives. Diffusion models trained on such data have emerged as a promising approach, but their samples are not inherently better than the data and must be steered toward the Pareto front. Existing methods guide or condition every sampling step. We instead act on the initial noise and leave the sampling process unchanged. Across Off-MOO-Bench, we observe that the objectives, as functions of the noise, are sensitive to only a few directions. We estimate these directions once per task via a Recursive Feature Machine using function values alone, and a small cache serves every trade-off, so each candidate costs one noise displacement and one ODE solve. We prove that this displacement increases the learned scalarized objective in expectation, and that sweeping trade-offs recovers the flow's attainable front up to proxy and steering errors. With additional guidance, for which we introduce novel data-adaptive and Pareto-aware operators, our method attains the best average hypervolume rank among generative methods on 47 tasks, at comparable or lower sampling cost. Steering alone outranks the best prior generative method at a fraction of its sampling cost.
展开完整摘要收起摘要↓
Offline multi-objective optimization (MOO) seeks solutions with better objective trade-offs using only a fixed dataset, without querying the objectives. Diffusion models trained on such data have emerged as a promising approach, but their samples are not inherently better than the data and must be steered toward the Pareto front. Existing methods guide or condition every sampling step. We instead act on the initial noise and leave the sampling process unchanged. Across Off-MOO-Bench, we observe that the objectives, as functions of the noise, are sensitive to only a few directions. We estimate these directions once per task via a Recursive Feature Machine using function values alone, and a small cache serves every trade-off, so each candidate costs one noise displacement and one ODE solve. We prove that this displacement increases the learned scalarized objective in expectation, and that sweeping trade-offs recovers the flow's attainable front up to proxy and steering errors. With additional guidance, for which we introduce novel data-adaptive and Pareto-aware operators, our method attains the best average hypervolume rank among generative methods on 47 tasks, at comparable or lower sampling cost. Steering alone outranks the best prior generative method at a fraction of its sampling cost.
Diffusion models generate data by reversing a forward corruption process that typically approaches a simple Gaussian prior. Recent work has extended this framework to signals supported on fixed graphs, e.g., road-network traffic and sensor-network measurements. Many graph signals have nonuniform spectral energy, whereas isotropic corruption adds the same conditional noise variance to every graph-frequency mode. Driving all modes to near-zero terminal signal-to-noise ratio (SNR) requires strong corruption, which increases the noise range that must be covered under a fixed sampling budget. We introduce Graph Residual Conjugate Diffusion (GRCD), which replaces the shared clock of graph heat diffusion with a mode-dependent clock that gives every graph-Fourier mode the same conditional SNR. GRCD fits a zero-mean graph-spectral Gaussian reference on the training split and stops at a finite terminal SNR at which the propagated reference still carries the fitted spectral variances. The Gaussian component has an exact modewise propagator in the probability-flow ODE, so sampling advances it analytically and integrates only the learned residual score numerically. We evaluate GRCD on five settings (METR-LA traffic, Molene weather, and three stochastic block models) against seven comparators under a matched protocol: Graph-Aware Diffusion (GAD), EDM (graph backbone), two adaptations of Whitened Score Diffusion (WSD), and three preconditioning controls. At four function evaluations (NFEs), GRCD lowers averaged maximum mean discrepancy (aMMD) by 22 to 36 times over the best comparator on all five settings, reaching 0.054 on METR-LA, where it clears an aMMD 0.1 target with 87% less sampling wall-clock time than the cheapest comparator that reaches it. Fitting the terminal reference reduces aMMD by 2.7 to 7.3 times at finite terminal SNR, while the factors shrink to 1.00 to 1.01 near zero.
展开完整摘要收起摘要↓
Diffusion models generate data by reversing a forward corruption process that typically approaches a simple Gaussian prior. Recent work has extended this framework to signals supported on fixed graphs, e.g., road-network traffic and sensor-network measurements. Many graph signals have nonuniform spectral energy, whereas isotropic corruption adds the same conditional noise variance to every graph-frequency mode. Driving all modes to near-zero terminal signal-to-noise ratio (SNR) requires strong corruption, which increases the noise range that must be covered under a fixed sampling budget. We introduce Graph Residual Conjugate Diffusion (GRCD), which replaces the shared clock of graph heat diffusion with a mode-dependent clock that gives every graph-Fourier mode the same conditional SNR. GRCD fits a zero-mean graph-spectral Gaussian reference on the training split and stops at a finite terminal SNR at which the propagated reference still carries the fitted spectral variances. The Gaussian component has an exact modewise propagator in the probability-flow ODE, so sampling advances it analytically and integrates only the learned residual score numerically. We evaluate GRCD on five settings (METR-LA traffic, Molene weather, and three stochastic block models) against seven comparators under a matched protocol: Graph-Aware Diffusion (GAD), EDM (graph backbone), two adaptations of Whitened Score Diffusion (WSD), and three preconditioning controls. At four function evaluations (NFEs), GRCD lowers averaged maximum mean discrepancy (aMMD) by 22 to 36 times over the best comparator on all five settings, reaching 0.054 on METR-LA, where it clears an aMMD 0.1 target with 87% less sampling wall-clock time than the cheapest comparator that reaches it. Fitting the terminal reference reduces aMMD by 2.7 to 7.3 times at finite terminal SNR, while the factors shrink to 1.00 to 1.01 near zero.
Flow Matching (FM) learns a velocity field whose ODE transports a simple source distribution to a target law. Existing finite-sample theory largely treats ambient-space regularity or data supported exactly on low-dimensional sets. We study linear FM under a noisy latent-generator model, where a low-dimensional Hölder map is perturbed by nondegenerate ambient Gaussian noise, so the target law is full-dimensional despite its latent structure. We construct a spatially regular ReLU velocity class and establish non-asymptotic high-probability approximation and estimation bounds whose leading sample-size exponent is governed by the latent dimension rather than the ambient dimension, with ambient and noise dependence kept explicit. Fixed positive target noise keeps the interpolation nondegenerate over the full time interval. The same spatial regularity propagates the learned velocity error through the transport ODE, yielding a corresponding Wasserstein convergence guarantee. These results show that exact low-dimensional support is not necessary for Flow Matching to retain latent-dimensional statistical behavior.
展开完整摘要收起摘要↓
Flow Matching (FM) learns a velocity field whose ODE transports a simple source distribution to a target law. Existing finite-sample theory largely treats ambient-space regularity or data supported exactly on low-dimensional sets. We study linear FM under a noisy latent-generator model, where a low-dimensional Hölder map is perturbed by nondegenerate ambient Gaussian noise, so the target law is full-dimensional despite its latent structure. We construct a spatially regular ReLU velocity class and establish non-asymptotic high-probability approximation and estimation bounds whose leading sample-size exponent is governed by the latent dimension rather than the ambient dimension, with ambient and noise dependence kept explicit. Fixed positive target noise keeps the interpolation nondegenerate over the full time interval. The same spatial regularity propagates the learned velocity error through the transport ODE, yielding a corresponding Wasserstein convergence guarantee. These results show that exact low-dimensional support is not necessary for Flow Matching to retain latent-dimensional statistical behavior.
