The slow thinking paradigm has been widely validated to enhance the reasoning capabilities of Large Language Models (LLMs), but it introduces notable reasoning inefficiencies: models often overthink simple tasks while prematurely shifting their reasoning paths when addressing complex problems. To address this, we propose AdapThink, a simple yet efficient framework for adaptive reasoning preference control. Unlike methods imposing uniform length constraints, AdapThink dynamically adjusts reflection preferences based on group-level distributional statistics of reasoning length and reflection intensity. We further introduce a dispersion-based diversity sampling mechanism that maximizes the geometric spread of reasoning patterns, accelerating learning through exposure to diverse problem-solving strategies. Across mathematical reasoning and code generation benchmarks, AdapThink reduces average response length by 17.1%-21.4% while improving performance by 6.12-6.59 points under 32K token budgets, demonstrating superior efficiency and robustness against reward hacking compared to strong baselines.
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
This paper presents ltzGLUE, the first Natural Language Understanding (NLU) benchmark for Luxembourgish (LTZ) based on the popular GLUE benchmark for English. Although NLU tasks are available for many european languages nowadays, LTZ is one of the official national languages that is often overlooked. We introduce new tasks and reuse existing ones to introduce the first official NLU benchmark and accompanying evaluation of encoder models for the language. Our tasks include common natural language processing tasks in binary and multi-class classification settings, including named entity recognition, topic classification, and intent classification. We evaluate various pre-trained language models for LTZ to present an overview of the current capabilities of these models on the LTZ language.
GRASS: Gradient-based Adaptive Layer-wise Importance Sampling for Memory-efficient Large Language Model Fine-tuning
PDF ↗Full-parameter fine-tuning of large language models is constrained by substantial GPU memory demands. Low-rank adaptation methods mitigate this challenge by updating only a subset of parameters. However, these approaches often limit model expressiveness and yield lower performance than full-parameter fine-tuning. Layer-wise fine-tuning methods have emerged as an alternative, enabling memory-efficient training through static layer importance sampling strategies. However, these methods overlook variations in layer importance across tasks and training stages, resulting in suboptimal performance on downstream tasks. To address these limitations, we propose GRASS, a gradient-based adaptive layer-wise importance sampling framework. GRASS utilizes mean gradient norms as a task-aware and training-stage-aware metric for estimating layer importance. Furthermore, GRASS adaptively adjusts layer sampling probabilities through an adaptive training strategy. We also introduce a layer-wise optimizer state offloading mechanism to further reduce memory usage while maintaining comparable training throughput. Extensive experiments across multiple models and benchmarks demonstrate that GRASS consistently outperforms state-of-the-art methods, achieving an average accuracy improvement of up to 4.38 points and reducing memory usage by up to 19.97%.
Large language models (LLMs) exhibit strong reasoning when guided by chain-of-thought exemplars, yet collecting large, high-quality reasoning datasets remains laborious and resource-intensive. We introduce Prompting Test-Time Scaling (P-TTS), a prompt-space data augmentation framework for enhancing LLM reasoning via fine-tuning. In P-TTS, scaling refers to systematic expansion of the prompt space during offline teacher-data generation, not to increased inference-time compute for the deployed student. Rather than collecting thousands of examples, P-TTS starts from a small pool of 90 manually selected reasoning instances and applies principled instruction templates and paraphrased prompt variants to elicit diverse reasoning trajectories from a teacher model, producing a compact synthetic training set. We fine-tune Qwen-2.5 models of multiple sizes on the resulting data. On reasoning benchmarks including AIME25, MATH500, and GPQA-Diamond, P-TTS consistently improves accuracy over competitive small-data baselines such as S1 and S1.1 (1K-shot), with the largest gains on AIME25 while remaining strong on MATH500 and GPQA-Diamond. P-TTS also improves generalization on out-of-domain reasoning evaluations. Ablations show that exemplar diversity and prompt-space scaling are critical drivers of improvement, suggesting that prompt scaling explores the latent space of reasoning patterns, amplifying LLM problem-solving with minimal annotation overhead. P-TTS offers a practical, low-cost way to elicit strong LLM reasoning in resource-constrained or rapidly evolving domains. Our code and data are available at https://github.com/VILA-Lab/PTTS.
