Transformer Grammars (TGs) enhance language modeling by incorporating syntactic tree structures. Despite the potentially significant impact on model performance of how syntactic trees are linearized in TGs, existing studies rely solely on Depth-First Traversal (DFT) for linearization. In this paper, we expand the traversal design space by exploring Breadth-First Traversal (BFT) and a novel hybrid traversal strategy, Production-Rule Traversal (PRT), which combines the structural lookahead of BFT with the early lexical generation of DFT. We integrate these traversal methods with varying tree configurations and masking strategies, and empirically evaluate their performance on language modeling, syntactic generalization and summarization. We reveal the inherent trade-offs between nested composition and global lookahead, providing actionable recommendations for designing task-aware Transformer Grammars.
展开完整摘要收起摘要↓
Transformer Grammars (TGs) enhance language modeling by incorporating syntactic tree structures. Despite the potentially significant impact on model performance of how syntactic trees are linearized in TGs, existing studies rely solely on Depth-First Traversal (DFT) for linearization. In this paper, we expand the traversal design space by exploring Breadth-First Traversal (BFT) and a novel hybrid traversal strategy, Production-Rule Traversal (PRT), which combines the structural lookahead of BFT with the early lexical generation of DFT. We integrate these traversal methods with varying tree configurations and masking strategies, and empirically evaluate their performance on language modeling, syntactic generalization and summarization. We reveal the inherent trade-offs between nested composition and global lookahead, providing actionable recommendations for designing task-aware Transformer Grammars.
Do different LLM architectures encode high-level concepts in structurally compatible ways? We systematically characterize a geometric-functional universality dissociation: across multiple concept domains and architectural families, moderate geometric convergence coexists with near-perfect functional transfer. Using contrastive-difference CKA (CKA_Delta), a training-free diagnostic that computes kernel alignment on per-sample contrastive differences, we isolate concept-specific convergence from generic similarity — achieving significant discrimination where standard CKA cannot. The dissociation replicates across all six concept domains we test (five with p <= 0.017 geometric discrimination and safety as a converging-functional trend, p = 0.08), including two non-instruction concepts (code-vs-NL, reasoning-vs-recall) validated without system prompts; a single 70B--70B pair provides an observational note that universality may strengthen with scale, requiring replication with additional >=70B models. We position CKA_Delta as a practical regime classifier and architectural outlier detector (Gemma: d = 1.08, AUC = 0.79) rather than an absolute transfer-accuracy predictor, providing a training-free diagnostic for cross-architecture concept monitoring.
展开完整摘要收起摘要↓
Do different LLM architectures encode high-level concepts in structurally compatible ways? We systematically characterize a geometric-functional universality dissociation: across multiple concept domains and architectural families, moderate geometric convergence coexists with near-perfect functional transfer. Using contrastive-difference CKA (CKA_Delta), a training-free diagnostic that computes kernel alignment on per-sample contrastive differences, we isolate concept-specific convergence from generic similarity — achieving significant discrimination where standard CKA cannot. The dissociation replicates across all six concept domains we test (five with p <= 0.017 geometric discrimination and safety as a converging-functional trend, p = 0.08), including two non-instruction concepts (code-vs-NL, reasoning-vs-recall) validated without system prompts; a single 70B--70B pair provides an observational note that universality may strengthen with scale, requiring replication with additional >=70B models. We position CKA_Delta as a practical regime classifier and architectural outlier detector (Gemma: d = 1.08, AUC = 0.79) rather than an absolute transfer-accuracy predictor, providing a training-free diagnostic for cross-architecture concept monitoring.
As Large Language Model (LLM) APIs become ubiquitous, users increasingly rely on black-box fingerprinting to verify that providers are serving the advertised premium models. However, these methods may overlook adversarial providers who manipulate model weights to cheat the fingerprint process. We introduce a novel threat termed fingerprint spoofing, where a malicious provider stealthily serves a weaker model that has been parameter-efficiently fine-tuned to mimic a stronger model, thereby evading user-side fingerprinting. We first formally prove that user-side resource constraints (i.e., finite query budgets and weak fingerprinting classifiers) make current fingerprinting vulnerable to fingerprint spoofing. Guided by this theoretical analysis, we propose GhostPrint, a cost-effective attack framework leveraging surrogate modeling, reward-ranked fine-tuning, and knowledge distillation. Extensive evaluations in both static and continual fingerprinting settings demonstrate that GhostPrint allows weak models to consistently bypass representative fingerprint methods while maintaining utility at a low fine-tuning cost, exposing a critical vulnerability in current LLM fingerprinting pipelines.
展开完整摘要收起摘要↓
As Large Language Model (LLM) APIs become ubiquitous, users increasingly rely on black-box fingerprinting to verify that providers are serving the advertised premium models. However, these methods may overlook adversarial providers who manipulate model weights to cheat the fingerprint process. We introduce a novel threat termed fingerprint spoofing, where a malicious provider stealthily serves a weaker model that has been parameter-efficiently fine-tuned to mimic a stronger model, thereby evading user-side fingerprinting. We first formally prove that user-side resource constraints (i.e., finite query budgets and weak fingerprinting classifiers) make current fingerprinting vulnerable to fingerprint spoofing. Guided by this theoretical analysis, we propose GhostPrint, a cost-effective attack framework leveraging surrogate modeling, reward-ranked fine-tuning, and knowledge distillation. Extensive evaluations in both static and continual fingerprinting settings demonstrate that GhostPrint allows weak models to consistently bypass representative fingerprint methods while maintaining utility at a low fine-tuning cost, exposing a critical vulnerability in current LLM fingerprinting pipelines.
Domain Generalizable (DG) person re-identification (Re-ID) has attracted growing research interest due to its potential for deployment in unseen real-world scenarios. Most existing approaches address DG Re-ID by focusing on training domain-generalizable encoders but ignore the possible refinements in inference stage. In contrast, this work explores an alternative direction which improves inference re-ranking to enhance DG Re-ID. Conventional re-ranking methods typically rely on neighborhood-based distances to refine the initial ranking list, inherently depending on features produced by the Re-ID encoder. However, they deteriorate on target domains since the encoder lacks sufficient generalizability to produce reliable feature distances on unseen scenarios. Inspired by the remarkable generalization capabilities of recent Multimodal Large Language Models (MLLMs), we propose an MLLM-empowered distance metric to improve re-ranking in DG Re-ID. Specifically, we first adapt an MLLM to Re-ID data through supervised fine-tuning, which incorporates a domain-agnostic prompt and a query-candidate hard mining scheme. Then, the adapted MLLM is employed to compute a $μ$-distance during inference, which is robust to domain gap and significantly enhances subsequent re-ranking performance. Our approach is model-agnostic and can be seamlessly integrated into previous re-ranking frameworks. Extensive experiments demonstrate that our approach consistently yields substantial performance improvements across multiple DG Re-ID benchmarks. The code of this work will be released at https://github.com/RikoLi/MUSE soon.
展开完整摘要收起摘要↓
Domain Generalizable (DG) person re-identification (Re-ID) has attracted growing research interest due to its potential for deployment in unseen real-world scenarios. Most existing approaches address DG Re-ID by focusing on training domain-generalizable encoders but ignore the possible refinements in inference stage. In contrast, this work explores an alternative direction which improves inference re-ranking to enhance DG Re-ID. Conventional re-ranking methods typically rely on neighborhood-based distances to refine the initial ranking list, inherently depending on features produced by the Re-ID encoder. However, they deteriorate on target domains since the encoder lacks sufficient generalizability to produce reliable feature distances on unseen scenarios. Inspired by the remarkable generalization capabilities of recent Multimodal Large Language Models (MLLMs), we propose an MLLM-empowered distance metric to improve re-ranking in DG Re-ID. Specifically, we first adapt an MLLM to Re-ID data through supervised fine-tuning, which incorporates a domain-agnostic prompt and a query-candidate hard mining scheme. Then, the adapted MLLM is employed to compute a $μ$-distance during inference, which is robust to domain gap and significantly enhances subsequent re-ranking performance. Our approach is model-agnostic and can be seamlessly integrated into previous re-ranking frameworks. Extensive experiments demonstrate that our approach consistently yields substantial performance improvements across multiple DG Re-ID benchmarks. The code of this work will be released at https://github.com/RikoLi/MUSE soon.
作者Sameera Ramasinghe, Ajanthan Thalaiyasingam, Hadi Mohaghegh Dolatabadi, Chamin Hewa Koneputugodage, Gil Avraham, Violetta Shevchenko, Yan Zuo, Karol Pajak, Alexander Long
Training billion-parameter Transformers is often brittle, with transient loss spikes and divergence that waste compute. Even though the recently developed Edge of Stability (EoS) theory provides a powerful tool to understand and control the stability of optimization methods via the (preconditioned) curvature, these curvature-controlling methods are not popular in large-scale Transformer training due to the complexity of curvature estimation. To this end, we first introduce a fast online estimator of the largest (preconditioned) Hessian eigenvalue (i.e., curvature) based on a warm-started variant for power iteration with Hessian-vector products. We show theoretically, and verify empirically, that the proposed method makes per-iteration curvature tracking feasible at billion parameter scale while being more accurate. Using this tool, we find that training instabilities coincide with surges in preconditioned curvature and that curvature grows with depth. Motivated by these observations, we propose architecture warm-up: progressively growing network depth to carefully control the preconditioned Hessian and stabilize training. Experiments on large Transformers validate that our approach enables efficient curvature tracking and reduces instabilities compared to existing state-of-the-art stabilization techniques without slowing down convergence.
