Reinforcement Learning with Verifiable Rewards (RLVR) is an essential paradigm that enhances the reasoning capabilities of Large Language Models (LLMs). However, existing methods typically rely on static policy optimization schemes that misalign with the model’s evolving reasoning capabilities. To address this issue, we propose Adaptive Power-Mean Policy Optimization (APMPO), which comprises two main innovations: Power-Mean Policy Optimization (PMPO) and Feedback-Adaptive Clipping (FAC). Specifically, PMPO introduces a generalized power-mean objective. This enables the model to adaptively transition from the signal-amplifying behavior of the arithmetic mean to the consistency-enforcing behavior of the geometric mean. FAC adaptively adjusts clipping bounds based on real-time reward statistics to overcome the limitations of static mechanisms. Capitalizing on these innovations, APMPO improves learning dynamics and reasoning performance. Extensive experiments on nine datasets across three reasoning tasks showcase the superiority of APMPO over state-of-the-art RLVR-based baselines. For instance, APMPO boosts the average Pass@1 score on mathematical reasoning benchmarks by 3.0 points compared to GRPO when using Qwen2.5-3B-Instruct.
论文检索
输入标题、作者或关键词，从 9,460 篇学术成果中精准定位
Reward-guided search methods have demonstrated strong potential in enhancing tool-using agents by effectively guiding sampling and exploration over complex action spaces. As a core design, those search methods utilize process reward models (PRMs) to provide step-level rewards, enabling more fine-grained monitoring. However, there is a lack of systematic and reliable evaluation benchmarks for PRMs in tool-use settings. In this paper, we introduce ToolPRMBench, a large-scale benchmark specifically designed to evaluate PRMs for tool-using agents. ToolPRMBench is built on top of several representative tool-use benchmarks and converts agent trajectories into step-level test cases. Each case contains the interaction history, a correct action, a plausible but incorrect alternative, and relevant tool metadata. We respectively utilize offline sampling to isolate local single-step errors and online sampling to capture realistic multi-step failures from full agent rollouts. A multi-LLM verification pipeline is proposed to reduce label noise and ensure data quality. We conduct extensive experiments across large language models, general PRMs, and tool-specialized PRMs on ToolPRMBench. The results reveal clear differences in PRM effectiveness and highlight the potential of specialized PRMs for tool-using. Our code and dataset are available at: https://github.com/David-Li0406/ToolPRMBench[More resources on LLM-as-a-judge are on the website: <https://llm-as-a-judge.github.io>].
Preference optimization is fundamental for aligning large language models. While existing methods use sample weighting, they typically rely on static functions of instantaneous model states and ignore temporal learning dynamics. We contend that a sample’s value evolves throughout training, characterized by patterns such as stable convergence or noisy oscillation. We propose MetaPO, a framework that meta-learns adaptive weights using three temporal features: reward margin evolution, learning volatility, and reference deviation. Through bilevel optimization on validation data, MetaPO automatically discovers weighting strategies tailored to specific datasets. Experiments on models ranging from 7B to 70B parameters demonstrate statistically significant improvements over strong baselines, achieving gains of up to 2.4 points on AlpacaEval 2.0 and Arena-Hard. Interpretability analysis confirms that temporal features drive over 70% of the weighting decisions and that the learned weights correlate strongly with sample quality.
Defending Large Language Models (LLMs) against backdoor attacks has long been trapped in a "cat-and-mouse" dilemma, where defenders passively react to ever-shifting attack strategies. To break this cycle, we posit that proactive immunization is inherently superior to reactive sanitization. In this study, we propose Poison-to-Poison (P2P), a general and effective defense algorithm that introduces a paradigm shift. Instead of waiting to detect malicious samples, P2P strategically implants benign triggers to reshape the model’s decision boundary, redirecting latent feature activation from malicious trajectories to a safe, controllable output space. This enforces the model to associate trigger-induced representations with safe outputs, thereby overriding the effects of original malicious triggers. Thanks to this robust and generalizable trigger-based fine-tuning, P2P is effective across task settings and attack types. Theoretically and empirically, we show that P2P can neutralize malicious backdoors while preserving task performance. We conduct extensive experiments on classification, mathematical reasoning, and summary generation tasks, involving multiple state-of-the-art LLMs. The results demonstrate that our P2P algorithm significantly reduces the attack success rate compared with baseline models. We hope that P2P can serve as a practical guideline for defending against backdoor attacks in the Model as a Service (MaaS) scenario, where benign prompts are embedded within the system to regulate model behavior.
A Herd of Language Models Makes a Better Zero-shot Annotator for Clinical Named Entity Recognition
PDF ↗Clinical named entity recognition (NER) remains difficult to scale due to the high cost of manual annotation. Although large language models (LLMs) enable zero-shot annotation, their performance on clinical NER is still limited. To this end, we improve the annotation quality by aggregating annotations from *a herd of diverse LLMs*, including general-purpose, medically adapted, and NER-specialized models. A key challenge in this multi-LLM setting is effectively leveraging entities extracted by only a minority of models: although they account for a substantial portion of true positives, they are heavily intermixed with noise. To address this, we introduce **MARY**, a label-modeling method for **M**ulti-LLM **A**nnotation using **R**epresentation learning to capture contextual similarit**Y**. During aggregation, MARY selectively incorporates minority-extracted entities whose contexts are similar to those of majority-extracted entities, yielding more reliable and comprehensive annotations. Experimental results show that MARY improves the average F1 score by 8.6% over vanilla zero-shot baselines while reducing annotation costs.