No-Worse Context-Aware Decoding: Preventing Neutral Regression in Context-Conditioned Generation
PDF ↗Large language models (LLMs) can answer questions and summarize documents when conditioned on external contexts (e.g., retrieved evidence), yet context use remains unreliable: models may overwrite an already-correct output (neutral regression) even when the context is non-informative. We formalize neutral regression as a do-no-harm requirement and quantify it by measuring accuracy drops on baseline-correct items under answer-consistent contexts. We propose No-Worse Context-Aware Decoding (NWCAD), a decode-time adapter built on a two-stream setup with a two-stage gate: it backs off to no-context decoding when the context is non-informative, and otherwise uses context-conditioned decoding with a contrastive fallback under uncertainty. We evaluate NWCAD on benchmarks that separate do-no-harm reliability from context utilization (accuracy gains on genuinely helpful contexts). NWCAD prevents neutral regression on baseline-correct items while preserving strong context-driven accuracy on helpful contexts.
Traditional automatic evaluation methods have been shown to be unsuitable for modern Chinese poetry because of the distinct nature of this literary genre. Human evaluation remains reliable, but is expensive and not applicable to large-scale data. In this paper, we propose Poller (Poetry LLM Evaluator), a novel method leveraging Large Language Models (LLMs) to evaluate the poetry understanding task. Specifically, our method requires LLMs to play the role of a poem’s author with detailed information, thereby emulating human evaluation and judgment by adopting the poet’s perspective. We conducted comprehensive experiments on multiple LLMs, evaluating the interpretations of poems across eight specialized dimensions. Experimental results demonstrate that our method effectively reduces the evaluation error between LLMs and humans. Especially for specific dimension evaluation, Poller-based LLMs achieve a 94.55% and 89.53% error reduction for rhetorical techniques and defamiliarization, respectively, compared to baseline methods. These performances are unattainable by conventional LLM evaluation methods. Experimental results from multiple LLMs across various dimensions validate the efficacy of our method. This work bridges the gap between automated efficiency and human expertise, establishing a foundation for automated evaluation in poetry-related tasks.
PersonaArena: Dynamic Simulation for Evaluating and Enhancing Persona-Level Role-Playing in Large Language Models
PDF ↗Large language models (LLMs) increasingly serve as interactive social agents, yet their ability to maintain coherent and authentic persona-level role-playing remains limited, particularly in realistic social scenarios. Existing research predominantly focuses on character-level settings and relies on static evaluation formats, failing to capture the complexity of everyday social interactions. In this work, we present PersonaArena, a dynamic simulation framework for evaluating and improving persona-level role-playing in LLMs. PersonaArena leverages a large, filtered corpus of user-generated social content to construct a nuanced persona bank, and elicits multi-turn, context-rich interactions within simulated social environments. Our framework features a multi-agent debating judge for holistic and unbiased assessment. Through extensive experiments, we demonstrate that PersonaArena enables rigorous evaluation and enhancement of LLMs’ role-playing capabilities, advancing the development of more authentic and socially adept AI agents. Our codes and long appendix are available at https://anonymous.4open.science/r/PersonaArena-B323/.
Large Language Models (LLMs) augmented with external tools have demonstrated remarkable capabilities in complex reasoning tasks. However, existing frameworks rely heavily on natural language reasoning to determine when tools can be invoked and whether their results should be committed, lacking formal guarantees for logical safety and verifiability. We present ToolGate, a forward execution framework that provides logical safety guarantees and verifiable state evolution for LLM tool calling. ToolGate maintains an explicit symbolic state space as a typed key-value mapping representing trusted world information throughout the reasoning process. Each tool is formalized as a Hoare-style contract consisting of a precondition and a postcondition, where the precondition gates tool invocation by checking whether the current state satisfies the required conditions, and the postcondition determines whether the tool’s result can be committed to update the state through runtime verification. Our approach guarantees that the symbolic state evolves only through verified tool executions, preventing invalid or hallucinated results from corrupting the world representation. Experimental validation demonstrates that ToolGate significantly improves the reliability and verifiability of tool-augmented LLM systems while maintaining competitive performance on complex multi-step reasoning tasks. This work establishes a foundation for building more trustworthy and debuggable AI systems that integrate language models with external tools.