展开完整摘要收起摘要↓
Training billion-parameter Transformers is often brittle, with transient loss spikes and divergence that waste compute. Even though the recently developed Edge of Stability (EoS) theory provides a powerful tool to understand and control the stability of optimization methods via the (preconditioned) curvature, these curvature-controlling methods are not popular in large-scale Transformer training due to the complexity of curvature estimation. To this end, we first introduce a fast online estimator of the largest (preconditioned) Hessian eigenvalue (i.e., curvature) based on a warm-started variant for power iteration with Hessian-vector products. We show theoretically, and verify empirically, that the proposed method makes per-iteration curvature tracking feasible at billion parameter scale while being more accurate. Using this tool, we find that training instabilities coincide with surges in preconditioned curvature and that curvature grows with depth. Motivated by these observations, we propose architecture warm-up: progressively growing network depth to carefully control the preconditioned Hessian and stabilize training. Experiments on large Transformers validate that our approach enables efficient curvature tracking and reduces instabilities compared to existing state-of-the-art stabilization techniques without slowing down convergence.
作者Xuran Li, Guanqin Zhang, Imran Razzak, Hakim Hacid, Eleanna Kafeza, Hao Xue, Flora D. Salim
Current evaluation paradigms for Large Language Model (LLM) personalization rely heavily on brittle surface-matching metrics or computationally expensive LLM-as-a-judge protocols, both of which lack interpretability. To address these limitations, we introduce Natural Language Inference Constraint Verification (NLICV), a scalable, semantically invariant framework that maps sentence meanings to truth-condition sets to verify personalization constraints via a Natural Language Inference (NLI) model. Moving beyond binary scoring, NLICV categorizes LLM behaviors into four distinct modes: personalization, generalization, sycophancy, and failure. Extensive experiments demonstrate that NLICV aligns closely with human annotations while drastically reducing the latency and token costs associated with LLM judges (up to 2100 inference speedup). Finally, through an ablation-based procedure, NLICV pinpoints the exact sentences driving the constraint verification, yielding faithful, understandable evidence for its evaluations.
展开完整摘要收起摘要↓
Current evaluation paradigms for Large Language Model (LLM) personalization rely heavily on brittle surface-matching metrics or computationally expensive LLM-as-a-judge protocols, both of which lack interpretability. To address these limitations, we introduce Natural Language Inference Constraint Verification (NLICV), a scalable, semantically invariant framework that maps sentence meanings to truth-condition sets to verify personalization constraints via a Natural Language Inference (NLI) model. Moving beyond binary scoring, NLICV categorizes LLM behaviors into four distinct modes: personalization, generalization, sycophancy, and failure. Extensive experiments demonstrate that NLICV aligns closely with human annotations while drastically reducing the latency and token costs associated with LLM judges (up to 2100 inference speedup). Finally, through an ablation-based procedure, NLICV pinpoints the exact sentences driving the constraint verification, yielding faithful, understandable evidence for its evaluations.
Large language models (LLMs) remain limited in multi-agent planning because independently generated plans can create coordination failures such as spatial collisions, resource contention, and temporal deadlocks. We introduce Tensor-Coord, a multilinear algebra framework that represents the joint plan of N agents as a third-order tensor \(T \in R^{N \times H \times A}\) over agents, timesteps, and actions. Canonical Polyadic (CP) and Tucker decompositions are used to identify latent coordination structure. The minimal epsilon-approximate CP rank R* defines a computable coordination complexity measure, with \(CC(Pi)=(R*-N)/N\). We prove that R*=N is necessary and sufficient for plan independence. The residual \(E=T-T_{R*}\) defines a conflict score over agent pairs, timesteps, and actions, localizing failures without domain-specific rules. Tucker factors provide interpretable agent roles, temporal phases, and action clusters that are converted into natural language constraints for iterative LLM replanning. Experiments on multi-robot delivery tasks across Easy (2 agents, 5x5 grid), Medium (3 agents, 5x5 grid), and Hard (4 agents, 5x5 grid) settings show convergence to conflict-free plans in 100% of 2-agent cases within 1.4 iterations on average, 80% of 3-agent cases within 3.2 iterations, and 60% of 4-agent cases within 4.0 iterations. CP rank scaled approximately linearly as \(R*(N) = 3.9N + 0.5\), supporting its use as a predictor of coordination complexity.
展开完整摘要收起摘要↓
Large language models (LLMs) remain limited in multi-agent planning because independently generated plans can create coordination failures such as spatial collisions, resource contention, and temporal deadlocks. We introduce Tensor-Coord, a multilinear algebra framework that represents the joint plan of N agents as a third-order tensor \(T \in R^{N \times H \times A}\) over agents, timesteps, and actions. Canonical Polyadic (CP) and Tucker decompositions are used to identify latent coordination structure. The minimal epsilon-approximate CP rank R* defines a computable coordination complexity measure, with \(CC(Pi)=(R*-N)/N\). We prove that R*=N is necessary and sufficient for plan independence. The residual \(E=T-T_{R*}\) defines a conflict score over agent pairs, timesteps, and actions, localizing failures without domain-specific rules. Tucker factors provide interpretable agent roles, temporal phases, and action clusters that are converted into natural language constraints for iterative LLM replanning. Experiments on multi-robot delivery tasks across Easy (2 agents, 5x5 grid), Medium (3 agents, 5x5 grid), and Hard (4 agents, 5x5 grid) settings show convergence to conflict-free plans in 100% of 2-agent cases within 1.4 iterations on average, 80% of 3-agent cases within 3.2 iterations, and 60% of 4-agent cases within 4.0 iterations. CP rank scaled approximately linearly as \(R*(N) = 3.9N + 0.5\), supporting its use as a predictor of coordination complexity.
Recent work motivates moving large language model (LLM) evaluation from mean-based to tail-aware metrics, including conditional value-at-risk and tail-index estimates of reward-model error. We ask whether the canonical extreme-value-theory tail-index parameter, which isolates how heavy a tail is from how large the tail mass is, adds discriminative information beyond the mean and a standard tail-magnitude statistic in LLM evaluation. We pre-register a protocol covering admissibility, goodness-of-fit, threshold-stability, and effect-size requirements for any positive tail-shape claim. The protocol is the contribution of this paper; the empirical study below is a demonstration of what its gates catch. Applied to a standard LLM toxicity-evaluation setup under two structurally different scorer families, the protocol catches three distinct modes of false positives that a naive analysis would have published, and rejects the headline tail-shape claim on both scorers. We conclude that tail-shape estimation in the LLM toxicity-evaluation setups we examined is more fragile than the recent literature suggests, and recommend the protocol as a starting point for tail-index claims in similar setups.
展开完整摘要收起摘要↓
Recent work motivates moving large language model (LLM) evaluation from mean-based to tail-aware metrics, including conditional value-at-risk and tail-index estimates of reward-model error. We ask whether the canonical extreme-value-theory tail-index parameter, which isolates how heavy a tail is from how large the tail mass is, adds discriminative information beyond the mean and a standard tail-magnitude statistic in LLM evaluation. We pre-register a protocol covering admissibility, goodness-of-fit, threshold-stability, and effect-size requirements for any positive tail-shape claim. The protocol is the contribution of this paper; the empirical study below is a demonstration of what its gates catch. Applied to a standard LLM toxicity-evaluation setup under two structurally different scorer families, the protocol catches three distinct modes of false positives that a naive analysis would have published, and rejects the headline tail-shape claim on both scorers. We conclude that tail-shape estimation in the LLM toxicity-evaluation setups we examined is more fragile than the recent literature suggests, and recommend the protocol as a starting point for tail-index claims in similar setups.
Equipping Large Language Model (LLM) agents with effective skills is crucial for solving complex tasks in real-world systems like OpenClaw. In this work, we aim to develop a framework that automatically constructs such reusable skills to enhance LLMs in tool use, multi-step reasoning, and dynamic environment interaction. To this end, we propose Collective Skill Tree Search (CSTS), a novel tree-search-based skill construction framework that constructs structured, diverse and generalizable tree of skills. The core idea of CSTS is to leverage collective intelligence to jointly search, identify and compose effective skills via two iterative phases: Collective Skill Node Generation (CSN-Gen) and Collective Skill Node Assessment (CSN-Assess). CSN-Gen exploits collective knowledge from multiple models to explore diverse candidate skills for each subtask, enabling comprehensive skill exploration. CSN-Assess employs multiple models as judges to evaluate and select skill nodes with two scoring mechanisms: (1) collective quality scoring that aggregates independent evaluations to produce a robust estimate of skill effectiveness, and (2) collective transferability scoring that explicitly verifies whether a skill generalizes well across different models. With CSTS, we construct a set of comprehensive tree of skills along with skill-augmented training data, enabling models to effectively learn and utilize skills. Besides, we introduce Collective Skill Reinforcement Learning, which actively selects multiple relevant skills from the tree to broaden solution-space exploration, avoid being trapped by a single skill and its resulting homogeneous or suboptimal solutions. As a result, our trained model, OpenClaw-Skill, exhibits outstanding agentic capabilities in long-horizon planning, tool use and generalization over challenging benchmarks.