Flow matching generates samples by gradually transforming noise into data. In practice, using a finite number of sampling steps introduces a numerical error that depends on the chosen schedule. We study this dependence for Gaussian targets and the explicit midpoint sampling method, using the exact flow field. We measure sampling error by the squared Wasserstein distance between the target distribution and the final distribution produced by the midpoint sampler. We show that the standard conditional optimal transport (CondOT) schedule cancels the leading midpoint error and improves the general convergence bound, even when the sampling steps are unequally spaced. On a uniform grid of $S$ sampling steps, we fix the signal schedule at $α_t=t$ and prove the existence of scalar noise schedules $β_t$ that approach the CondOT noise schedule $1-t$ at rate $1/S$ and yield exact Gaussian sampling for every sufficiently large $S$. Controlled Gaussian experiments illustrate the convergence rates and exact calibration.
展开完整摘要收起摘要↓
Flow matching generates samples by gradually transforming noise into data. In practice, using a finite number of sampling steps introduces a numerical error that depends on the chosen schedule. We study this dependence for Gaussian targets and the explicit midpoint sampling method, using the exact flow field. We measure sampling error by the squared Wasserstein distance between the target distribution and the final distribution produced by the midpoint sampler. We show that the standard conditional optimal transport (CondOT) schedule cancels the leading midpoint error and improves the general convergence bound, even when the sampling steps are unequally spaced. On a uniform grid of $S$ sampling steps, we fix the signal schedule at $α_t=t$ and prove the existence of scalar noise schedules $β_t$ that approach the CondOT noise schedule $1-t$ at rate $1/S$ and yield exact Gaussian sampling for every sufficiently large $S$. Controlled Gaussian experiments illustrate the convergence rates and exact calibration.
Multi-image understanding requires MLLMs not only to recognise the content of individual images, but also to organise visual evidence distributed across them. We study this problem through multi-image re-representation, viewing prompted Chain-of-Thought reasoning and agentic visual tool use as different ways of re-organising visual evidence during reasoning. We introduce Mosaic, a general-purpose multi-image visual harness that enables an MLLM to actively construct visual intermediates with ten composable image operations. We compare five re-representation settings on existing multi-image benchmarks and on MosaicBench, a new grounding-focused benchmark for fine-grained multi-image understanding. Our experiments show that the relative benefits of textual and visual re-representation are strongly task-dependent. Visual re-representation is particularly effective for tasks requiring precise visual evidence, including hypothesis testing, precision comparison, and orientation-sensitive reasoning, while tasks dominated by higher-level semantic content show smaller or less consistent gains. Building on this finding, we train MosaicAgent-8B to use Mosaic with reinforcement learning using only accuracy and format rewards. Without demonstration trajectories or rewards for specific tool-use, the agent learns to compose visual operations over multiple steps and exhibits diverse problem-solving patterns unpromptedly. Code and data will be released at https://github.com/gengyuanmax/Mosaic.
展开完整摘要收起摘要↓
Multi-image understanding requires MLLMs not only to recognise the content of individual images, but also to organise visual evidence distributed across them. We study this problem through multi-image re-representation, viewing prompted Chain-of-Thought reasoning and agentic visual tool use as different ways of re-organising visual evidence during reasoning. We introduce Mosaic, a general-purpose multi-image visual harness that enables an MLLM to actively construct visual intermediates with ten composable image operations. We compare five re-representation settings on existing multi-image benchmarks and on MosaicBench, a new grounding-focused benchmark for fine-grained multi-image understanding. Our experiments show that the relative benefits of textual and visual re-representation are strongly task-dependent. Visual re-representation is particularly effective for tasks requiring precise visual evidence, including hypothesis testing, precision comparison, and orientation-sensitive reasoning, while tasks dominated by higher-level semantic content show smaller or less consistent gains. Building on this finding, we train MosaicAgent-8B to use Mosaic with reinforcement learning using only accuracy and format rewards. Without demonstration trajectories or rewards for specific tool-use, the agent learns to compose visual operations over multiple steps and exhibits diverse problem-solving patterns unpromptedly. Code and data will be released at https://github.com/gengyuanmax/Mosaic.
In hospital workflows, electronic health records (EHRs) are often noisy, and may not contain the evidence needed to confirm events or measurements referenced in a clinical query. Even when database retrieval succeeds, clinical agents can overlook such discrepancies and return plausible but unsupported answers. We introduce EHR-RobustGym, a scalable and interactive environment for evaluating and training robust clinical agents grounded in noisy EHRs. Built on MIMIC-IV hospital records (365K patients, 31 tables, and over 500M records), EHR-RobustGym comprises 5,486 Clean-Noise pairs spanning six clinical intents and both patient-level and population-level queries. The pairs test robustness to Record-level, Value-level, and Query-level noise, while interactive SQL/Python execution and outcome verification support trajectory collection and training. Evaluating multiple LLMs reveals substantial robustness gaps: average task success across proprietary and large-scale open-weight models drops from 62.2% on Clean questions to 37.9% on Noise questions. At k=4, pass^k consistency falls below 50% for most evaluated models, exposing instability in clinical task completion. Supervised fine-tuning and reinforcement learning in EHR-RobustGym improve performance, with gains generalizing to five external EHR benchmarks. Together, these results position EHR-RobustGym as a testbed for evaluating and improving the evidence-grounded robustness of clinical agents.
展开完整摘要收起摘要↓
In hospital workflows, electronic health records (EHRs) are often noisy, and may not contain the evidence needed to confirm events or measurements referenced in a clinical query. Even when database retrieval succeeds, clinical agents can overlook such discrepancies and return plausible but unsupported answers. We introduce EHR-RobustGym, a scalable and interactive environment for evaluating and training robust clinical agents grounded in noisy EHRs. Built on MIMIC-IV hospital records (365K patients, 31 tables, and over 500M records), EHR-RobustGym comprises 5,486 Clean-Noise pairs spanning six clinical intents and both patient-level and population-level queries. The pairs test robustness to Record-level, Value-level, and Query-level noise, while interactive SQL/Python execution and outcome verification support trajectory collection and training. Evaluating multiple LLMs reveals substantial robustness gaps: average task success across proprietary and large-scale open-weight models drops from 62.2% on Clean questions to 37.9% on Noise questions. At k=4, pass^k consistency falls below 50% for most evaluated models, exposing instability in clinical task completion. Supervised fine-tuning and reinforcement learning in EHR-RobustGym improve performance, with gains generalizing to five external EHR benchmarks. Together, these results position EHR-RobustGym as a testbed for evaluating and improving the evidence-grounded robustness of clinical agents.
Large language models (LLMs) encode substantial latent geographic knowledge, yet they reason poorly over space and are unreliable when queried from coordinates alone. Useful behaviour emerges only when structured spatial context is supplied in the prompt. This raises a question geographic evaluation has left unexamined: once the right context is supplied, does the model reason from it, or override it with its own parametric recall? We take up this question with an open pipeline for network-based spatial context retriev-al. In it, the surroundings of a selected point are defined by the pedestrian street network, the area actually reachable on foot. Using only open data and open-weight models, the pipeline retrieves features from OpenStreetMap and the GHS-POP population grid, computes indicators over the network catchment in code, and injects them as a compact spatial brief. On this basis we build a faithfulness benchmark. It labels every claim a model makes by its source (grounded in the brief, or drawn from training knowledge) and its correctness, and it probes each case with a planted false premise that the brief refutes. We evaluate sixteen open-weight model configurations across three families (Qwen, Gemma and Llama, with Gemma in two generations), four size classes and, where available, both thinking and non-thinking modes, on three con-trasting cities, resampling every case over ten seeds. The results show that resistance to the planted premise varies more strongly by model family and generation than by scale, while brief-reading competence forms a partly separate dimension. These behaviours are not captured by conventional world-correctness scores or single-shot evaluation. We release the implementation, spatial briefs, model outputs, and claim-level labels as a reproducible workflow at github.com/perezjoan/NSCR-LLM.
