Benchmarks are often used as a standard to understand LLM capabilities in different domains. However, aggregate benchmark scores provide limited insight into compositional skill gaps of LLMs and how to improve them. To make these weaknesses visible, we propose Scaffolded Task Design (STaD) framework. STaD generates controlled variations of benchmark tasks based on the concept of scaffolding, which introduces structured, incremental support in a step-by-step manner. Rather than inspecting failures individually, this approach enables systematic and scalable probing of model behavior by identifying the specific reasoning skill compositions they lack. Treating the LLM as a black box, our experiments on six models of varying sizes reveal multiple failure points in three reasoning benchmarks and highlight each model’s unique and distinct skill gaps.
论文检索
输入标题、作者或关键词，从 100,903 篇学术成果中精准定位
Evaluating software engineering capabilities has become a core component of modern large language models (LLMs); however, the key bottleneck hindering further scaling lies not in the scarcity of high-quality solutions, but in the lack of high-quality test suites. Test suites are indispensable both for synthesizing program repair trajectories and for providing precise feedback signals in reinforcement learning. Unfortunately, due to the high cost and difficulty of annotation, high-quality test suites have long been hard to obtain, while those automatically generated by LLMs tend to be superficial and lack sufficient discriminative power. As a first step toward constructing high-quality test suites, we introduce SWE-Mutation, a benchmark for evaluating LLM-generated test suites. The benchmark characterizes test suites by introducing systematically mutated solutions that attempt to “fool” the test suites and pass validation. We further propose an agentic, language-agnostic framework for automatically generating complex mutants. Our benchmark consists of 2,636 mutated variants derived from 800 original instances and includes a multilingual subset spanning nine programming languages. Experiments on seven LLMs reveal that even DeepSeek-V3.1 achieves only 10.20% verification and 36.15% detection rates, highlighting the inadequacy of current LLMs. Additionally, our agentic mutation strategy enhances realism, reducing average detection rates from 71.04% to 39.81% compared to conventional methods. These findings expose persistent deficiencies in the ability of current LLMs to generate reliable and discriminative test suites.
Nash-Pruned CredMAS: Dynamic Panel Pruning for VLM-MAS using Nash-based Selection and Doubly-Robust Credits
PDF ↗Multi-round Vision-Language Model (VLM) Multi-Agent Systems (MAS) offer powerful reasoning capabilities but suffer from prohibitive costs due to static panel designs, where all N agents communicate at every T round. This approach is fundamentally inefficient, as it ignores the context-dependent and diminishing marginal utility of specific agents. To address this, we propose Nash-CredMAS, an economic framework that transforms agent selection into a dynamic resource allocation game. Unlike heuristic routing or one-time pruning, our method operates in two phases: (1) Offline Causal Value Learning, where we employ a doubly-robust (AIPW) estimator to train a context-aware value function from biased interaction logs, effectively learning the true marginal contribution of agents; and (2) Online Dynamic Auctions, where agents bid for communication slots based on their predicted utility. We formulate the inference-time selection as a submodular maximization problem under budget constraints, theoretically guaranteeing a (1 - 1/e)-approximation of the optimal coalition via a greedy strategy. Empirically, Nash-CredMAS achieves state-of-the-art results on challenging benchmarks, including MMMU and V*-Bench, while reducing token consumption by over 25% compared to static baselines. The system naturally converges to an economic equilibrium where agents actively remain silent when their marginal value does not justify the cost.
As large language model agents tackle increasingly complex long-horizon tasks, effective post-training becomes critical. Prior work faces fundamental challenges: outcome-only rewards fail to precisely attribute credit to intermediate steps, estimated step-level rewards introduce systematic noise, and Monte Carlo sampling approaches for step reward estimation incur prohibitive computational cost. Inspired by findings that only a small fraction of high-entropy tokens drive effective RL for reasoning, we propose Critical Step Optimization (CSO), which focuses preference learning on verified critical steps—decision points where alternate actions demonstrably flip task outcomes from failure to success. Crucially, our method starts from failed policy trajectories rather than expert demonstrations, directly targeting the policy model’s weaknesses. We use a process reward model (PRM) to identify candidate critical steps, leverage expert models to propose high-quality alternatives, then continue execution from these alternatives using the policy model itself until task completion. Only alternatives that the policy successfully executes to correct outcomes are verified and used as DPO training data, ensuring both quality and policy reachability. This yields fine-grained, verifiable supervision at critical decisions while avoiding trajectory-level coarseness and step-level noise. Experiments on GAIA-Text-103 and XBench-DeepSearch show that CSO achieves 37% and 26% relative improvement over the SFT baseline and substantially outperforms other post-training methods, while requiring supervision at only 16% of trajectory steps. This demonstrates the effectiveness of selective verification-based learning for agent post-training.