展开完整摘要收起摘要↓
Equipping Large Language Model (LLM) agents with effective skills is crucial for solving complex tasks in real-world systems like OpenClaw. In this work, we aim to develop a framework that automatically constructs such reusable skills to enhance LLMs in tool use, multi-step reasoning, and dynamic environment interaction. To this end, we propose Collective Skill Tree Search (CSTS), a novel tree-search-based skill construction framework that constructs structured, diverse and generalizable tree of skills. The core idea of CSTS is to leverage collective intelligence to jointly search, identify and compose effective skills via two iterative phases: Collective Skill Node Generation (CSN-Gen) and Collective Skill Node Assessment (CSN-Assess). CSN-Gen exploits collective knowledge from multiple models to explore diverse candidate skills for each subtask, enabling comprehensive skill exploration. CSN-Assess employs multiple models as judges to evaluate and select skill nodes with two scoring mechanisms: (1) collective quality scoring that aggregates independent evaluations to produce a robust estimate of skill effectiveness, and (2) collective transferability scoring that explicitly verifies whether a skill generalizes well across different models. With CSTS, we construct a set of comprehensive tree of skills along with skill-augmented training data, enabling models to effectively learn and utilize skills. Besides, we introduce Collective Skill Reinforcement Learning, which actively selects multiple relevant skills from the tree to broaden solution-space exploration, avoid being trapped by a single skill and its resulting homogeneous or suboptimal solutions. As a result, our trained model, OpenClaw-Skill, exhibits outstanding agentic capabilities in long-horizon planning, tool use and generalization over challenging benchmarks.
The adoption of AI agents is increasing rapidly. Terminal AI agents, i.e., AI agents that run in terminal environments, are a widely used type of AI agents. Terminal AI agents rely heavily on shell command execution to interact with the host systems. They adopt a three-list command-gating mechanism to mitigate security risks introduced by command execution, with denylists serving as the load-bearing component. However, modern operating systems often ship a large, ever-expanding set of shell commands with complex functionalities. Our observation is that even a built-in denylist of Claude Code, well-maintained by its developers, can overlook bypass commands that invalidate its effectiveness. Such negligence leads to fragile command denylists that cannot even block operations that practitioners expect them to block. This paper presents the first systematic characterization of command denylist fragility in terminal AI agents. The paper formalizes the command denylist fragility problem and proposes an LLM-driven pipeline, CmdNeedle, to detect such fragility. It prompts the LLM to propose possible bypasses and iteratively repairs them using feedback from a validator that executes them in a sandbox. In the evaluation, we applied CmdNeedle to 1,709 real-world command denylists (containing 13,332 denylist rules) collected from GitHub. The evaluation shows several key findings, including that 69.0--98.6% of the denylists are fragile, that this fragility occurs consistently across projects and agents, and the validity of several possible root causes for this fragility. Our pipeline and findings will hopefully facilitate future research and practice regarding the command denylists used by AI agents.
展开完整摘要收起摘要↓
The adoption of AI agents is increasing rapidly. Terminal AI agents, i.e., AI agents that run in terminal environments, are a widely used type of AI agents. Terminal AI agents rely heavily on shell command execution to interact with the host systems. They adopt a three-list command-gating mechanism to mitigate security risks introduced by command execution, with denylists serving as the load-bearing component. However, modern operating systems often ship a large, ever-expanding set of shell commands with complex functionalities. Our observation is that even a built-in denylist of Claude Code, well-maintained by its developers, can overlook bypass commands that invalidate its effectiveness. Such negligence leads to fragile command denylists that cannot even block operations that practitioners expect them to block. This paper presents the first systematic characterization of command denylist fragility in terminal AI agents. The paper formalizes the command denylist fragility problem and proposes an LLM-driven pipeline, CmdNeedle, to detect such fragility. It prompts the LLM to propose possible bypasses and iteratively repairs them using feedback from a validator that executes them in a sandbox. In the evaluation, we applied CmdNeedle to 1,709 real-world command denylists (containing 13,332 denylist rules) collected from GitHub. The evaluation shows several key findings, including that 69.0--98.6% of the denylists are fragile, that this fragility occurs consistently across projects and agents, and the validity of several possible root causes for this fragility. Our pipeline and findings will hopefully facilitate future research and practice regarding the command denylists used by AI agents.
作者Wasi Uddin Ahmad, Nikolai Ludwig, Somshubra Majumdar, Boris Ginsburg
The path toward autonomous software engineering is currently bottlenecked by a severe deficit of diverse, large-scale trajectory data. We address this by introducing \ourdataset, an expansive dataset of 207,489 agentic trajectories spanning nine programming languages (Python, Go, TS, JS, Rust, Java, PHP, C, C++). Sourced from 20,000 real-world PRs via OpenHands and SWE-agent harnesses, the dataset utilizes a hybrid-reasoning synthesis: Minimax-M2.5 generates trajectories with explicit "thinking" processes, while Qwen3.5-122B provides high-quality "non-thinking" traces. Filtered for permissive licenses (MIT, Apache, BSD) from SWE-rebench-V2, this data facilitates the training of models capable of long-horizon reasoning. We validate the dataset by fine-tuning the Qwen3-30B-A3B series (Thinking, Instruct, and Coder). The best performing model achieves resolve rates of 61.7% on SWE-bench Verified, 57.1% on SWE-bench Multilingual, and 36.8% on SWE-bench Pro. These results establish Open-SWE-Traces as a premier resource for distilling human-level software engineering capabilities into efficient, open-source agentic LLMs.
展开完整摘要收起摘要↓
The path toward autonomous software engineering is currently bottlenecked by a severe deficit of diverse, large-scale trajectory data. We address this by introducing \ourdataset, an expansive dataset of 207,489 agentic trajectories spanning nine programming languages (Python, Go, TS, JS, Rust, Java, PHP, C, C++). Sourced from 20,000 real-world PRs via OpenHands and SWE-agent harnesses, the dataset utilizes a hybrid-reasoning synthesis: Minimax-M2.5 generates trajectories with explicit "thinking" processes, while Qwen3.5-122B provides high-quality "non-thinking" traces. Filtered for permissive licenses (MIT, Apache, BSD) from SWE-rebench-V2, this data facilitates the training of models capable of long-horizon reasoning. We validate the dataset by fine-tuning the Qwen3-30B-A3B series (Thinking, Instruct, and Coder). The best performing model achieves resolve rates of 61.7% on SWE-bench Verified, 57.1% on SWE-bench Multilingual, and 36.8% on SWE-bench Pro. These results establish Open-SWE-Traces as a premier resource for distilling human-level software engineering capabilities into efficient, open-source agentic LLMs.
作者Yuhan Wang, Manqing Wang, Yixuan Lu, Zhaoyue Peng, Shengda Lin
Audit risk assessment increasingly benefits from combining heterogeneous evidence sources, yet existing approaches typically produce point predictions without quantifying how well different evidence streams agree. We propose UMAR (Uncertainty-Aware Multi-Agent Risk Assessment), a framework that employs three specialized agents: an MD&A Text Agent, a Financial Ratio Agent, and a CAM Agent, each producing independent risk scores with calibrated uncertainty estimates. An Uncertainty Aggregator based on Dempster-Shafer evidence theory fuses these scores while explicitly measuring inter-agent conflict. We evaluate UMAR on a U.S. dataset of 3,200 firm-year observations from SEC 10-K filings (2019-2023), with financial restatement as the target label. Experimental results show that UMAR achieves an AUROC of 0.782 and a PR-AUC of 0.341, outperforming logistic regression, XGBoost, FinBERT, and single-agent and dual-agent LLM baselines. UMAR attains the lowest expected calibration error (ECE = 0.052) among all methods and identifies evidence-conflict patterns that correlate with actual restatement risk, offering auditors potentially actionable and interpretable risk signals.