展开完整摘要收起摘要↓
Large language models (LLMs) encode substantial latent geographic knowledge, yet they reason poorly over space and are unreliable when queried from coordinates alone. Useful behaviour emerges only when structured spatial context is supplied in the prompt. This raises a question geographic evaluation has left unexamined: once the right context is supplied, does the model reason from it, or override it with its own parametric recall? We take up this question with an open pipeline for network-based spatial context retriev-al. In it, the surroundings of a selected point are defined by the pedestrian street network, the area actually reachable on foot. Using only open data and open-weight models, the pipeline retrieves features from OpenStreetMap and the GHS-POP population grid, computes indicators over the network catchment in code, and injects them as a compact spatial brief. On this basis we build a faithfulness benchmark. It labels every claim a model makes by its source (grounded in the brief, or drawn from training knowledge) and its correctness, and it probes each case with a planted false premise that the brief refutes. We evaluate sixteen open-weight model configurations across three families (Qwen, Gemma and Llama, with Gemma in two generations), four size classes and, where available, both thinking and non-thinking modes, on three con-trasting cities, resampling every case over ten seeds. The results show that resistance to the planted premise varies more strongly by model family and generation than by scale, while brief-reading competence forms a partly separate dimension. These behaviours are not captured by conventional world-correctness scores or single-shot evaluation. We release the implementation, spatial briefs, model outputs, and claim-level labels as a reproducible workflow at github.com/perezjoan/NSCR-LLM.
Multi-agent collaboration lets large language models (LLMs) improve question answering through deliberation and feedback. Yet shared discussion couples correction with exposure to the same mistakes, which can erode the diversity needed for voting. Self-consistency offers sampling diversity without feedback, while single-pair Actor-Critic collaboration refines only one candidate. We introduce SEPAL, which assigns three private Actor-Critic teams to direct reasoning, evidence grounding, and verification. Role-specific training gives the teams different reasoning objectives beyond sampling variation. Each Critic guides revisions within its own team, preventing feedback from carrying errors across candidates. Once revision ends, majority voting combines only the final answers, keeping the reasoning histories separate until the decision. Across five open-weight backbones and five question-answering benchmarks, SEPAL improves mean accuracy by 1.81 percentage points over a matched single Actor-Critic pair, with improvements across all five backbones. Code is available at https://github.com/zhansan114514/SEPAL.
展开完整摘要收起摘要↓
Multi-agent collaboration lets large language models (LLMs) improve question answering through deliberation and feedback. Yet shared discussion couples correction with exposure to the same mistakes, which can erode the diversity needed for voting. Self-consistency offers sampling diversity without feedback, while single-pair Actor-Critic collaboration refines only one candidate. We introduce SEPAL, which assigns three private Actor-Critic teams to direct reasoning, evidence grounding, and verification. Role-specific training gives the teams different reasoning objectives beyond sampling variation. Each Critic guides revisions within its own team, preventing feedback from carrying errors across candidates. Once revision ends, majority voting combines only the final answers, keeping the reasoning histories separate until the decision. Across five open-weight backbones and five question-answering benchmarks, SEPAL improves mean accuracy by 1.81 percentage points over a matched single Actor-Critic pair, with improvements across all five backbones. Code is available at https://github.com/zhansan114514/SEPAL.
Cooperative language-model agents must coordinate over long horizons and adapt to changing environments and to partners with unfamiliar conventions, yet existing agents map observations to actions without separating persistent coordination strategies from their tactical execution. We introduce OverForge, a training-free hierarchical architecture that separates strategic reasoning over roles and divisions of labour from tactical reasoning over actions within each agent's private, partner-conditioned world model. A metacognitive Prefrontal Cortex Module couples the two levels by forming strategy-action branches, imagining their consequences with a forward model, and committing when confident. In OvercookedV2, OverForge delivers 7 soups in a connected kitchen versus 3 for each flat LLM baseline, retains agreed roles, and adopts roles proposed by unfamiliar partners. Ablations and a fixed-strategy probe show that persistent strategies guide tactical adaptation while each reasoning level contributes to coordination. Memory restarts show that cross-episode partner knowledge supports task performance and partner prediction, linking the hierarchy to continual adaptation.
展开完整摘要收起摘要↓
Cooperative language-model agents must coordinate over long horizons and adapt to changing environments and to partners with unfamiliar conventions, yet existing agents map observations to actions without separating persistent coordination strategies from their tactical execution. We introduce OverForge, a training-free hierarchical architecture that separates strategic reasoning over roles and divisions of labour from tactical reasoning over actions within each agent's private, partner-conditioned world model. A metacognitive Prefrontal Cortex Module couples the two levels by forming strategy-action branches, imagining their consequences with a forward model, and committing when confident. In OvercookedV2, OverForge delivers 7 soups in a connected kitchen versus 3 for each flat LLM baseline, retains agreed roles, and adopts roles proposed by unfamiliar partners. Ablations and a fixed-strategy probe show that persistent strategies guide tactical adaptation while each reasoning level contributes to coordination. Memory restarts show that cross-episode partner knowledge supports task performance and partner prediction, linking the hierarchy to continual adaptation.
Agent memory faces heterogeneous access needs: a single-hop question may require one piece of evidence, whereas a multi-hop question must combine evidence from multiple sources. Predefined memory workflows cannot adapt to these varying needs. Recent adaptive methods search or learn over memory components and their compositions, but the design space itself remains predefined. We introduce MemCodex, a self-evolving hierarchical memory system that organizes experience into executable memory programs for summaries, relational knowledge, reusable skills, and latent memory. Open-ended program evolution searches the open design space of layer programs by rewriting how each layer is constructed, indexed, retrieved, and routed, thereby adapting both within-layer implementations and cross-layer composition. At query time, reads traverse the hierarchy from coarse to fine and stop once sufficient evidence is found, descending to the original history when needed. We further develop MemArena, a unified runtime that places heterogeneous data and memory systems behind a common interface. MemCodex improves average task success by 10.1% relative to the strongest adaptive-memory baseline, while using 3.4x fewer context tokens and achieving 2.1x faster inference.
展开完整摘要收起摘要↓
Agent memory faces heterogeneous access needs: a single-hop question may require one piece of evidence, whereas a multi-hop question must combine evidence from multiple sources. Predefined memory workflows cannot adapt to these varying needs. Recent adaptive methods search or learn over memory components and their compositions, but the design space itself remains predefined. We introduce MemCodex, a self-evolving hierarchical memory system that organizes experience into executable memory programs for summaries, relational knowledge, reusable skills, and latent memory. Open-ended program evolution searches the open design space of layer programs by rewriting how each layer is constructed, indexed, retrieved, and routed, thereby adapting both within-layer implementations and cross-layer composition. At query time, reads traverse the hierarchy from coarse to fine and stop once sufficient evidence is found, descending to the original history when needed. We further develop MemArena, a unified runtime that places heterogeneous data and memory systems behind a common interface. MemCodex improves average task success by 10.1% relative to the strongest adaptive-memory baseline, while using 3.4x fewer context tokens and achieving 2.1x faster inference.