Language-model-based agents operating over extended interaction horizons face persistent challenges in preserving temporally grounded information and maintaining behavioral consistency across sessions, a failure mode we term "soul erosion." We present BMAM (Brain-inspired Multi-Agent Memory), a general-purpose memory architecture that models agent memory as a set of functionally specialized subsystems rather than a single unstructured store. Inspired by cognitive memory systems, BMAM decomposes memory into episodic, semantic, salience-aware, and control-oriented components that operate at complementary time scales, organised as a six-phase memory lifecycle. To support long-horizon reasoning, BMAM organises episodic memories along explicit timelines and retrieves evidence by fusing multiple complementary signals. Experiments on the LoCoMo benchmark show that BMAM achieves 78.45% accuracy, outperforming seven memory-augmented baselines. Pairwise ablations reveal super-additive synergy between brain-region components rather than redundant stacking, and a Soul Portability Test demonstrates 87.5% identity-integrity across full memory export, clear, and restore. A targeted refinement of the temporal-trigger heuristics raises LongMemEval multi-session accuracy from 45.2% to 56.4%, validating the architectural decomposition behind BMAM.Code is available at https://github.com/innovation64/BMAM.
As numerous instruction-tuning datasets continue to emerge, dynamically balancing and optimizing their mixtures has become a criticalchallenge. To address this, we propose DynamixSFT, a dynamic and automated method for instruction-tuning dataset mixture optimization. We formulate the problem as a multi-armed bandit setup and introduce a Prior-scaled Boltzmann Exploration that softly anchors the updated sampling distribution to the original dataset proportions, thereby preserving the inherent diversity and coverage of the collection. Sampling probabilities are updated using a lightweight 1-Step Look-ahead Reward, reflecting how much the dataset contributes to improving the model’s performance at its current state. We demonstrate that DynamixSFT effectively optimizes the TÜLU-2-mixture andTÜLU-3-mixture collections across 10 benchmarks, while introducing minimal computational overhead over naive sampling. Furthermore, we provide a comprehensive analysis and visualizations to offer deeper insights into the adaptive dynamics of our method.
The increasing integration of large language models (LLMs) into mental health applications necessitates robust frameworks for evaluating professional safety alignment. Current evaluative approaches primarily rely on refusal-based safety signals, which offer limited insight into the nuanced behaviors required in clinical practice. In mental health, clinically inadequate refusals can be perceived as unempathetic and discourage help-seeking. To address this gap, we move beyond refusal-centric metrics and introduce PsychEthicsBench, the first principle-grounded benchmark based on Australian psychology and psychiatry guidelines, designed to evaluate LLMs’ ethical knowledge and behavioral responses through multiple-choice and open-ended tasks with fine-grained ethicality annotations. Empirical results across 14 models reveal that refusal rates are poor indicators of ethical behavior, revealing a significant divergence between safety triggers and clinical appropriateness. Notably, we find that domain-specific fine-tuning can degrade ethical robustness, as several specialized models underperform their base backbones in ethical alignment. PsychEthicsBench provides a foundation for systematic, jurisdiction-aware evaluation of LLMs in mental health, encouraging more responsible development in this domain.
PhaseMI: A Motivational Interviewing Dataset for Enhancing Phase Progression in LLM-based Counseling
PDF ↗The growing demand for scalable mental health support has increased interest in AI-based counseling systems grounded in Motivational Interviewing (MI). However, existing MI datasets do not explicitly model the structured progression of MI phases, which is essential for effective and goal-oriented counseling. To address this gap, we introduce PhaseMI, a phase-structured MI dataset, together with a data generation framework that employs therapist, client, and supervisor LLMs to explicitly control phase transitions. Compared to the best alternative baseline, PhaseMI achieves improved coverage of MI phases, with gains of 12.3% in exploring, 37.6% in guiding, and 61.1% in choosing, and experimental evaluations demonstrate that it yields higher overall counseling quality than baseline datasets.