展开完整摘要收起摘要↓
Audit risk assessment increasingly benefits from combining heterogeneous evidence sources, yet existing approaches typically produce point predictions without quantifying how well different evidence streams agree. We propose UMAR (Uncertainty-Aware Multi-Agent Risk Assessment), a framework that employs three specialized agents: an MD&A Text Agent, a Financial Ratio Agent, and a CAM Agent, each producing independent risk scores with calibrated uncertainty estimates. An Uncertainty Aggregator based on Dempster-Shafer evidence theory fuses these scores while explicitly measuring inter-agent conflict. We evaluate UMAR on a U.S. dataset of 3,200 firm-year observations from SEC 10-K filings (2019-2023), with financial restatement as the target label. Experimental results show that UMAR achieves an AUROC of 0.782 and a PR-AUC of 0.341, outperforming logistic regression, XGBoost, FinBERT, and single-agent and dual-agent LLM baselines. UMAR attains the lowest expected calibration error (ECE = 0.052) among all methods and identifies evidence-conflict patterns that correlate with actual restatement risk, offering auditors potentially actionable and interpretable risk signals.
We present TickingCollabBench, a Minecraft-based multi-agent benchmark for a novel class of time-sensitive complementary collaboration tasks. Our benchmark reflects four core characteristics of real-world collaboration: agent heterogeneity, mandatory collaboration, dynamic environments, and strict real-time constraints with failure risks. To enable this, we develop the TickingCollab framework, which supports the generation of diverse dynamic environments and abstracts Minecraft's primitive APIs to enable declarative YAML task specifications for composing these events. Building on this, we design a feasibility-aware automated benchmark generation pipeline, where an LLM drafts structurally diverse task configurations and feasibility verifier filters out invalid ones using approximate constraints. Evaluations demonstrate that lang latency and inherent difficulty of coordinating under partial observability and agent heterogeneity cause LLMs to frequently fail under dynamic environments and fall significantly short of a global-knowledge oracle.
展开完整摘要收起摘要↓
We present TickingCollabBench, a Minecraft-based multi-agent benchmark for a novel class of time-sensitive complementary collaboration tasks. Our benchmark reflects four core characteristics of real-world collaboration: agent heterogeneity, mandatory collaboration, dynamic environments, and strict real-time constraints with failure risks. To enable this, we develop the TickingCollab framework, which supports the generation of diverse dynamic environments and abstracts Minecraft's primitive APIs to enable declarative YAML task specifications for composing these events. Building on this, we design a feasibility-aware automated benchmark generation pipeline, where an LLM drafts structurally diverse task configurations and feasibility verifier filters out invalid ones using approximate constraints. Evaluations demonstrate that lang latency and inherent difficulty of coordinating under partial observability and agent heterogeneity cause LLMs to frequently fail under dynamic environments and fall significantly short of a global-knowledge oracle.
Where an LLM sits in an agent memory pipeline — between the recall plane that retrieves stored facts (extensively benchmarked) and the control plane that mutates them via supersede, release, purge (largely untested) — shapes which forgetting failure modes the system recovers. Comparing thirteen system configurations on a 385-case adversarial surface, we observe three placement regimes with partly complementary coverage: deterministic primitives suffice for lexical/temporal categories but fail canonicalization (5% on identifier-obfuscation, 0% on cross-lingual); inscribe-time LLM recovers canonicalization (100%) but cannot help intent-aware deletion (0% on prefix-collision and compound-fact); a mutation-time hook recovers intent-aware deletion (78-85%) and brightens nearly all categories simultaneously (91.7-93.2% overall, $0.17 per 385-case run, 2.3s/case mutation latency vs. 64-191ms/case deterministic, recall path unchanged). We expose the trade-off via ForgetEval, a 1000-case templated suite plus a 385-case adversarial layer (132 hand-crafted + 253 LLM-drafted oracle-validated) scored by deterministic substring match, paired with a six-method Adapter Protocol with honest N/A scoring that lets heterogeneous memory stores enter in 130 lines. Admission is corroborated by 10-annotator IAA (Fleiss' kappa = 0.958) and a 77-case external-authored subset (four blind contributors) that replicates the canonicalization asymmetry and amplifies the joint-placement lift (+27.8 pt). Production failures are predominantly forgetting failures rather than recall failures, yet existing benchmarks measure only recall. ForgetEval and all adapters are released under MIT.
展开完整摘要收起摘要↓
Where an LLM sits in an agent memory pipeline — between the recall plane that retrieves stored facts (extensively benchmarked) and the control plane that mutates them via supersede, release, purge (largely untested) — shapes which forgetting failure modes the system recovers. Comparing thirteen system configurations on a 385-case adversarial surface, we observe three placement regimes with partly complementary coverage: deterministic primitives suffice for lexical/temporal categories but fail canonicalization (5% on identifier-obfuscation, 0% on cross-lingual); inscribe-time LLM recovers canonicalization (100%) but cannot help intent-aware deletion (0% on prefix-collision and compound-fact); a mutation-time hook recovers intent-aware deletion (78-85%) and brightens nearly all categories simultaneously (91.7-93.2% overall, $0.17 per 385-case run, 2.3s/case mutation latency vs. 64-191ms/case deterministic, recall path unchanged). We expose the trade-off via ForgetEval, a 1000-case templated suite plus a 385-case adversarial layer (132 hand-crafted + 253 LLM-drafted oracle-validated) scored by deterministic substring match, paired with a six-method Adapter Protocol with honest N/A scoring that lets heterogeneous memory stores enter in 130 lines. Admission is corroborated by 10-annotator IAA (Fleiss' kappa = 0.958) and a 77-case external-authored subset (four blind contributors) that replicates the canonicalization asymmetry and amplifies the joint-placement lift (+27.8 pt). Production failures are predominantly forgetting failures rather than recall failures, yet existing benchmarks measure only recall. ForgetEval and all adapters are released under MIT.
作者Linghua Zhang, Jun Wang, Jingtong Wu, Zhisong Zhang
Large language model (LLM) agents have made rapid progress on short-horizon, well-scoped tasks, yet their ability to sustain coherent decisions in dynamic long-horizon environments remains uncertain. We introduce RetailBench, a data-grounded simulation benchmark for evaluating tool-using LLM agents in single-store supermarket operation. RetailBench models retail management as a partially observable decision process and is designed to support thousand-day-scale simulations. In this environment, agents must manage pricing, replenishment, supplier selection, shelf assortment, inventory aging, customer feedback, external events, and cash-flow constraints. We evaluate seven contemporary LLMs under representative agent frameworks over a 180-day evaluation horizon and compare them with a privileged oracle policy. Results show substantial variation across models: only a small subset survives the full evaluation horizon, and even the strongest LLM runs remain substantially behind the oracle policy in final net worth and sales outcomes. Behavioral analysis attributes these gaps to incomplete evidence acquisition, surface-level decision making, and the lack of a consistent long-horizon policy. RetailBench provides a controlled testbed for studying reliable autonomy in economically grounded long-horizon decision-making.
展开完整摘要收起摘要↓
Large language model (LLM) agents have made rapid progress on short-horizon, well-scoped tasks, yet their ability to sustain coherent decisions in dynamic long-horizon environments remains uncertain. We introduce RetailBench, a data-grounded simulation benchmark for evaluating tool-using LLM agents in single-store supermarket operation. RetailBench models retail management as a partially observable decision process and is designed to support thousand-day-scale simulations. In this environment, agents must manage pricing, replenishment, supplier selection, shelf assortment, inventory aging, customer feedback, external events, and cash-flow constraints. We evaluate seven contemporary LLMs under representative agent frameworks over a 180-day evaluation horizon and compare them with a privileged oracle policy. Results show substantial variation across models: only a small subset survives the full evaluation horizon, and even the strongest LLM runs remain substantially behind the oracle policy in final net worth and sales outcomes. Behavioral analysis attributes these gaps to incomplete evidence acquisition, surface-level decision making, and the lack of a consistent long-horizon policy. RetailBench provides a controlled testbed for studying reliable autonomy in economically grounded long-horizon decision-making.
Open-source LLM agent ecosystems are growing rapidly, yet the security of community-contributed skills - modular tool definitions that extend agent capabilities - remains largely unvetted. The gap we fill: existing scanners operate at the code layer and are structurally blind to instruction-layer and multi-agent risk - natural-language directives that hijack an agent, exfiltrate data through encoded side channels, or chain harm across pipelines - so what is needed is a semantic, multi-dimensional vetting system rather than another signature matcher. We present SKILLVETBENCH, a live public leaderboard on Hugging Face that uses an LLM-as-Judge to vet agent skills. What is new: SARS (Skill Agentic Risk Score), a five-dimensional agentic-risk metric with a principled weighted formula for instruction-following systems. What is integrated: full CVSS v4.0 vector decomposition and a ClawHub dual-view that places our LLM-generated review beside the official marketplace verdict. What is demonstrated: drawing on our companion benchmark paper [ 1], the LLM-as-Judge stage achieves zero false negatives across 78 confirmed-malicious skills and zero false positives across 22 benign controls, while the best static baseline (SKILLSIEVE) still misses 15%; for instruction-layer categories such as Prompt Injection and Memory Poisoning, conventional tools miss between 89% and 100% of threats (e.g., CODEBERT detects none of nine memory-poisoning skills). Detection rates vary from 35% to 95% across four LLM evaluators, motivating ensemble scoring in production deployments.