Large language models (LLMs) have shown promise in automating interactive theorem proving, yet verification of real-world C codebases requires more than discharging individual proof goals. The task involves jointly constructing expressive function specifications and their proofs, and ensuring that library interfaces compose along intended call sequences even without a designated client. This paper presents CCV, an LLM-assisted framework for building machine-checked assurance cases: structured, auditable artifacts supporting the claim that a C codebase meets its intended requirements. To model intended cross-interface use in open libraries, CCV constructs an interface protocol that exposes permitted call sequences and resource assumptions for review, with a conditional safety guarantee under verified contracts and caller obligations. CCV coordinates two complementary phases: (i) requirement-guided analysis and bottom-up construction of candidate specifications and protocols; and (ii) modular proof construction with feedback that revises the specifications and proofs. Implemented using VST in Rocq, CCV verifies memory safety and leak freedom for all 299 function definitions across six C benchmarks, including industrial cryptographic components, with less than one person-day of reported human effort per benchmark. The guarantees depend on disclosed contracts and assumptions; human review supplies the conformance judgments connecting the formal artifacts to the intended requirements.
展开完整摘要收起摘要↓
Large language models (LLMs) have shown promise in automating interactive theorem proving, yet verification of real-world C codebases requires more than discharging individual proof goals. The task involves jointly constructing expressive function specifications and their proofs, and ensuring that library interfaces compose along intended call sequences even without a designated client. This paper presents CCV, an LLM-assisted framework for building machine-checked assurance cases: structured, auditable artifacts supporting the claim that a C codebase meets its intended requirements. To model intended cross-interface use in open libraries, CCV constructs an interface protocol that exposes permitted call sequences and resource assumptions for review, with a conditional safety guarantee under verified contracts and caller obligations. CCV coordinates two complementary phases: (i) requirement-guided analysis and bottom-up construction of candidate specifications and protocols; and (ii) modular proof construction with feedback that revises the specifications and proofs. Implemented using VST in Rocq, CCV verifies memory safety and leak freedom for all 299 function definitions across six C benchmarks, including industrial cryptographic components, with less than one person-day of reported human effort per benchmark. The guarantees depend on disclosed contracts and assumptions; human review supplies the conformance judgments connecting the formal artifacts to the intended requirements.
作者Guanqun Yang, Wenlong Zhang, Tian Shi, Ping Wang
Anthropic's Agent Skills package reusable procedural know-how for an LLM agent into SKILL.md directories, and open-source aggregations have grown past 230,000 skills, making selection rather than authoring the bottleneck. The standing answer in the literature outsources selection to the agent itself: an LLM-mediated retrieval loop that rewrites queries and refines candidates inside the agent's decision loop, paying LLM tokens on every task. We present SkillSeek, an open-source two-stage skill retriever built from the standard IR recipe (a BGE-base bi-encoder feeding a small cross-encoder, exposed over MCP). Across a $4 \times 11$ grid of pool, backbone, and method on the 89-task SkillsBench benchmark, SkillSeek reaches observed parity with the LLM-mediated loop of Liu et al. at essentially no extra cost: plain bm25 alone records a pass rate at or above their refined loop on three of four settings, and a small cross-encoder covers the remaining difference on the fourth. A first-stage recall ceiling explains the pattern, and total per-trial spend drops from USD 51.30 to USD 27.54 (within fifty cents of the no-skill baseline). Under the SkillsBench tasks and OpenHands harness we tested, this positions the standard IR recipe as a strong default for agent-skill retrieval, with LLM-mediated alternatives a natural fit for cases where deterministic methods fall short.
展开完整摘要收起摘要↓
Anthropic's Agent Skills package reusable procedural know-how for an LLM agent into SKILL.md directories, and open-source aggregations have grown past 230,000 skills, making selection rather than authoring the bottleneck. The standing answer in the literature outsources selection to the agent itself: an LLM-mediated retrieval loop that rewrites queries and refines candidates inside the agent's decision loop, paying LLM tokens on every task. We present SkillSeek, an open-source two-stage skill retriever built from the standard IR recipe (a BGE-base bi-encoder feeding a small cross-encoder, exposed over MCP). Across a $4 \times 11$ grid of pool, backbone, and method on the 89-task SkillsBench benchmark, SkillSeek reaches observed parity with the LLM-mediated loop of Liu et al. at essentially no extra cost: plain bm25 alone records a pass rate at or above their refined loop on three of four settings, and a small cross-encoder covers the remaining difference on the fourth. A first-stage recall ceiling explains the pattern, and total per-trial spend drops from USD 51.30 to USD 27.54 (within fifty cents of the no-skill baseline). Under the SkillsBench tasks and OpenHands harness we tested, this positions the standard IR recipe as a strong default for agent-skill retrieval, with LLM-mediated alternatives a natural fit for cases where deterministic methods fall short.
作者Chenglin Chen, Lujia Wang, Xinhu Zheng, Jun Ma, Haoang Li
Safe and efficient trajectory planning is essential in autonomous driving. However, existing end-to-end approaches often fall short in both computational efficiency and safety guarantees. Methods based on imitation learning suffer from causal confusion, while rule-based scoring approaches often incur heavy computational overhead and suffer from objective misalignment. Additionally, preference-based methods rely on strict pairwise annotations, limiting data utilization. To overcome these limitations, we propose EMPlan, an efficient multi-modal trajectory planning method powered by reward-guided fine-tuning. We design a hybrid architecture that combines sparse anchors with an offset refinement module for efficient multi-modal trajectory prediction. Sparse anchors provide coarse trajectory candidates with low latency, which are subsequently refined by the offset module for higher prediction accuracy. To enhance safety without incurring additional inference costs, we adopt a two-stage training paradigm consisting of pretraining and reward-guided fine-tuning. During fine-tuning, we leverage rule-based reward signals and unpaired preference supervision to refine the pretrained policy toward safer trajectory selection. We evaluate EMPlan on the non-reactive NAVSIM benchmark, where it strikes a favorable balance between planning accuracy and efficiency, demonstrating superior performance under real-time constraints.
展开完整摘要收起摘要↓
Safe and efficient trajectory planning is essential in autonomous driving. However, existing end-to-end approaches often fall short in both computational efficiency and safety guarantees. Methods based on imitation learning suffer from causal confusion, while rule-based scoring approaches often incur heavy computational overhead and suffer from objective misalignment. Additionally, preference-based methods rely on strict pairwise annotations, limiting data utilization. To overcome these limitations, we propose EMPlan, an efficient multi-modal trajectory planning method powered by reward-guided fine-tuning. We design a hybrid architecture that combines sparse anchors with an offset refinement module for efficient multi-modal trajectory prediction. Sparse anchors provide coarse trajectory candidates with low latency, which are subsequently refined by the offset module for higher prediction accuracy. To enhance safety without incurring additional inference costs, we adopt a two-stage training paradigm consisting of pretraining and reward-guided fine-tuning. During fine-tuning, we leverage rule-based reward signals and unpaired preference supervision to refine the pretrained policy toward safer trajectory selection. We evaluate EMPlan on the non-reactive NAVSIM benchmark, where it strikes a favorable balance between planning accuracy and efficiency, demonstrating superior performance under real-time constraints.