Disambiguation-Centric Finetuning Makes Enterprise Tool-Calling LLMs More Realistic and Less Risky
PDF ↗Large language models (LLMs) are increasingly tasked with invoking enterprise APIs, yet they routinely falter when near-duplicate tools vie for the same user intent or when required arguments are left underspecified. We introduce **DiaFORGE** (**Dia**logue **F**ramework for **O**rganic **R**esponse **G**eneration **E**valuation), a disambiguation-centric, three-stage pipeline that (i) synthesizes persona-driven, multi-turn dialogues in which the assistant must distinguish among highly similar tools, (ii) performs supervised fine-tuning of open-source models with reasoning traces across 3B - 70B parameters, and (iii) evaluates real-world readiness via a dynamic suite that redeploys each model in a live agentic loop and reports end-to-end goal completion alongside conventional static metrics. On our dynamic benchmark DiaBENCH, models trained with DiaFORGE raise tool-invocation success by **27 pp over GPT-4o** and by **49 pp over Claude-3.5-Sonnet**, both under optimized prompting. To spur further research, we release an open corpus of **5000 production-grade enterprise API** specifications paired with rigorously validated, disambiguation-focused dialogues, offering a practical blueprint for building reliable, enterprise-ready tool-calling agents.
Large language models (LLMs) can generate fluent dialogue, but prior works lack situational grounding, dynamic strategy control, and evaluation aligned with clinical standards in motivational interviewing (MI). We introduce StoryMI, a multi-LLM agent framework for controllable MI dialogue generation, where questionnaire-based client profiles are expanded into situational stories that provide narrative context for the dialogue. Therapist and client agents generate MI-coded utterances guided by MI codes selected by the interaction agent, while an interaction agent dynamically coordinates exchanges to control MI strategies during a multi-turn conversation. We propose a two-level evaluation protocol: lexical metrics and MI-specific measures of macro-level counseling strategies, alongside LLM-as-judge and human expert assessments. We construct a dataset of 6K simulated MI dialogues grounded in 1K questionnaire-story pairs, covering 12 MI codes and 13 symptom domains, and benchmark six open- and closed-source LLMs. Our results show that situational grounding and macro-level control can improve MI adherence and clinical plausibility, demonstrating the effectiveness of a structured multi-agent workflow for psychotherapy dialogue generation. We provide code and data for reproducibility.
Group Relative Policy Optimization (GRPO) has become a standard approach for training mathematical reasoning models; however, GRPO training is computationally intensive and usually takes a long time, which consumes substantial computational resources and creates barriers for academic researchers and smaller organizations with limited GPU budgets. In this paper, we propose MMR-GRPO to accelerate GRPO training and reduce the overall training time required to reach peak performance, and the approach adopts Maximal Marginal Relevanceto reweigh rewards of multiple rollouts by balancing rollout quality with diversity to reduce rollout redundancy. The rationale is that redundant or similar completions, if repeatedly used to train a model, will create an “exploitation trap” and slow down model convergence in GRPO style reinforcement learning. Extensive evaluations across three model sizes (1.5B, 7B, 8B), three GRPO variants, and five mathematical reasoning benchmarks show that MMR-GRPO achieves comparable peak performance while requiring on average 47.9% fewer training steps and 70.2% less wall-clock time. These gains are consistent across models, methods, and benchmarks. Our code is released at: https://github.com/WeiKangda/MMR-GRPO.