展开完整摘要收起摘要↓
Open-source LLM agent ecosystems are growing rapidly, yet the security of community-contributed skills - modular tool definitions that extend agent capabilities - remains largely unvetted. The gap we fill: existing scanners operate at the code layer and are structurally blind to instruction-layer and multi-agent risk - natural-language directives that hijack an agent, exfiltrate data through encoded side channels, or chain harm across pipelines - so what is needed is a semantic, multi-dimensional vetting system rather than another signature matcher. We present SKILLVETBENCH, a live public leaderboard on Hugging Face that uses an LLM-as-Judge to vet agent skills. What is new: SARS (Skill Agentic Risk Score), a five-dimensional agentic-risk metric with a principled weighted formula for instruction-following systems. What is integrated: full CVSS v4.0 vector decomposition and a ClawHub dual-view that places our LLM-generated review beside the official marketplace verdict. What is demonstrated: drawing on our companion benchmark paper [ 1], the LLM-as-Judge stage achieves zero false negatives across 78 confirmed-malicious skills and zero false positives across 22 benign controls, while the best static baseline (SKILLSIEVE) still misses 15%; for instruction-layer categories such as Prompt Injection and Memory Poisoning, conventional tools miss between 89% and 100% of threats (e.g., CODEBERT detects none of nine memory-poisoning skills). Detection rates vary from 35% to 95% across four LLM evaluators, motivating ensemble scoring in production deployments.
Large Language Models (LLMs) are increasingly assessed and utilized in the field of Argument Mining (AM), thanks to their strong general reasoning capabilities. However, standard training-free models often miss sophisticated details, specifically in contexts where two parts of the text have to be analyzed together. Furthermore, self-correction mechanisms tend to reinforce initial hallucinations in reasoning. Overcoming these limitations typically requires expensive, domain-specific supervised fine-tuning. Recent work has shown that a multi-agent paradigm can address such weaknesses for the component classification task through dialectical refinement with a Proponent-Opponent-Judge architecture, setting a promising direction for training-free approaches in the field. In this paper, we extend and evaluate this framework on the Argument Relation Identification and Classification (ARIC) task, reformulating it as a debate over component pairs. Besides that, we introduce a confidence gating mechanism that enables debating only on the uncertain cases and accepting the initial prediction when confidence is high. On the UKP Argument Annotated Essays v2 corpus, we demonstrate that the selective debate achieves the highest Macro F1 among all training-free methods, while debate over all samples degrades performance below that of one of the baselines. All generative approaches also outperform fine-tuned RoBERTa models on Macro F1, suggesting that the under-representation of the Attack class was more damaging to supervised fine-tuning than to inference-only models. Additionally, our framework produces human-readable debate transcripts, offering interpretability absent from both single-agent and supervised classifiers.
展开完整摘要收起摘要↓
Large Language Models (LLMs) are increasingly assessed and utilized in the field of Argument Mining (AM), thanks to their strong general reasoning capabilities. However, standard training-free models often miss sophisticated details, specifically in contexts where two parts of the text have to be analyzed together. Furthermore, self-correction mechanisms tend to reinforce initial hallucinations in reasoning. Overcoming these limitations typically requires expensive, domain-specific supervised fine-tuning. Recent work has shown that a multi-agent paradigm can address such weaknesses for the component classification task through dialectical refinement with a Proponent-Opponent-Judge architecture, setting a promising direction for training-free approaches in the field. In this paper, we extend and evaluate this framework on the Argument Relation Identification and Classification (ARIC) task, reformulating it as a debate over component pairs. Besides that, we introduce a confidence gating mechanism that enables debating only on the uncertain cases and accepting the initial prediction when confidence is high. On the UKP Argument Annotated Essays v2 corpus, we demonstrate that the selective debate achieves the highest Macro F1 among all training-free methods, while debate over all samples degrades performance below that of one of the baselines. All generative approaches also outperform fine-tuned RoBERTa models on Macro F1, suggesting that the under-representation of the Attack class was more damaging to supervised fine-tuning than to inference-only models. Additionally, our framework produces human-readable debate transcripts, offering interpretability absent from both single-agent and supervised classifiers.
作者Zixin Rao, Wentian Zhu, Chan Aristella Lu, Zhaorun Chen, Wei Niu, Le Guan, Bo Li, Zhen Xiang
Large language model (LLM) agents increasingly rely on long-term memory to support complex task execution, user personalization, and domain adaptation. Meanwhile, emerging access-control mechanisms for LLM agents are being explored to block policy-violating requests and prevent misuse. We reveal a novel attack surface arising from agent memory operations: prohibited content that would trigger access control can be fragmented across interactions, stored in long-term memory in benign-appearing form, and later reconstructed through memory retrieval without appearing explicitly in the final user query. We propose FragFuse, the first attack that enables unprivileged users to bypass agent access control by exploiting this temporal channel introduced by long-term memory. FragFuse operates in three stages: (1) identifying rejection-responsive fragments via black-box adaptive querying with fragment masking; (2) injecting these fragments into memory using marker carrier queries; and (3) retrieving and fusing the stored fragments through a follow-up attack query. Although FragFuse can be instantiated manually for individual agents, we further develop a surrogate-based optimization scheme that tunes fusion instructions and marker designs, enabling automated attack generation without violating the attacker's threat-model assumptions. We evaluate FragFuse across four representative agent settings and task domains, covering three state-of-the-art agent access-control mechanisms. FragFuse achieves an average bypass success rate of 86.3% and an average end-to-end harmful task success rate of 41.1% across all settings, with only 4.4% average task-success degradation compared with configurations without access control. We also show that alternative defenses, including state-of-the-art prompt-injection detectors and perplexity detectors, do not effectively address this attack.
展开完整摘要收起摘要↓
Large language model (LLM) agents increasingly rely on long-term memory to support complex task execution, user personalization, and domain adaptation. Meanwhile, emerging access-control mechanisms for LLM agents are being explored to block policy-violating requests and prevent misuse. We reveal a novel attack surface arising from agent memory operations: prohibited content that would trigger access control can be fragmented across interactions, stored in long-term memory in benign-appearing form, and later reconstructed through memory retrieval without appearing explicitly in the final user query. We propose FragFuse, the first attack that enables unprivileged users to bypass agent access control by exploiting this temporal channel introduced by long-term memory. FragFuse operates in three stages: (1) identifying rejection-responsive fragments via black-box adaptive querying with fragment masking; (2) injecting these fragments into memory using marker carrier queries; and (3) retrieving and fusing the stored fragments through a follow-up attack query. Although FragFuse can be instantiated manually for individual agents, we further develop a surrogate-based optimization scheme that tunes fusion instructions and marker designs, enabling automated attack generation without violating the attacker's threat-model assumptions. We evaluate FragFuse across four representative agent settings and task domains, covering three state-of-the-art agent access-control mechanisms. FragFuse achieves an average bypass success rate of 86.3% and an average end-to-end harmful task success rate of 41.1% across all settings, with only 4.4% average task-success degradation compared with configurations without access control. We also show that alternative defenses, including state-of-the-art prompt-injection detectors and perplexity detectors, do not effectively address this attack.
We introduce GRACE-DS, a Guarded Reward-guided Agent Correction Environment in Data Science for pre-deployment evaluation of LLM-powered AutoML agents. GRACE-DS is a set of evaluation metrics in an isolated environment that can be applied to tabular ML tasks specific to a particular organization. It exposes agents to realistic workflow stages, from planning and data inspection through feature engineering, model development, validation, and code repair to final submission, while hidden executable validators measure not only final predictive performance but also leakage avoidance, reproducibility, protocol validity, correction behavior, and reward alignment. The strongest structured regime, flexible iterative interaction (our approach), achieves higher end-to-end normalized hidden-test quality than single-shot generation, unstructured interaction, and restart-based baselines, while also improving protocol-valid completion. Validated across more than 7,000 episodes, these results establish GRACE-DS as a robust platform for assessing the capacity of LLM-based AutoML agents to execute machine learning workflows under production-like conditions and in accordance with organization-specific requirements.
展开完整摘要收起摘要↓
We introduce GRACE-DS, a Guarded Reward-guided Agent Correction Environment in Data Science for pre-deployment evaluation of LLM-powered AutoML agents. GRACE-DS is a set of evaluation metrics in an isolated environment that can be applied to tabular ML tasks specific to a particular organization. It exposes agents to realistic workflow stages, from planning and data inspection through feature engineering, model development, validation, and code repair to final submission, while hidden executable validators measure not only final predictive performance but also leakage avoidance, reproducibility, protocol validity, correction behavior, and reward alignment. The strongest structured regime, flexible iterative interaction (our approach), achieves higher end-to-end normalized hidden-test quality than single-shot generation, unstructured interaction, and restart-based baselines, while also improving protocol-valid completion. Validated across more than 7,000 episodes, these results establish GRACE-DS as a robust platform for assessing the capacity of LLM-based AutoML agents to execute machine learning workflows under production-like conditions and in accordance with organization-specific requirements.