POLARIS: A Gödel Agent Framework for Small Language Models through Experience-Abstracted Policy Repair
PDF ↗Gödel agent realize recursive self-improvement: an agent inspects its own policy and traces and then modifies that policy in a tested loop. We introduce Polaris, a Gödel agent for compact models that performs policy repair via experience abstraction, turning failures into policy updates through a structured cycle of analysis, strategy formation, abstraction, and minimal code patch repair with conservative checks. Unlike response level self-correction or parameter tuning, Polaris makes policy level changes with small, auditable patches that persist in the policy and are reused on unseen instances within each benchmark. As part of the loop, the agent engages in meta reasoning: it explains its errors, proposes concrete revisions to its own policy, and then updates the policy. To enable cumulative policy refinement, we introduce experience abstraction, which distills failures into compact, reusable strategies that transfer to unseen instances. On MGSM, DROP, GPQA, and LitBench (covering arithmetic reasoning, compositional inference, graduate-level problem solving, and creative writing evaluation), a 7-billion-parameter model equipped with Polaris achieves consistent gains over the base policy and competitive baselines.
Preference alignment methods such as RLHF and Direct Preference Optimization (DPO) improve instruction following, but they can also reinforce hallucinations when preference judgments reward fluency and confidence over factual correctness. We introduce F-DPO (Factuality-aware Direct Preference Optimization), a simple extension of DPO that uses only binary factuality labels. F-DPO (i) applies a label-flipping transformation that corrects misordered preference pairs so the chosen response is never less factual than the rejected one, and (ii) adds a factuality-aware margin that emphasizes pairs with clear correctness differences, while reducing to standard DPO when both responses share the same factuality. We construct factuality-aware preference data by augmenting DPO pairs with binary factuality indicators and synthetic hallucinated variants. Across seven open-weight LLMs (1B–14B), F-DPO consistently improves factuality and reduces hallucination rates relative to both base models and standard DPO. On Qwen3-8B, F-DPO reduces hallucination rates by 5×(from 0.424 to 0.084) while improving factuality scores by 50% (from 5.26 to 7.90). F-DPO also generalizes to out-of-distribution benchmarks: on TruthfulQA, Qwen2.5-14B achieves +17% MC1 accuracy (0.500 to 0.585) and +49% MC2 accuracy (0.357 to 0.531). F-DPO requires no auxiliary reward model, token-level annotations, or multi-stage training.
The Model Context Protocol (MCP) aims to standardize the integration of Large Language Models (LLMs) with external tools, yet existing research primarily evaluates functional capabilities while treating the underlying protocol as an opaque black box. This oversight obscures critical inefficiencies in token flows and latency distributed across MCP’s decoupled Host-Client-Server architecture. In this paper, we introduce ProMCP, an end-to-end profiling and instrumentation framework that decomposes the MCP workflow into a six-stage communication pipeline, enabling granular attribution of computational costs. We evaluate widely varying deployment topologies—from air-gapped local models to commercial off-the-shelf (OTS) clients—across 20 servers and 169 tools from MCP-Bench and MCP-Universe. Our analysis reveals a distinct inversion in performance bottlenecks: topologies with customized clients devote 56–72% of total tokens and 60–67% of latency to planning and schema injection, whereas OTS clients concentrate over 85% of latency in final answer synthesis. Crucially, actual tool execution constitutes a negligible fraction of the total cost across all configurations. These findings establish a quantitative baseline for protocol overhead and demonstrate that future optimization must target schema orchestration and transport efficiency rather than tool execution speed. The code is available at: https://github.com/ResponsibleAILab/ProMCP.
Characterizing the Robustness of Black-Box LLM Planners Under Perturbed Observations with Adaptive Stress Testing
PDF ↗Large language models (LLMs) have recently demonstrated success in decision-making tasks including planning, control, and prediction, but their tendency to hallucinate unsafe and undesired outputs poses risks. This unwanted behavior is further exacerbated in environments where sensors are noisy or unreliable. Characterizing the behavior of LLM planners to varied observations is necessary to proactively avoid failures in safety-critical scenarios. We specifically investigate the response of LLMs along two different perturbation dimensions. Like prior works, one dimension generates semantically similar prompts with varied phrasing by randomizing order of details, modifying access to few-shot examples, etc. Unique to our work, the second dimension simulates access to varied sensors and noise to mimic raw sensor or detection algorithm failures. An initial case study in which perturbations are manually applied show that both dimensions lead LLMs to hallucinate in a multi-agent driving environment. However, manually covering the entire perturbation space for several scenarios is infeasible. As such, we propose a novel method for efficiently searching the space of prompt perturbations using adaptive stress testing (AST) with Monte-Carlo tree search (MCTS). Our AST formulation enables discovery of scenarios, sensor configurations, and prompt phrasing that cause language models to act with high uncertainty or even crash. By generating MCTS prompt perturbation trees across diverse scenarios, we show through extensive experiments that offline analyses can be used to proactively understand potential failures that may arise at runtime. Code is available at https://sites.google.com/illinois.edu/astllm/.