作者Wei Zhao, Yangshuo Zou, Chengxiang Ding, Yifan Wu, Xuchuan Wang, Zimu Mao, Lei Zhang, Tao Luo
We present FYAN, a human--AI harness for document-level mathematical formalization. Rather than treating theorems in isolation, FYAN coordinates an end-to-end workflow spanning specification, proof planning, logical review, Lean proof construction, knowledge curation, and validation, with support for independent supervision and human guidance. A central component is evidence-grounded semantic auditing, which assesses whether formal statements faithfully preserve their informal specifications. A language model constructs structured evidence over local correspondences, omissions, scope, and logical relations, while a deterministic validator checks this evidence and produces reproducible judgments. When a substantive but admissible deviation is accepted, FYAN requires an explicit proof-transfer obligation connecting the formal statement back to a source-facing interpretation. With the same model (DeepSeek-V4.1-Flash) in every stage, FYAN proves 86 of 143 FormalTCS theorems under a strict Lean check, against 69 for a general agent harness, and raises the natural-language proof score from 0.501 to 0.851. On ConsistencyCheck, its semantic audit catches more inconsistent statements than a direct LLM judge, both on labels verified against the source (recall 0.777 vs. 0.636) and on the original labels (0.873 vs. 0.820), and localizes each mismatch it reports to a specific hypothesis, conclusion, or scope. FYAN also built ODENumLib, a 9,355-line Lean library for the numerical analysis of ordinary differential equation.
展开完整摘要收起摘要↓
We present FYAN, a human--AI harness for document-level mathematical formalization. Rather than treating theorems in isolation, FYAN coordinates an end-to-end workflow spanning specification, proof planning, logical review, Lean proof construction, knowledge curation, and validation, with support for independent supervision and human guidance. A central component is evidence-grounded semantic auditing, which assesses whether formal statements faithfully preserve their informal specifications. A language model constructs structured evidence over local correspondences, omissions, scope, and logical relations, while a deterministic validator checks this evidence and produces reproducible judgments. When a substantive but admissible deviation is accepted, FYAN requires an explicit proof-transfer obligation connecting the formal statement back to a source-facing interpretation. With the same model (DeepSeek-V4.1-Flash) in every stage, FYAN proves 86 of 143 FormalTCS theorems under a strict Lean check, against 69 for a general agent harness, and raises the natural-language proof score from 0.501 to 0.851. On ConsistencyCheck, its semantic audit catches more inconsistent statements than a direct LLM judge, both on labels verified against the source (recall 0.777 vs. 0.636) and on the original labels (0.873 vs. 0.820), and localizes each mismatch it reports to a specific hypothesis, conclusion, or scope. FYAN also built ODENumLib, a 9,355-line Lean library for the numerical analysis of ordinary differential equation.
Real-world embodied tasks, from everyday activities to professional procedures, require agents to act under physical constraints while tracking evolving object and task states. Tool use sits at the heart of such tasks, as many everyday and professional activities are tool-mediated. Understanding them requires reasoning about affordances, hand-tool-object geometry, procedural progress, and causal effects on target objects. Yet despite strong performance on perception-oriented video tasks such as captioning and general video QA, current multimodal video models remain limited in this form of tool-centric embodied reasoning. Progress in this direction has been limited by the lack of real-world egocentric data and diagnostic benchmarks. To address this gap, we introduce EgoTools, the first comprehensive suite for egocentric tool-use understanding. It consists of two complementary components: EgoTools-Data, a large-scale corpus of 100 hours of tool-centric egocentric recordings with synchronized audio, dense captions, reasoning-heavy narrations, and supplementary 3D information; and EgoTools-Bench, a diagnostic benchmark of 1,000 QA pairs across four tracks that cover tool-use understanding from perception and geometry to procedure and causal reasoning. Experimental results show that current models still struggle to ground tool use in visual evidence: Gemini-3.1-Pro achieves 66.9% overall accuracy but only 51.7% on Perception & Grounding. Beyond evaluation, we validate EgoTools-Data as a training resource. On the full 1,000-question benchmark, full supervised fine-tuning improves Qwen3-VL-8B-Instruct from 50.0% to 60.9%, under strict source-video separation. Together, these results establish EgoTools as a unified resource for both training and diagnostic evaluation of real-world egocentric tool-use understanding.
展开完整摘要收起摘要↓
Real-world embodied tasks, from everyday activities to professional procedures, require agents to act under physical constraints while tracking evolving object and task states. Tool use sits at the heart of such tasks, as many everyday and professional activities are tool-mediated. Understanding them requires reasoning about affordances, hand-tool-object geometry, procedural progress, and causal effects on target objects. Yet despite strong performance on perception-oriented video tasks such as captioning and general video QA, current multimodal video models remain limited in this form of tool-centric embodied reasoning. Progress in this direction has been limited by the lack of real-world egocentric data and diagnostic benchmarks. To address this gap, we introduce EgoTools, the first comprehensive suite for egocentric tool-use understanding. It consists of two complementary components: EgoTools-Data, a large-scale corpus of 100 hours of tool-centric egocentric recordings with synchronized audio, dense captions, reasoning-heavy narrations, and supplementary 3D information; and EgoTools-Bench, a diagnostic benchmark of 1,000 QA pairs across four tracks that cover tool-use understanding from perception and geometry to procedure and causal reasoning. Experimental results show that current models still struggle to ground tool use in visual evidence: Gemini-3.1-Pro achieves 66.9% overall accuracy but only 51.7% on Perception & Grounding. Beyond evaluation, we validate EgoTools-Data as a training resource. On the full 1,000-question benchmark, full supervised fine-tuning improves Qwen3-VL-8B-Instruct from 50.0% to 60.9%, under strict source-video separation. Together, these results establish EgoTools as a unified resource for both training and diagnostic evaluation of real-world egocentric tool-use understanding.
Large language models are strong general-purpose code generators, but executable algorithmic trading remains a demanding specialization target: a model must translate a natural-language strategy specification into correct program logic for a specialized trading framework, execute on historical data, produce trades, and remain semantically faithful to the request. We study two complementary mechanisms for specializing language models for this setting: continued pretraining on algorithmic-trading framework code and supervised fine-tuning (SFT) on agent-validated request-to-code pairs. Evaluation is centered on QuantCode-Bench, our 400-task benchmark for Backtrader strategy generation, together with a repository-level SWE-bench-like track. Continued pretraining improves single-turn Judge Pass from 41.5% to 47.5% for Qwen3.5-397B-A17B and from 27.8% to 33.0% for Qwen3.6-35B-A3B. SFT applied after continued pretraining yields a larger gain for Qwen3.6-35B-A3B, reaching 58.2% Judge Pass and 83.5% successful backtests; in agentic evaluation it raises first-turn success from 22.3% to 58.3% and final success after up to 10 turns from 47.5% to 79.5%. Continued pretraining alone improves first-turn agentic success but lowers final success after repair from 47.5% to 32.5%, consistent with degraded instruction following, whereas SFT improves both. We also identify a capability-retention failure: domain specialization degrades parser-conformant structured tool calling, and targeted recovery SFT restores tool-call formatting but not the base checkpoint's repository-level agent performance. The results show that framework-oriented pretraining, validated SFT, and explicit capability-retention evaluation address distinct failure modes in domain-specific executable code generation.