Vision-Language Models (VLMs) are AI systems that process both images and text, yet they often struggle with compositional visual reasoning questions that require chaining multiple steps together, such as identifying objects, counting them, and comparing the results. Existing approaches improve this reasoning by training models on human-written step-by-step explanations, but creating these annotations is expensive and difficult to scale. We propose a self-questioning framework that trains a VLM to break visual questions into smaller sub-questions and answer each one before producing a final response, using a reinforcement learning algorithm called Group Relative Policy Optimization (GRPO). The model is never shown examples of how to decompose questions, it discovers this behavior on its own, guided by a reward signal that scores whether the output contains sub-questions and whether the final answer is correct. We apply this framework to a 3-billion-parameter model, training on both synthetic scenes of geometric shapes (CLEVR) and real-world photographs (A-OKVQA). On A-OKVQA, both self-questioning and standard reinforcement learning substantially improve accuracy over the untrained model (52.2% and 51.6% vs. 46.8%). We introduce the first self-questioning VLM by rewarding not only the final answer like standard RL but additionally for generating intermediate sub-questions, enabling it to discover compositional decomposition strategies. These results suggest that teaching AI systems to ask themselves intermediate questions is a promising strategy for complex visual reasoning, particularly when the difficulty of a question warrants explicit step-by-step decomposition.
展开完整摘要收起摘要↓
Vision-Language Models (VLMs) are AI systems that process both images and text, yet they often struggle with compositional visual reasoning questions that require chaining multiple steps together, such as identifying objects, counting them, and comparing the results. Existing approaches improve this reasoning by training models on human-written step-by-step explanations, but creating these annotations is expensive and difficult to scale. We propose a self-questioning framework that trains a VLM to break visual questions into smaller sub-questions and answer each one before producing a final response, using a reinforcement learning algorithm called Group Relative Policy Optimization (GRPO). The model is never shown examples of how to decompose questions, it discovers this behavior on its own, guided by a reward signal that scores whether the output contains sub-questions and whether the final answer is correct. We apply this framework to a 3-billion-parameter model, training on both synthetic scenes of geometric shapes (CLEVR) and real-world photographs (A-OKVQA). On A-OKVQA, both self-questioning and standard reinforcement learning substantially improve accuracy over the untrained model (52.2% and 51.6% vs. 46.8%). We introduce the first self-questioning VLM by rewarding not only the final answer like standard RL but additionally for generating intermediate sub-questions, enabling it to discover compositional decomposition strategies. These results suggest that teaching AI systems to ask themselves intermediate questions is a promising strategy for complex visual reasoning, particularly when the difficulty of a question warrants explicit step-by-step decomposition.
4-bit quantization significantly reduces the memory footprint and accelerates the inference of large language models (LLMs). However, its limited bit-width representation struggles to faithfully capture both dense common values (inliers) and rare large-magnitude values (outliers), causing substantial accuracy degradation. Existing mixed-precision methods mitigate this by retaining outliers in high precision, but at the cost of breaking the uniformity of low-bit execution, introducing precision conversion and extra data movement that undermine practical speedup. We propose MosaicQuant, a unified 4-bit LLM quantization paradigm built on a novel principle of inlier--outlier disaggregation. Rather than elevating outlier precision, MosaicQuant quantizes the full weight matrix into a dense 4-bit base component, where inliers are captured faithfully while outlier are inevitably quantized. A sparse 4-bit residual component is then introduced to compensate for these quantization errors, selectively targeting the most error-critical weight blocks where output distortion is shown to be concentrated. However, a unified representation alone is insufficient, as naïvely executing the sparse residual as a separate kernel still breaks the unified low-bit inference pipeline. To bridge this gap, we introduce ZipperEngine, which fuses sparse block computation into the dense 4-bit GEMM kernel via an overlapped pipeline, unifying not only the representation but also the execution into a single coherent low-bit inference pipeline. Extensive experiments on LLaMA3 and Qwen3 demonstrate that MosaicQuant preserves near-FP16 accuracy while achieving up to $1.24\times$ speedup over the W16A16 baseline.
展开完整摘要收起摘要↓
4-bit quantization significantly reduces the memory footprint and accelerates the inference of large language models (LLMs). However, its limited bit-width representation struggles to faithfully capture both dense common values (inliers) and rare large-magnitude values (outliers), causing substantial accuracy degradation. Existing mixed-precision methods mitigate this by retaining outliers in high precision, but at the cost of breaking the uniformity of low-bit execution, introducing precision conversion and extra data movement that undermine practical speedup. We propose MosaicQuant, a unified 4-bit LLM quantization paradigm built on a novel principle of inlier--outlier disaggregation. Rather than elevating outlier precision, MosaicQuant quantizes the full weight matrix into a dense 4-bit base component, where inliers are captured faithfully while outlier are inevitably quantized. A sparse 4-bit residual component is then introduced to compensate for these quantization errors, selectively targeting the most error-critical weight blocks where output distortion is shown to be concentrated. However, a unified representation alone is insufficient, as naïvely executing the sparse residual as a separate kernel still breaks the unified low-bit inference pipeline. To bridge this gap, we introduce ZipperEngine, which fuses sparse block computation into the dense 4-bit GEMM kernel via an overlapped pipeline, unifying not only the representation but also the execution into a single coherent low-bit inference pipeline. Extensive experiments on LLaMA3 and Qwen3 demonstrate that MosaicQuant preserves near-FP16 accuracy while achieving up to $1.24\times$ speedup over the W16A16 baseline.
作者Ruicheng Ao, Jing Dong, Gan Luo, David Simchi-Levi
In large language model (LLM) serving, each request accumulates persistent graphics processing unit (GPU) memory during service as its key-value cache grows with every generated token. Under high concurrency, aggregate memory usage therefore increases endogenously over time: the service process itself creates future capacity pressure. When memory capacity is exceeded, systems evict active requests, discarding cached state and restarting them later, which wastes computation and reduces throughput. We develop a discrete-time dynamical model of memory-constrained LLM inference that captures admission, memory growth, and eviction under continuous batching. In the saturated-input regime, the system admits both eviction-free fixed points and limit cycles with evictions. For homogeneous workloads, we show that the eviction-free equilibrium is unstable and that, except for a Lebesgue-measure-zero exact-capture set, the system converges to a unique worst-case limit cycle that is asymptotically stable outside this exceptional set, with throughput losses as large as 50%. For heterogeneous workloads, we prove a stability criterion in the two-class common-input setting and explain how the survival-polynomial mechanism generalizes to multiple classes and heterogeneous-input lengths. Under an input-dominated scaling regime, coprime decoding lengths stabilize the eviction-free equilibrium, while non-coprime lengths create synchronized modes that drive instability. These results characterize when workload heterogeneity desynchronizes completions and helps stabilize memory-constrained serving. More broadly, we identify service-induced congestion as a structural instability mechanism and derive scheduling design principles for sustaining high throughput.
展开完整摘要收起摘要↓
In large language model (LLM) serving, each request accumulates persistent graphics processing unit (GPU) memory during service as its key-value cache grows with every generated token. Under high concurrency, aggregate memory usage therefore increases endogenously over time: the service process itself creates future capacity pressure. When memory capacity is exceeded, systems evict active requests, discarding cached state and restarting them later, which wastes computation and reduces throughput. We develop a discrete-time dynamical model of memory-constrained LLM inference that captures admission, memory growth, and eviction under continuous batching. In the saturated-input regime, the system admits both eviction-free fixed points and limit cycles with evictions. For homogeneous workloads, we show that the eviction-free equilibrium is unstable and that, except for a Lebesgue-measure-zero exact-capture set, the system converges to a unique worst-case limit cycle that is asymptotically stable outside this exceptional set, with throughput losses as large as 50%. For heterogeneous workloads, we prove a stability criterion in the two-class common-input setting and explain how the survival-polynomial mechanism generalizes to multiple classes and heterogeneous-input lengths. Under an input-dominated scaling regime, coprime decoding lengths stabilize the eviction-free equilibrium, while non-coprime lengths create synchronized modes that drive instability. These results characterize when workload heterogeneity desynchronizes completions and helps stabilize memory-constrained serving. More broadly, we identify service-induced congestion as a structural instability mechanism and derive scheduling design principles for sustaining high throughput.
作者Janghwan Lee, Sihwa Lee, Jinseok Kim, Yongjik Kim, Jieun Lim, Jinwook Oh, Jungwook Choi
Large Reasoning Models (LRMs) achieve strong problem-solving through long chain-of-thought, but their deployment is constrained by the high cost of full-precision inference and growing KV cache footprints. Microscaled FP4 formats enable efficient FP4 deployment; however, fully quantizing weights, activations, and KV caches (W4A4KV4) causes severe reasoning degradation that existing PTQ and QAT fail to recover. We identify that FP4 failures concentrate on low-entropy tokens--precise symbolic commitments such as digits and operators--where quantization noise inflates sampling errors that cascade through reasoning traces. Based on this insight, we propose ReQAT, a reasoning-centric FP4 training framework with three components: (i) Trace-Aligned QAT (TAQ), which revisits identical reasoning traces to focus updates on critical low-entropy decisions; (ii) Selective Entropy Minimization (SEM), which reinforces confidence at low-entropy positions; and (iii) Q-FIT, a quantization-friendly initialization that jointly calibrates RoPE-consistent KV cache transformations to stabilize QAT. Under the same training budget, ReQAT not only recovers but surpasses BF16 fine-tuning accuracy, while delivering up to 3.9x throughput speedup on NVIDIA DGX Spark and 3.1x on B200.