EduMARS: Can Vision-Language Models Grade Like Teachers? Benchmarking Multimodal, Rubric-Based Assessment on Chinese K-12 Answers
PDF ↗Automated grading of student work is a critical application of AI in education. However, existing benchmarks fall short in evaluating models on realistic, cognitively demanding tasks. Most rely on synthetic, well-structured text inputs, overlooking the multimodal, error-prone, and often handwritten nature of real student responses, especially in K-12 settings. We introduce EduMARS, a multimodal benchmark designed for rubric-aligned evaluation of real Chinese K-12 student answers. The dataset contains over 4,500 authentic responses from high-stakes exams across eight subjects, featuring noisy handwriting,mixed-layout diagrams,mathematical expressions, and narrative reasoning. Each response is meticulously annotated by expert teachers using step-wise scoring rubrics, error classifications, and key-point mappings, providing fine-grained supervision aligned with real-world pedagogical practices. We evaluated existing SOTA MLLMs across the dimensions of final score and the reasoning process of grading, reveals a significant gap between existing SOTA MLLMs and human-level performance. To bridge this performance gap, we propose the Retrieval-Augmented Adaptive-Rubric Grading (RARG), enabling models to emulate expert grading logic by dynamically synthesizing case-specific evaluation schemas. RARG effectively enhances the performance and interpretability of various MLLMs on EduMARS, surpassing in-context learning and chain-of-thought.
NSL-MT: Linguistically Informed Negative Samples for Efficient Machine Translation in African Low-Resource Languages
PDF ↗We introduce negative space learning machine translation (NSL-MT), a training method for underresourced languages, that augments limited parallel data with synthetically generated violations of the target language’s grammar and explicitly penalizes the model when it assigns high probability to these linguistically invalid outputs. NSL-MT delivers improvements across all baselines we tested, including 3-12% BLEU gains for well-performing models and 56-89% gains for models lacking decent initial support. Furthermore, NSL-MT provides a 5x data efficiency multiplier: training with 1,000 examples matches or exceeds normal training with 5,000 examples. NSL-MT thus provides a data-efficient alternative training method for settings where parallel data is limited.
Multilingual speech foundation models such as Whisper are trained on web-scale data, where data for each language consists of a myriad of regional varieties. However, different regional varieties often employ different scripts to write the same language, rendering speech recognition output also subject to non-determinism in the output script. To mitigate this problem, we show that script is linearly encoded in the activation space of multilingual speech models, and that modifying activations at inference time enables direct control over output script. We find the addition of such script vectors to activations at test time can induce script change even in unconventional language-script pairings (e.g. Italian in Cyrillic and Japanese in Latin script). We apply this approach to inducing post-hoc control over the script of speech recognition output, where we observe competitive performance across all model sizes of Whisper.
Large Reasoning Models (LRMs) have achieved remarkable performance across diverse domains, yet their decision-making under conflicting objectives remains insufficiently understood. This work investigates how LRMs respond to harmful queries when confronted with two categories of conflicts: internal conflicts that pit alignment values against each other and dilemmas, which impose mutually contradictory choices, including sacrificial, duress, agent-centered, and social forms. Using over 1,300 prompts across five benchmarks, we evaluate three representative LRMs - Llama-3.1-Nemotron-8B, QwQ-32B, and DeepSeek R1 - and find that conflicts significantly increase attack success rates, even under single-round non-narrative queries without sophisticated auto-attack techniques. Our findings reveal through layerwise and neuron-level analyses that safety-related and functional representations shift and overlap under conflict, interfering with safety-aligned behavior. This study highlights the need for deeper alignment strategies to ensure the robustness and trustworthiness of next-generation reasoning models. Our code is available at https://github.com/DataArcTech/ConflictHarm. Warning: This paper contains inappropriate, offensive and harmful content.
One paradigm of language model (LM) fine-tuning relies on creating large training datasets, under the assumption that high quantity and diversity will enable models to generalize to novel tasks after post-training. In practice, gathering large sets of data is inefficient, and training on them is prohibitively expensive; worse, there is no guarantee that the resulting model will handle complex scenarios or generalize better. Moreover, existing techniques rarely assess whether a training sample provides novel information, resulting in unnecessary costs. In this work, we explore a new Test-Time Self-Improvement (TT-SI) algorithm to create more effective and generalizable agentic LMs on-the-fly. TT-SI can be summarized in three steps: (i) first it identifies the samples that model struggles with (self-awareness), (ii) then generates similar examples from detected uncertain samples (self-data augmentation), and (iii) uses these newly generated samples at test-time training (self-improvement). We further explore Test-Time Distillation (TT-D), which leverages a stronger supervisor for targeted data generation. Empirical evaluations across different agent benchmarks demonstrate that TT-SI improves the performance with +5.48% absolute accuracy gain on average across all benchmarks and surpasses other standard learning methods more efficiently. Our findings highlight the promise of TT-SI, demonstrating the potential of self-improvement algorithms at test-time as a new paradigm for building more capable agents toward self-evolution.