展开完整摘要收起摘要↓
Large language models are strong general-purpose code generators, but executable algorithmic trading remains a demanding specialization target: a model must translate a natural-language strategy specification into correct program logic for a specialized trading framework, execute on historical data, produce trades, and remain semantically faithful to the request. We study two complementary mechanisms for specializing language models for this setting: continued pretraining on algorithmic-trading framework code and supervised fine-tuning (SFT) on agent-validated request-to-code pairs. Evaluation is centered on QuantCode-Bench, our 400-task benchmark for Backtrader strategy generation, together with a repository-level SWE-bench-like track. Continued pretraining improves single-turn Judge Pass from 41.5% to 47.5% for Qwen3.5-397B-A17B and from 27.8% to 33.0% for Qwen3.6-35B-A3B. SFT applied after continued pretraining yields a larger gain for Qwen3.6-35B-A3B, reaching 58.2% Judge Pass and 83.5% successful backtests; in agentic evaluation it raises first-turn success from 22.3% to 58.3% and final success after up to 10 turns from 47.5% to 79.5%. Continued pretraining alone improves first-turn agentic success but lowers final success after repair from 47.5% to 32.5%, consistent with degraded instruction following, whereas SFT improves both. We also identify a capability-retention failure: domain specialization degrades parser-conformant structured tool calling, and targeted recovery SFT restores tool-call formatting but not the base checkpoint's repository-level agent performance. The results show that framework-oriented pretraining, validated SFT, and explicit capability-retention evaluation address distinct failure modes in domain-specific executable code generation.
作者Yong Du, Tongbo Chen, Zhengxi Lu, Yizhou Liu, Bofan Chen, Tao Jiang, Wenhao Xu, Yongliang Shen
Online training enables computer-use agents (CUAs) to improve through interaction with executable environments. However, existing methods primarily rely on sparse outcome rewards, which provide no supervision for intermediate actions. On-policy self-distillation (OPSD) offers token-level learning signals through privileged rescoring, but directly applying it to CUA online training presents two challenges: fixed guidance may become misaligned with the student's current state, and guidance-induced probability shifts may conflict with step-level correctness. We introduce ComputerSD, an online self-distillation method for CUAs that converts real-time feedback from executed GUI transitions into guidance for policy learning. A fine-tuned GUI analyzer produces guidance and a step-level value score after each action; the guidance provides privileged context, while the score regulates the resulting OPSD signals. ComputerSD jointly optimizes token-level OPSD and trajectory-level GRPO in a fully asynchronous training framework. On OSWorld-Verified, ComputerSD outperforms outcome-only GRPO by 1.9 and 4.1 percentage points on the general-purpose Qwen3-VL-8B-Thinking and specialized EvoCUA-8B backbones, respectively. Evaluation in out-of-distribution settings further supports the generalizability of ComputerSD. These results demonstrate the effectiveness of learning from real-time feedback through online self-distillation for CUAs.
展开完整摘要收起摘要↓
Online training enables computer-use agents (CUAs) to improve through interaction with executable environments. However, existing methods primarily rely on sparse outcome rewards, which provide no supervision for intermediate actions. On-policy self-distillation (OPSD) offers token-level learning signals through privileged rescoring, but directly applying it to CUA online training presents two challenges: fixed guidance may become misaligned with the student's current state, and guidance-induced probability shifts may conflict with step-level correctness. We introduce ComputerSD, an online self-distillation method for CUAs that converts real-time feedback from executed GUI transitions into guidance for policy learning. A fine-tuned GUI analyzer produces guidance and a step-level value score after each action; the guidance provides privileged context, while the score regulates the resulting OPSD signals. ComputerSD jointly optimizes token-level OPSD and trajectory-level GRPO in a fully asynchronous training framework. On OSWorld-Verified, ComputerSD outperforms outcome-only GRPO by 1.9 and 4.1 percentage points on the general-purpose Qwen3-VL-8B-Thinking and specialized EvoCUA-8B backbones, respectively. Evaluation in out-of-distribution settings further supports the generalizability of ComputerSD. These results demonstrate the effectiveness of learning from real-time feedback through online self-distillation for CUAs.
Textual skills enable large language model (LLM) based agents to accumulate reusable procedural knowledge without updating model parameters. Yet existing skill evolution remains largely confined to the text space: an optimizer must diagnose success and failure patterns, and revise skills solely from long execution trajectories and sparse task outcomes. This text-only paradigm leaves the agent's internal representations, which contain rich records of its evolving execution state, outside the skill optimization loop. We ask whether an agent can improve its external textual skills by reflecting on its own internal representations. We introduce Rep2Skill, a representation-guided framework for self-evolution on agent skills. Specifically, upon the collected agent rollouts, Rep2Skill models their internal model representation trajectories to localize turns that deviate from successful execution dynamics, and it further interprets these signals alongside the execution contexts as actionable textual feedback for targeted skill revision. Experiments on two agent environments with two open-source LLMs show that Rep2Skill consistently outperforms text-only approaches in the self-evolution setting, where the same LLM serves as both executor and optimizer without a stronger external model. This establishes a promising direction moving agent self-improvement beyond text-only reflection.
展开完整摘要收起摘要↓
Textual skills enable large language model (LLM) based agents to accumulate reusable procedural knowledge without updating model parameters. Yet existing skill evolution remains largely confined to the text space: an optimizer must diagnose success and failure patterns, and revise skills solely from long execution trajectories and sparse task outcomes. This text-only paradigm leaves the agent's internal representations, which contain rich records of its evolving execution state, outside the skill optimization loop. We ask whether an agent can improve its external textual skills by reflecting on its own internal representations. We introduce Rep2Skill, a representation-guided framework for self-evolution on agent skills. Specifically, upon the collected agent rollouts, Rep2Skill models their internal model representation trajectories to localize turns that deviate from successful execution dynamics, and it further interprets these signals alongside the execution contexts as actionable textual feedback for targeted skill revision. Experiments on two agent environments with two open-source LLMs show that Rep2Skill consistently outperforms text-only approaches in the self-evolution setting, where the same LLM serves as both executor and optimizer without a stronger external model. This establishes a promising direction moving agent self-improvement beyond text-only reflection.
Autonomous AI agents can turn authors' goals into interactive narratives by independently organizing and carrying out generation and revision. As agents generate and revise extensive content, authors struggle to grasp its overall structure, local details, and relationships, complicating continued guidance. We present NarrativeSteward, an authoring environment that organizes outlines, worldbuilding, and narrative graphs as linked artifacts for agent implementation and author guidance. Agent dialogue and project-wide structural review help authors understand the evolving work and guide local and cross-layer revisions, while change records and execution verification help authors assess the resulting work. Technical tests validated the system's change records, recovery mechanisms, and execution diagnostics. In a 12-participant within-subject study, NarrativeSteward supported easier formulation of revision requests and inspection of changes, and greater perceived understanding of changes and story structure, than general-purpose agents. Qualitative findings show how reviewing the work and feedback helps authors develop requirements and guide subsequent delegation. We open-source NarrativeSteward at https://github.com/Tencent/NarrativeSteward.