Chain-of-Thought (CoT) has significantly enhanced LLM reasoning, yet often incurs substantial computational overhead due to “overthinking”: generating excessively long rationales without commensurate accuracy gains. Existing efficiency methods typically apply uniform compression, which overlooks a critical observation that reasoning complexity is heterogeneous at two distinct granularities: across different problems and within individual reasoning steps. This motivates our principle of Thinking Economically: intelligently allocating computational resources based on intrinsic task and step demands rather than pursuing uniform brevity. We propose Hierarchical Adaptive Budgeter (HAB), a training framework that operationalizes this principle through coarse-to-fine budgeting. At the inter-step level, HAB predicts the optimal reasoning depth for each problem. At the intra-step level, HAB learns step-specific token budgeting signals from PPL-derived step comparisons and an adaptive Pareto optimization objective that captures the local quality-efficiency trade-off, while a Fisher Information-based pruner further provides fine-grained training-time guidance, thereby encouraging the generator to internalize more economical reasoning patterns. Experiments on GSM8K and MATH500 show that HAB not only surpasses standard CoT in accuracy but also reduces token usage, achieving a stronger performance-efficiency trade-off than the compared baselines.
Despite displaying semantic competence, large language models’ internal mechanisms that ground abstract semantic structure remain insufficiently characterised. To investigate whether and how LLMs develop causally functional representations of semantic roles, we introduce a causal-temporal methodology combining contrastive minimal pairs, edge-attribution circuit discovery, and training-time tracking. Our analysis reveals that LLMs encode semantic roles through highly localised circuits (89–92% attribution within 28 nodes) that emerge gradually via structural refinement rather than phase transitions. These circuits exhibit moderate cross-scale conservation (24–51% component overlap) alongside high spectral similarity, with larger models reusing similar components while rewiring connections. These findings suggest that LLMs form compact, causally isolated mechanisms for abstract semantic structure that exhibit partial transfer across scales and architectures.
Although Large Language Models (LLMs) have made remarkable progress, current preference optimization methods still struggle to align directional consistency while preserving reasoning diversity. To address this limitation, we propose Directional-Groupwise Preference Optimization (DGPO), a lightweight framework that aggregates supervision signals at the group level and explicitly models direction-aware alignment through multi-candidate comparisons. DGPO organizes forward and reverse question-answer instances into structured sets and optimizes a margin-based likelihood objective that separates coherent reasoning paths from inconsistent alternatives. This groupwise formulation captures richer relative information than pairwise objectives and reinforces consistency across diverse reasoning pathways. Empirical results show that our constructed reverse data yields a 3.2% average improvement across five benchmarks, while DGPO further delivers consistent gains across multiple datasets and model families, achieving average accuracy improvements of up to 3.6%. Our code and data are available at https://github.com/Demi-deng2/DGPO.
Automated Program Repair (APR) is vital for software maintenance. Despite notable advancements, existing methods still face challenges of insufficient bug dependency modeling and inadequate global repair planning when addressing semantically complex multi-location bugs. We propose CascadeFix, a multi-location automatic repair method via cascading planning and generation. Firstly, to improve the modeling of semantic and structural dependencies among bugs, three types of bug relationships-Use, Copy, and Nearby-are defined to characterize semantic connection, patch reusability, and contextual interference. Then, to address inadequate global repair planning, a cascading repair planning algorithm is designed to effectively cluster strongly correlated bugs and intelligently assign reasonable repair priorities and operations to each cluster, ensuring the rationality and consistency of global repair. Finally, taking clusters as the basic repair units, a cascading patch generation mechanism is proposed to dynamically integrate intra-cluster dependency information and cross-cluster repair knowledge, producing patches that maintain syntactic correctness and semantic consistency under global dependency constraints. Experiments on Defects4J show that CascadeFix resolves 84 multi-location bugs, achieving a 31% improvement over current state-of-the-art methods.