展开完整摘要收起摘要↓
Large Reasoning Models (LRMs) achieve strong problem-solving through long chain-of-thought, but their deployment is constrained by the high cost of full-precision inference and growing KV cache footprints. Microscaled FP4 formats enable efficient FP4 deployment; however, fully quantizing weights, activations, and KV caches (W4A4KV4) causes severe reasoning degradation that existing PTQ and QAT fail to recover. We identify that FP4 failures concentrate on low-entropy tokens--precise symbolic commitments such as digits and operators--where quantization noise inflates sampling errors that cascade through reasoning traces. Based on this insight, we propose ReQAT, a reasoning-centric FP4 training framework with three components: (i) Trace-Aligned QAT (TAQ), which revisits identical reasoning traces to focus updates on critical low-entropy decisions; (ii) Selective Entropy Minimization (SEM), which reinforces confidence at low-entropy positions; and (iii) Q-FIT, a quantization-friendly initialization that jointly calibrates RoPE-consistent KV cache transformations to stabilize QAT. Under the same training budget, ReQAT not only recovers but surpasses BF16 fine-tuning accuracy, while delivering up to 3.9x throughput speedup on NVIDIA DGX Spark and 3.1x on B200.
Many undergraduate students in Computer Science (CS) and Software Engineering (SWE) struggle to identify suitable career paths, particularly when their academic performance, abilities, and interests do not fully align. To address this issue, this study proposes an AI-driven Student Assessment and Career Prediction System that integrates a Career Guidance Expert (CGE) system with a Web-Based Student Assessment (WBSA) platform. Within the integrated framework, CGE enhances personalized career recommendations using AI while also assisting students after graduation in identifying suitable jobs, research domains, and higher study opportunities aligned with their skills and interests. The WBSA platform further strengthens interaction between students and faculty through assessments, personalized tasks, mentorship activities, and a secure real-time chat application. The CGE system employs a Multilayer Perceptron (MLP) model trained on real-world academic and extracurricular data collected using the snowball sampling method from the students of universities, achieving a validation accuracy of 94.71% in predicting personalized career paths. A pre-survey was conducted across universities to evaluate the proposed model before deployment. The WBSA system was developed as a modern web application using technologies such as Node.js, Next.js, and PostgreSQL to ensure scalability, responsiveness, and secure data management. The overall system is supported by a secure cloud-based infrastructure, the platform provides reliable performance while assisting graduates to select suitable career path in IT sector. In addition, a post-survey involving both students and faculty was conducted to gather feedback and further improve the overall effectiveness and usability of the system.
展开完整摘要收起摘要↓
Many undergraduate students in Computer Science (CS) and Software Engineering (SWE) struggle to identify suitable career paths, particularly when their academic performance, abilities, and interests do not fully align. To address this issue, this study proposes an AI-driven Student Assessment and Career Prediction System that integrates a Career Guidance Expert (CGE) system with a Web-Based Student Assessment (WBSA) platform. Within the integrated framework, CGE enhances personalized career recommendations using AI while also assisting students after graduation in identifying suitable jobs, research domains, and higher study opportunities aligned with their skills and interests. The WBSA platform further strengthens interaction between students and faculty through assessments, personalized tasks, mentorship activities, and a secure real-time chat application. The CGE system employs a Multilayer Perceptron (MLP) model trained on real-world academic and extracurricular data collected using the snowball sampling method from the students of universities, achieving a validation accuracy of 94.71% in predicting personalized career paths. A pre-survey was conducted across universities to evaluate the proposed model before deployment. The WBSA system was developed as a modern web application using technologies such as Node.js, Next.js, and PostgreSQL to ensure scalability, responsiveness, and secure data management. The overall system is supported by a secure cloud-based infrastructure, the platform provides reliable performance while assisting graduates to select suitable career path in IT sector. In addition, a post-survey involving both students and faculty was conducted to gather feedback and further improve the overall effectiveness and usability of the system.
作者Feng Qiao, Zhaochong An, Zhexiao Xiong, Serge Belongie, Nathan Jacobs
Re-rendering an existing video from a novel camera viewpoint requires the output to follow the prescribed camera trajectory while preserving the appearance and dynamics of the original scene across every frame. Existing methods rely on per-frame pose embeddings, noisy point-cloud renderings, or implicit learned correspondences, none of which provides an explicit, temporally continuous link between source and target pixels. We propose Track2View, which conditions a video diffusion transformer on paired 3D point tracks: sparse trajectories of scene points projected into both the source and target camera views. These tracks provide explicit spatiotemporal correspondences that are temporally continuous by construction, encoding what content should appear where and when. At the core of Track2View is a dual-view track conditioner that transfers visual context from source to target view through parameter-free geometric operations and learned temporal aggregation, ensuring generalization to arbitrary camera trajectories without memorizing specific motions. We further introduce a data curation pipeline that extracts one-to-one track correspondences by running a 3D point tracker on temporally concatenated multi-camera view pairs. On a 400-video benchmark spanning static and dynamic scenes, Track2View achieves state-of-the-art results across visual quality, view synchronization, and camera accuracy, reducing rotation error by 30-65% and translation error by 61-72% relative to leading baselines. Project page is available at this https URL: https://qjizhi.github.io/track2view
展开完整摘要收起摘要↓
Re-rendering an existing video from a novel camera viewpoint requires the output to follow the prescribed camera trajectory while preserving the appearance and dynamics of the original scene across every frame. Existing methods rely on per-frame pose embeddings, noisy point-cloud renderings, or implicit learned correspondences, none of which provides an explicit, temporally continuous link between source and target pixels. We propose Track2View, which conditions a video diffusion transformer on paired 3D point tracks: sparse trajectories of scene points projected into both the source and target camera views. These tracks provide explicit spatiotemporal correspondences that are temporally continuous by construction, encoding what content should appear where and when. At the core of Track2View is a dual-view track conditioner that transfers visual context from source to target view through parameter-free geometric operations and learned temporal aggregation, ensuring generalization to arbitrary camera trajectories without memorizing specific motions. We further introduce a data curation pipeline that extracts one-to-one track correspondences by running a 3D point tracker on temporally concatenated multi-camera view pairs. On a 400-video benchmark spanning static and dynamic scenes, Track2View achieves state-of-the-art results across visual quality, view synchronization, and camera accuracy, reducing rotation error by 30-65% and translation error by 61-72% relative to leading baselines. Project page is available at this https URL: https://qjizhi.github.io/track2view
We present SLS^2, a framework for safe feedback motion planning from pixels using robust model predictive control (MPC) in learned latent world models. Our approach trains an action-conditioned joint-embedding world model with compact Markovian latent states, enabling efficient gradient-based trajectory optimization through learned latent dynamics. To enforce safety for the true system despite imperfect latent predictions, we inform a GPU-accelerated system level synthesis (SLS) robust MPC scheme with conformal prediction to obtain calibrated latent error bounds and robust latent-space constraint sets. We further learn and conformalize a latent constraint checker, allowing the SLS planner to impose probabilistic safety constraints during closed-loop execution. We evaluate our method on vision-based control tasks, where it improves both goal-reaching performance and safety over latent world-model and safe-planning baselines.
展开完整摘要收起摘要↓
We present SLS^2, a framework for safe feedback motion planning from pixels using robust model predictive control (MPC) in learned latent world models. Our approach trains an action-conditioned joint-embedding world model with compact Markovian latent states, enabling efficient gradient-based trajectory optimization through learned latent dynamics. To enforce safety for the true system despite imperfect latent predictions, we inform a GPU-accelerated system level synthesis (SLS) robust MPC scheme with conformal prediction to obtain calibrated latent error bounds and robust latent-space constraint sets. We further learn and conformalize a latent constraint checker, allowing the SLS planner to impose probabilistic safety constraints during closed-loop execution. We evaluate our method on vision-based control tasks, where it improves both goal-reaching performance and safety over latent world-model and safe-planning baselines.