展开完整摘要收起摘要↓
Autonomous AI agents can turn authors' goals into interactive narratives by independently organizing and carrying out generation and revision. As agents generate and revise extensive content, authors struggle to grasp its overall structure, local details, and relationships, complicating continued guidance. We present NarrativeSteward, an authoring environment that organizes outlines, worldbuilding, and narrative graphs as linked artifacts for agent implementation and author guidance. Agent dialogue and project-wide structural review help authors understand the evolving work and guide local and cross-layer revisions, while change records and execution verification help authors assess the resulting work. Technical tests validated the system's change records, recovery mechanisms, and execution diagnostics. In a 12-participant within-subject study, NarrativeSteward supported easier formulation of revision requests and inspection of changes, and greater perceived understanding of changes and story structure, than general-purpose agents. Qualitative findings show how reviewing the work and feedback helps authors develop requirements and guide subsequent delegation. We open-source NarrativeSteward at https://github.com/Tencent/NarrativeSteward.
Textual skills provide reusable guidance for large language model agents, but existing approaches often rely on manually curated skill banks or reinforcement learning with indirect and delayed feedback. We introduce SkillFM (Skill Flow Matching), a generative framework that synthesizes task-conditioned textual skills directly without test-time skill retrieval. Our framework combines a codec for encoding and reconstructing textual skills in a continuous latent space with a conditional flow model trained using improved MeanFlow. At inference time, the learned velocity field enables single-step latent sampling, and an LLM-based decoder converts the sampled representation into textual guidance for a frozen downstream agent. We evaluate the framework on embodied tasks, question answering, and web shopping. On ALFWorld and Search-QA, our method achieves the best overall performance among the compared vector-based skill approaches. Our analyses further demonstrate that latent skill generation is an effective alternative to retrieval-based skill augmentation. Our code and training skill libraries are available at https://github.com/lulushang999/SkillFM.
展开完整摘要收起摘要↓
Textual skills provide reusable guidance for large language model agents, but existing approaches often rely on manually curated skill banks or reinforcement learning with indirect and delayed feedback. We introduce SkillFM (Skill Flow Matching), a generative framework that synthesizes task-conditioned textual skills directly without test-time skill retrieval. Our framework combines a codec for encoding and reconstructing textual skills in a continuous latent space with a conditional flow model trained using improved MeanFlow. At inference time, the learned velocity field enables single-step latent sampling, and an LLM-based decoder converts the sampled representation into textual guidance for a frozen downstream agent. We evaluate the framework on embodied tasks, question answering, and web shopping. On ALFWorld and Search-QA, our method achieves the best overall performance among the compared vector-based skill approaches. Our analyses further demonstrate that latent skill generation is an effective alternative to retrieval-based skill augmentation. Our code and training skill libraries are available at https://github.com/lulushang999/SkillFM.
作者Max Otto, Haci Ismail Aslan, Joel Witzke, Jonathan Bader, Odej Kao
Efficient resource provisioning for large-scale workflows on cloud infrastructures is a critical performance engineering challenge. These workflows are often structured as directed acyclic graphs (DAGs), where under-provisioning can cause critical bottlenecks and over-provisioning leads to unnecessary costs. Accurate, task-level prediction of resource intensity (e.g., CPU load and memory usage) is essential for mitigating these issues. While task-level features are commonly used for prediction, the performance impact of the workflow's overall topological structure is often overlooked or assumed. The central question of our work is: To what extent does what part of the DAG topology influence task-level resource intensity, and what is the most effective way to model this influence? This paper presents a comprehensive benchmark to systematically quantify the impact of graph topology on task intensity prediction. We evaluate and compare a spectrum of modeling approaches. Our findings demonstrate that topology is a critical feature for accurate prediction. Models incorporating important topological information, even through simple handcrafted features, significantly outperform baseline models. We show that graph-native models provide the highest accuracy, achieving low mean absolute errors for both CPU and memory predictions, and can still be combined with simple topological features that they do not learn for better performance.
展开完整摘要收起摘要↓
Efficient resource provisioning for large-scale workflows on cloud infrastructures is a critical performance engineering challenge. These workflows are often structured as directed acyclic graphs (DAGs), where under-provisioning can cause critical bottlenecks and over-provisioning leads to unnecessary costs. Accurate, task-level prediction of resource intensity (e.g., CPU load and memory usage) is essential for mitigating these issues. While task-level features are commonly used for prediction, the performance impact of the workflow's overall topological structure is often overlooked or assumed. The central question of our work is: To what extent does what part of the DAG topology influence task-level resource intensity, and what is the most effective way to model this influence? This paper presents a comprehensive benchmark to systematically quantify the impact of graph topology on task intensity prediction. We evaluate and compare a spectrum of modeling approaches. Our findings demonstrate that topology is a critical feature for accurate prediction. Models incorporating important topological information, even through simple handcrafted features, significantly outperform baseline models. We show that graph-native models provide the highest accuracy, achieving low mean absolute errors for both CPU and memory predictions, and can still be combined with simple topological features that they do not learn for better performance.
作者Kunlun Zhu, Xuyan Ye, Yibo Li, Cheng Qian, Beibin Li, Heng Ji
An unsuccessful LLM agent rollout contains more information than its final reward: the observations available to the agent, the actions it chose, and the environment's responses. Reusing this experience for learning requires identifying a decision to revise and testing a concrete alternative. We introduce the Agent Error Dataset (AED), comprising 50,228 error-diagnosis pairs from 9,961 source tasks across 33 environments, 19 harness families, and 23 policy models in text-based agent systems. We retain source traces and execution metadata to support cross-setting failure analysis and re-diagnosis without repeating the original rollout. Our five-stage Agentic Error-to-Training (AET) pipeline collects natural failures, generates diagnoses and proposed corrections, and checks them against recorded evidence. Where replay is supported, we compare corrections with original-action retries from the same checkpoint under matched execution settings. We then construct separate training views for diagnosis and actor recovery. Across 3,062 matched replay pairs, first-proposal corrections raise verifier pass rates from 18.4% to 51.1%, a gain of 32.7 percentage points. Using a separately frozen diagnosis release, full-diagnosis fine-tuning on 1,656 source tasks raises Qwen3-8B's exact-step agreement with internal teacher labels from 47.2% to 63.6%, averaged over three seeds on a 943-case holdout. The strongest prompted reference in this comparison scores 54.7%, and mean agreement improves at each of four increasing training-set sizes. In a single-seed comparison of actor-training recipes, action-only repair training scores 6.67 percentage points higher on WebShop-lite than success-only training.
展开完整摘要收起摘要↓
An unsuccessful LLM agent rollout contains more information than its final reward: the observations available to the agent, the actions it chose, and the environment's responses. Reusing this experience for learning requires identifying a decision to revise and testing a concrete alternative. We introduce the Agent Error Dataset (AED), comprising 50,228 error-diagnosis pairs from 9,961 source tasks across 33 environments, 19 harness families, and 23 policy models in text-based agent systems. We retain source traces and execution metadata to support cross-setting failure analysis and re-diagnosis without repeating the original rollout. Our five-stage Agentic Error-to-Training (AET) pipeline collects natural failures, generates diagnoses and proposed corrections, and checks them against recorded evidence. Where replay is supported, we compare corrections with original-action retries from the same checkpoint under matched execution settings. We then construct separate training views for diagnosis and actor recovery. Across 3,062 matched replay pairs, first-proposal corrections raise verifier pass rates from 18.4% to 51.1%, a gain of 32.7 percentage points. Using a separately frozen diagnosis release, full-diagnosis fine-tuning on 1,656 source tasks raises Qwen3-8B's exact-step agreement with internal teacher labels from 47.2% to 63.6%, averaged over three seeds on a 943-case holdout. The strongest prompted reference in this comparison scores 54.7%, and mean agreement improves at each of four increasing training-set sizes. In a single-seed comparison of actor-training recipes, action-only repair training scores 6.67 percentage points higher on WebShop-lite than success-only training.