Evaluating the instruction-following (IF) capabilities of Multimodal Large Language Models (MLLMs) is essential for rigorously assessing how faithfully model outputs adhere to user-specified intentions. Nevertheless, existing benchmarks for evaluating MLLMs’ instruction-following capability primarily focus on verbal instructions in the textual modality. These limitations hinder a thorough analysis of instruction-following capabilities, as they overlook the implicit constraints embedded in the semantically rich visual modality. To address this gap, we introduce VC-IFEval, a new benchmark accompanied by a systematically constructed dataset that evaluates MLLMs’ instruction-following ability under multimodal settings. Our benchmark systematically incorporates vision-dependent constraints into instruction design, enabling a more rigorous and fine-grained assessment of how well MLLMs align their outputs with both visual input and textual instructions. Furthermore, by fine-tuning MLLMs on our dataset, we achieve substantial gains in visual instruction-following accuracy and adherence. Through extensive evaluation across representative MLLMs, we provide new insights into the strengths and limitations of current models.
Despite strong performance on code generation tasks, it remains unclear whether large language models (LLMs) genuinely reason about code execution. Existing code reasoning benchmarks primarily evaluate final output correctness under a single canonical implementation, leaving two critical aspects underexplored: (1) whether LLMs predictions are consistent to functionally equivalent implementations, and (2) whether LLMs can accurately reason about intermediate execution states. We introduce CoRE, a Code Reasoning benchmark that evaluates code reasoning through implementation invariance and process transparency. Extensive evaluations on eight frontier LLMs reveal two fundamental limitations. First, models exhibit a substantial robustness gap, with performance varying significantly across equivalent implementations. Second, we observe superficial execution, where models arrive at correct final outputs without correctly reasoning about intermediate execution states. Together, these findings demonstrate that output-only evaluations are insufficient for assessing code reasoning and position CoRE as a necessary benchmark for evaluating robust and faithful code reasoning.
While LLM watermarking is essential for machine- generated content identification, existing paraphrase-based attacks struggle to balance watermark removal efficacy with text quality. We propose TSAPA, a training-free evolutionary framework that models watermark removal as a constrained multi-objective optimization problem. By leveraging genetic algorithms to navigate the Pareto front, TSAPA utilizes a Pseudo-Log-Likelihood (PLL)-guided mutation to precisely target and modify watermark-carrying tokens. Experiments on Qwen3 series (1.7B/8B/32B) across multiple watermark schemes show that TSAPA achieves over 90% attack success rate (ASR) while maintaining high text semantic fidelity, significantly outperforming baselines methods. This work exposes critical vulnerabilities in current watermarks and provides a new perspective for robust evaluation.
PhysPRM: A Generative Process Reward Model with Fine-grained Diagnosis for Physics Problem Solving
PDF ↗Despite the remarkable progress of Large Language Models (LLMs) in abstract reasoning tasks, they continue to struggle with physics problem solving due to difficulties in decoding implicit constraints and maintaining physical consistency. To address these challenges, Process Reward Models (PRMs) have emerged as a promising approach to verify intermediate reasoning steps. Existing PRMs attempt to mitigate reasoning errors but typically rely on scalar scoring, which lacks the explanatory power necessary to diagnose complex physical misconceptions. In this work, we introduce PhysPRM, a Generative PRM that treats evaluation as a generative task to produce fine-grained diagnoses comprising critiques, final judgments, and specific error types. To facilitate this, we develop an automated data synthesis pipeline to construct PhysPRM30K, a comprehensive training dataset, and PhysProcessBench, a rigorously human-verified benchmark. By employing a two-stage training paradigm that integrates Supervised Fine-Tuning with Group Relative Policy Optimization, PhysPRM significantly enhances the physics reasoning capabilities of various LLMs. Extensive experiments demonstrate that PhysPRM improves performance across seven benchmarks in both Best-of-N and critique refinement strategies.