Long Chain-of-Thought (CoT) reasoning is pivotal for the success of recent reasoning models but suffers from high computational overhead and latency. While prior works attempt to compress CoT via external compressor, they often fail to align with the model’s internal reasoning dynamics, resulting in the loss of critical logical steps. This paper presents Compressing Redundancy in Chain-of-Thought via Intrinsic Saliency Pruning (CRISP), a framework that compresses CoT by exploiting the model’s intrinsic saliency. Our analysis reveals a distinct phenomenon: the reasoning termination token </think> acts as an information anchor, where its attention pattern effectively demarcates essential reasoning from redundancy. Based on this finding, we design a policy that utilizes these intrinsic attention signals to guide atomic compression operations. In contrast to coarse-grained pruning strategies, CRISP strategically distills the reasoning chain to maximize information density while preserving logical coherence. Empirical results across various backbone models and mathematical datasets demonstrate that CRISP achieves a 50-60% reduction in token count without compromising accuracy, effectively mitigating the efficiency bottleneck of long-context reasoning. We open-source our implementation to facilitate further research in efficient reasoning.
As large language model (LLM) based systems take on high-stakes roles in real-world decision-making, they must reconcile competing instructions from multiple sources within a single prompt context. Enforcing an instruction hierarchy, where higher-level directives override lower-priority requests, is critical to the reliability and controllability of LLMs. In this work, we reframe instruction hierarchy resolution as a reasoning task. The model must first "think" about the relationship between a given user prompt and higher-priority instructions before generating a response. To enable this capability, we construct VerIH, a training dataset of constraint-following tasks with verifiable answers, comprising aligned and conflicting system–user instructions. We show that lightweight reinforcement learning with VerIH effectively transfers general reasoning capabilities of models to instruction prioritization. Our method leads to consistent improvements across multiple model families and scales on instruction following and instruction hierarchy benchmarks, achieving ~20% absolute improvement in conflict setups. Our method also generalizes to safety-critical scenarios beyond the training distribution, exhibiting increased robustness against jailbreak and prompt injection, reducing absolute attack success rates by up to 20%. Our results establish reasoning over instruction hierarchies as a practical mechanism for improving AI reliability, where targeted updates to system prompts produce predictable, controllable, and robust changes in model behavior.
Large Language Models have recently shown impressive capabilities in reasoning and code generation, making them promising tools for natural language interfaces to relational databases. However, existing approaches often fail to generalize in complex, real-world settings due to the highly database-specific nature of SQL reasoning, which requires deep familiarity with unique schemas, ambiguous semantics, and intricate join paths. To address this challenge, we introduce a novel two-stage LLM-based framework that decouples knowledge acquisition from query generation. In the Exploration Stage, the system autonomously constructs a database-specific knowledge base by navigating the schema with a Monte Carlo Tree Search–inspired strategy, generating triplets of schema fragments, executable queries, and natural language descriptions as usage examples. In the Deployment Stage, a dual-agent system leverages the collected knowledge as in-context examples to iteratively retrieve relevant information and generate accurate SQL queries in response to user questions. This design enables the agent to proactively familiarize itself with unseen databases and handle complex, multi-step reasoning. Extensive experiments on large-scale benchmarks demonstrate that our approach significantly improves accuracy over strong baselines, highlighting its effectiveness and generalizability.
Temporal Entity Alignment (TEA), which aims to identify equivalent entities across Temporal Knowledge Graphs (TKGs), is crucial for integrating knowledge facts from multiple sources. However, existing TEA models often fail to capture the orthogonal yet complementary effect between structural and temporal features, and typically overlook the importance of information richness—a key factor for effective message passing in the neural feature encoders. To address these limitations, we propose a RCTEA framework that jointly models both structural and temporal aspects of the TKGs for entity alignment. Specifically, we design a richness-guided attention mechanism along with an adaptive weighting strategy to facilitate effective feature fusion. To ensure robust alignment despite noisy entity contexts, we introduce a dual-view neighborhood consensus algorithm that jointly refines the feature encoders to enforce local structural consistency of the predicted alignments. Extensive experiments demonstrate the superiority of RCTEA, achieving state-of-the-art performance on public TEA benchmarks.