Learning to complete tasks in unfamiliar environments with unknown rules remains a key challenge for LLM agents. Current LLM agents often record their discoveries in prose, which may not provide a compact, explicit account of how the environment works. Inspired by how scientists organize observations into testable, predictive theories, we introduce Schema, an agent harness that organizes learning and action through interactive program induction. The LLM agent decides what to investigate and how to act, expressing its evolving understanding of the environment as executable programs. The harness consists of a persistent program workspace and a small set of interfaces for checking these programs against the interaction history, planning within them, and executing plans under step-by-step verification. Schema raises ARC-AGI-3 RHAE from 58.7% to 99.2% with the same base model, solves 100% of the public DiG-bench games, and reaches the median performance of the top-50 human players on MazeBench. Extensive analysis shows the effectiveness of Schema in unknown mechanism discovery, and ablations confirm the contribution of each component.
展开完整摘要收起摘要↓
Learning to complete tasks in unfamiliar environments with unknown rules remains a key challenge for LLM agents. Current LLM agents often record their discoveries in prose, which may not provide a compact, explicit account of how the environment works. Inspired by how scientists organize observations into testable, predictive theories, we introduce Schema, an agent harness that organizes learning and action through interactive program induction. The LLM agent decides what to investigate and how to act, expressing its evolving understanding of the environment as executable programs. The harness consists of a persistent program workspace and a small set of interfaces for checking these programs against the interaction history, planning within them, and executing plans under step-by-step verification. Schema raises ARC-AGI-3 RHAE from 58.7% to 99.2% with the same base model, solves 100% of the public DiG-bench games, and reaches the median performance of the top-50 human players on MazeBench. Extensive analysis shows the effectiveness of Schema in unknown mechanism discovery, and ablations confirm the contribution of each component.
作者Anmol Kabra, Swathi Saravana Selvam, Albert Gong, Chao Wan, Christian Belardi, Dongyoung Go, Katie Z. Luo, Kilian Q. Weinberger
Training LLM agents with reinforcement learning (RL) is bottlenecked by environments, which must provide verifiable rewards, support long-horizon interaction, and scale cheaply. Existing approaches rely on costly human-curated data or on LLM-generated environments that risk hallucinations and benchmark contamination. We show that LLMs can instead be trained into capable search agents using synthetic environments generated entirely by rules, whose generation requires no LLM and has zero marginal cost. We build PhantomEnvironments, multi-turn RL environments from fictional worlds, where agents must search a corpus of templated articles to answer multi-hop questions. Despite sharing no facts with the real world, these strikingly simple environments yield agents that transfer to real-world multi-hop search benchmarks, often outperforming real-world training data on newer benchmarks. Trained agents generalize to unseen fictional universes, and Qwen models learn to scale their search budget roughly linearly with question difficulty, suggesting emergent search scaling from environment interaction alone. Ablating environment complexity reveals that hop count drives transfer more than constraints or comparisons: even the simplest rule-generated environments are a surprisingly effective, free resource for training generalizable LLM agents.
展开完整摘要收起摘要↓
Training LLM agents with reinforcement learning (RL) is bottlenecked by environments, which must provide verifiable rewards, support long-horizon interaction, and scale cheaply. Existing approaches rely on costly human-curated data or on LLM-generated environments that risk hallucinations and benchmark contamination. We show that LLMs can instead be trained into capable search agents using synthetic environments generated entirely by rules, whose generation requires no LLM and has zero marginal cost. We build PhantomEnvironments, multi-turn RL environments from fictional worlds, where agents must search a corpus of templated articles to answer multi-hop questions. Despite sharing no facts with the real world, these strikingly simple environments yield agents that transfer to real-world multi-hop search benchmarks, often outperforming real-world training data on newer benchmarks. Trained agents generalize to unseen fictional universes, and Qwen models learn to scale their search budget roughly linearly with question difficulty, suggesting emergent search scaling from environment interaction alone. Ablating environment complexity reveals that hop count drives transfer more than constraints or comparisons: even the simplest rule-generated environments are a surprisingly effective, free resource for training generalizable LLM agents.
作者Zongwan Cao, Ziyuan Yang, Shangbin Feng, Michael Duan, Skyler Hallinan, Bingbing Wen, Lucy Lu Wang, Yulia Tsvetkov
While multi-agent and model collaboration algorithms gain traction to combine the strengths of diverse Large Language Models (LLMs), existing systems remain bottlenecked on pre-defined and hand-crafted model pools. In this work, we investigate the problem of model selection in multi-LLM systems. We propose and systematically evaluate a taxonomy of 9 selection algorithms ranging from diversity of model descriptions, capability-aware behavioral diversity, and LLM-based recruiters. We conduct extensive experiments across two candidate pools of 10 and 32 models, deployed in four model collaboration algorithms, and evaluated across tasks spanning math, coding, QA, and reasoning. Results demonstrate that successful selection algorithms greatly outperform random or heuristics-based teams such as merely selecting the models with top individual performance, by up to 36.1% across settings. Specifically, capability- and training-based selection strategies alleviate selection variance and achieve the best performance, which we recommend to employ before deploying real-world multi-LLM systems. Further analysis reveals that larger candidate pools pose greater challenges to shallow selection heuristics, while algorithms grounded in interacting with candidate models and understanding model capability robustly filter out misaligned, unsafe models, as well as generalizing to novel, out-of-distribution tasks. Together, we establish that principled and informed team selection is critical and present strong model selection algorithms for assembling effective multi-LLM systems.
展开完整摘要收起摘要↓
While multi-agent and model collaboration algorithms gain traction to combine the strengths of diverse Large Language Models (LLMs), existing systems remain bottlenecked on pre-defined and hand-crafted model pools. In this work, we investigate the problem of model selection in multi-LLM systems. We propose and systematically evaluate a taxonomy of 9 selection algorithms ranging from diversity of model descriptions, capability-aware behavioral diversity, and LLM-based recruiters. We conduct extensive experiments across two candidate pools of 10 and 32 models, deployed in four model collaboration algorithms, and evaluated across tasks spanning math, coding, QA, and reasoning. Results demonstrate that successful selection algorithms greatly outperform random or heuristics-based teams such as merely selecting the models with top individual performance, by up to 36.1% across settings. Specifically, capability- and training-based selection strategies alleviate selection variance and achieve the best performance, which we recommend to employ before deploying real-world multi-LLM systems. Further analysis reveals that larger candidate pools pose greater challenges to shallow selection heuristics, while algorithms grounded in interacting with candidate models and understanding model capability robustly filter out misaligned, unsafe models, as well as generalizing to novel, out-of-distribution tasks. Together, we establish that principled and informed team selection is critical and present strong model selection algorithms for assembling effective multi-LLM systems.