作者Yifei Dong, Mingen Zheng, Linquan Wu, Jeff Z. Pan, Jiaxin Bai
World-model synthesis aims to turn interaction experience into an internal model of environment dynamics. Existing symbolic approaches often fit observed transitions or mixtures of local rules, but they do not produce a complete executable program that can run independently of the real environment. We present Mind-Studio, a framework that synthesizes executable pygame-style world models from state-action-next-state trajectories using large language models. Mind-Studio combines entropy-selected traces with a lightweight game skill file containing object, action, and static scene information extracted from screenshots. We evaluate synthesis quality with a K-step lookahead fidelity protocol that compares generated world-model rollouts against Real-ALE rollouts from the same state. On Montezuma's Revenge, Mind-Studio improves chosen-action next-state prediction from 0.3% for PoE-World to 48.7% while verifying 5 of 8 subgoals; across Alien, Assault, and Skiing, it achieves stronger branch-level fidelity than prior learned lookahead sources.
展开完整摘要收起摘要↓
World-model synthesis aims to turn interaction experience into an internal model of environment dynamics. Existing symbolic approaches often fit observed transitions or mixtures of local rules, but they do not produce a complete executable program that can run independently of the real environment. We present Mind-Studio, a framework that synthesizes executable pygame-style world models from state-action-next-state trajectories using large language models. Mind-Studio combines entropy-selected traces with a lightweight game skill file containing object, action, and static scene information extracted from screenshots. We evaluate synthesis quality with a K-step lookahead fidelity protocol that compares generated world-model rollouts against Real-ALE rollouts from the same state. On Montezuma's Revenge, Mind-Studio improves chosen-action next-state prediction from 0.3% for PoE-World to 48.7% while verifying 5 of 8 subgoals; across Alien, Assault, and Skiing, it achieves stronger branch-level fidelity than prior learned lookahead sources.
作者Ruofei Bai, Jie Chen, Yuxin Cai, Jun Li, Wei-Yun Yau, Lihua Xie
The success of generative models in language and visual generation has inspired extensive applications to generative robot planning. However, most existing works either focus on single-robot planning, or generate multi-robot trajectories in a sequential manner with iterative post-processing to resolve inter-robot conflicts. In this work, we investigate whether coordinated multi-robot trajectories, as a special spatiotemporal distribution, can be learned and generated with a generative model in a feed-forward manner. We propose Robots as Tokens (Roken), a unified diffusion transformer that directly generates multi-robot trajectories that satisfy both (individual) safety and (global) connectivity constraints. The core design of Roken is to represent each robot as a discrete token, allowing them to naturally interact with each other through self-attention, and cross-attend to map tokens for environment layouts. We further introduce several auxiliary tasks based on Bayes' theorem to provide multi-scale spatial-temporal supervision for efficient learning of the conditional distribution. In training, Roken absorbs diverse expert trajectories from different team sizes. During inference, Roken behaves as a versatile multi-robot planner that can handle single-robot planning, coordinated multi-robot trajectory generation, and conditional trajectory generation by fixing some robot tokens as conditions. Experiments in diverse cluttered environments show that Roken can generate coordinated multi-robot trajectories to perform connectivity-constrained goal navigation tasks with high success rates, outperforming the baseline method used to generate the training dataset. Roken also demonstrates good scalability after training with mixed team sizes, and shows generalization to unseen or partially observed environments, verifying its potential to learn from diverse data and perform versatile tasks.
展开完整摘要收起摘要↓
The success of generative models in language and visual generation has inspired extensive applications to generative robot planning. However, most existing works either focus on single-robot planning, or generate multi-robot trajectories in a sequential manner with iterative post-processing to resolve inter-robot conflicts. In this work, we investigate whether coordinated multi-robot trajectories, as a special spatiotemporal distribution, can be learned and generated with a generative model in a feed-forward manner. We propose Robots as Tokens (Roken), a unified diffusion transformer that directly generates multi-robot trajectories that satisfy both (individual) safety and (global) connectivity constraints. The core design of Roken is to represent each robot as a discrete token, allowing them to naturally interact with each other through self-attention, and cross-attend to map tokens for environment layouts. We further introduce several auxiliary tasks based on Bayes' theorem to provide multi-scale spatial-temporal supervision for efficient learning of the conditional distribution. In training, Roken absorbs diverse expert trajectories from different team sizes. During inference, Roken behaves as a versatile multi-robot planner that can handle single-robot planning, coordinated multi-robot trajectory generation, and conditional trajectory generation by fixing some robot tokens as conditions. Experiments in diverse cluttered environments show that Roken can generate coordinated multi-robot trajectories to perform connectivity-constrained goal navigation tasks with high success rates, outperforming the baseline method used to generate the training dataset. Roken also demonstrates good scalability after training with mixed team sizes, and shows generalization to unseen or partially observed environments, verifying its potential to learn from diverse data and perform versatile tasks.
作者Kacper Wyrwal, İsmail İlkan Ceylan, Alexander Tong
Flow matching is a powerful generative modeling framework, valued for its simplicity and strong empirical performance. However, its standard formulation treats signals on structured spaces, such as fMRI data on brain graphs, as points in Euclidean space, overlooking the rich topological features of their domains. To address this, we introduce topological flow matching, a topology-aware generalization of flow matching. We interpret flow matching as a framework for solving a degenerate Schrödinger bridge problem and inject topological information by augmenting the reference process with a Laplacian-derived drift. This principled modification captures the structure of the underlying domain while preserving the desirable properties of flow matching: a stable, simulation-free objective and deterministic sample paths. As a result, our framework serves as a drop-in replacement for standard flow matching. We demonstrate its effectiveness on diverse structured datasets, including brain fMRIs, ocean currents, seismic events, and traffic flows.
展开完整摘要收起摘要↓
Flow matching is a powerful generative modeling framework, valued for its simplicity and strong empirical performance. However, its standard formulation treats signals on structured spaces, such as fMRI data on brain graphs, as points in Euclidean space, overlooking the rich topological features of their domains. To address this, we introduce topological flow matching, a topology-aware generalization of flow matching. We interpret flow matching as a framework for solving a degenerate Schrödinger bridge problem and inject topological information by augmenting the reference process with a Laplacian-derived drift. This principled modification captures the structure of the underlying domain while preserving the desirable properties of flow matching: a stable, simulation-free objective and deterministic sample paths. As a result, our framework serves as a drop-in replacement for standard flow matching. We demonstrate its effectiveness on diverse structured datasets, including brain fMRIs, ocean currents, seismic events, and traffic flows.
Mechanistic interpretability seeks to explain neural network behavior by decomposing model computations into interpretable features and circuits. While transcoder-based circuit tracing has recently enabled detailed causal analyses of large language models, multimodal diffusion transformers for image generation remain comparatively opaque. We still lack tools for understanding how semantic information propagates across denoising steps and how text and image representations interact within double-stream MM-DiT architectures. Existing methods provide only partial insight: attention maps expose a limited view of token interactions, while sparse autoencoders can discover interpretable features but do not directly reveal how these features are transformed and composed through nonlinear MLP layers. In this work, we extend transcoder-based circuit tracing to multimodal diffusion transformers. We train timestep-conditioned transcoders that faithfully approximate the input-output behavior of MLP sublayers in FLUX.1[schnell]. By replacing MLPs with transcoders and linearizing the remaining computation, we obtain exact feature-to-feature attribution and recover compact, interpretable circuits. Empirically, our transcoders match or slightly outperform sparse autoencoders on the sparsity-faithfulness tradeoff. The resulting circuits reveal mechanisms underlying attribute binding and cross-stream semantic propagation, and provide causal explanations for systematic generation errors. Moreover, circuit-guided interventions are substantially more precise and effective than standard SAE-based steering. Our results demonstrate that transcoder-based circuit analysis is feasible for state-of-the-art diffusion transformers and provides a powerful framework for understanding and controlling multimodal generative models. The code is available at https://github.com/Artalmaz31/DifFRACT
展开完整摘要收起摘要↓
Mechanistic interpretability seeks to explain neural network behavior by decomposing model computations into interpretable features and circuits. While transcoder-based circuit tracing has recently enabled detailed causal analyses of large language models, multimodal diffusion transformers for image generation remain comparatively opaque. We still lack tools for understanding how semantic information propagates across denoising steps and how text and image representations interact within double-stream MM-DiT architectures. Existing methods provide only partial insight: attention maps expose a limited view of token interactions, while sparse autoencoders can discover interpretable features but do not directly reveal how these features are transformed and composed through nonlinear MLP layers. In this work, we extend transcoder-based circuit tracing to multimodal diffusion transformers. We train timestep-conditioned transcoders that faithfully approximate the input-output behavior of MLP sublayers in FLUX.1[schnell]. By replacing MLPs with transcoders and linearizing the remaining computation, we obtain exact feature-to-feature attribution and recover compact, interpretable circuits. Empirically, our transcoders match or slightly outperform sparse autoencoders on the sparsity-faithfulness tradeoff. The resulting circuits reveal mechanisms underlying attribute binding and cross-stream semantic propagation, and provide causal explanations for systematic generation errors. Moreover, circuit-guided interventions are substantially more precise and effective than standard SAE-based steering. Our results demonstrate that transcoder-based circuit analysis is feasible for state-of-the-art diffusion transformers and provides a powerful framework for understanding and controlling multimodal generative models. The code is available at https://github.com/Artalmaz31/DifFRACT