A major challenge for the operation of large language models (LLMs) is how to predict whether a specific LLM will produce sufficiently high-quality output for a given query. Existing approaches rely on external classifiers, most commonly BERT-based models, which suffer from limited context windows, constrained representational capacity, and additional computational overhead. We propose IntroLM, a method that enables causal language models to predict their own output quality during the prefilling phase without affecting generation using [CPX] tokens. By introducing token-conditional LoRA that activates only for the introspective [CPX] token, the model learns to predict the output quality for a given query while preserving the original backbone behavior and avoiding external evaluators. On question-answering benchmarks, IntroLM applied to Qwen3-8B achieves a ROC–AUC of 90% for success prediction, outperforming a DeBERTa-v3-Large classifier by 14%. When integrated into multi-model routing systems, IntroLM achieves superior cost–performance trade-offs, reducing end-to-end latency by up to 33% and large-model usage by up to 50% at matched reliability. Our code is available at https://github.com/hhosseini1377/LLM_routing.
Double-Edged Sword or Sharp Tool? Designing and Evaluating Triadic LLM-Teacher Collaboration for K-12 Writing at Scale
PDF ↗The double-edged sword of integrating Large Language Models (LLMs) requires an effective triadic collaboration mechanism among LLMs, teachers and students, especially for K-12 education. By developing a triadic collaboration system to support K-12 writing learning, a multidimensional evaluation framework grounded in Systemic Functional Linguistics and the suggestion trajectory tracing pipeline, this paper contributes a large-scale empirical dataset involving 57,954 essays from 10,195 students across 120 schools over two years. Our findings confirm the efficacy of this system in improving writing quality through a strategic labor division: the LLM serves as a generative engine to mitigate teacher burnout, and the teacher acts as a pedagogical gatekeeper and bridge to guarantee feedback quality. While both LLM and teacher are critical for skill improvement, we uncover a ceiling effect where excessive linguistic expansion yields diminishing marginal utility. These suggest a dynamically adaptive LLM-teacher collaboration as student proficiency increases.
Tool-Integrated Reasoning (TIR) with search engines enables large language models to iteratively retrieve up-to-date external knowledge, enhancing adaptability and generalization in complex question-answering tasks. However, existing search agent pipelines typically depend on reinforcement learning based optimization, which often suffers from sparse outcome rewards, leading to inefficient exploration and unstable training. We introduce CriticSearch, a fine-grained credit-assignment framework that supplies dense, turn-level feedback via a retrospective critic mechanism. During training, a frozen, asymmetric critique LLM retrospectively evaluates each turn using privileged information from the full trajectory and gold answers, converting these assessments into stable, dense rewards that guide policy improvement. Experimental results across diverse multi-hop reasoning benchmarks demonstrate that CriticSearch consistently outperforms existing baselines, achieving faster convergence, improved training stability, and higher performance.
Current evaluations of agents remain centered around one-shot task completion, failing to account for the inherently iterative and collaborative nature of many real-world problems, where human goals are often underspecified and evolve. We argue for a shift from building and assessing task completion agents to developing collaborative agents, assessed not only by the quality of their final outputs but by how well they engage with and enhance human effort throughout the problem-solving process. To support this shift, we introduce collaborative effort scaling, a framework that captures how an agent’s utility grows with increasing user involvement. Through case studies and simulated evaluations, we show that state-of-the-art agents often underperform in multi-turn, real-world scenarios, revealing a missing ingredient in agent design: the ability to sustain engagement and scaffold user understanding. Collaborative effort scaling offers a lens for diagnosing agent behavior and guiding development toward more effective interactions.
Diagnosing LLMs via Information Spectrum Analysis: Tail Behavior and the Effects of Side Information
PDF ↗Large language models (LLMs) exhibit non-stationary generation: their output distributions shift with prompts, retrieved documents, and decoding conditions. Under such variability, average likelihood metrics can obscure heterogeneous behaviors across samples, especially in high-surprisal tails where failures often occur. We propose an information-spectrum-based diagnostic framework that treats LLMs as general sources without assuming stationarity, ergodicity, or the asymptotic equipartition property. We define sequence-level self-information density (coding rate; mean surprisal) and construct an empirical information spectrum from finite samples, enabling operational estimates of spectrum quantiles and width. We further introduce an information gain spectrum, a teacher-forced likelihood-based measure that evaluates the same generated sequence with and without side information. Across multiple Japanese LLMs and QA settings, we observe that correctness differences are often more visible in the high-surprisal tail than in the mean coding rate, and that side information can reshape tail behavior in heterogeneous ways across sequences. We also observe that instruction tuning changes the spectrum structure, making tail statistics and spectrum width more predictive of correctness than the mean coding rate. Overall, our analysis illustrates how spectrum-based diagnostics complement average-based metrics for understanding conditional generation.
Large-scale multitask benchmarks have driven rapid progress in language modeling, yet most emphasize high-resource languages such as English, leaving Bengali underrepresented. We present BnMMLU, a comprehensive benchmark for measuring massive multitask language understanding in Bengali. BnMMLU spans 41 domains across STEM, humanities, social sciences, and general knowledge, and contains 134,375 multiple-choice question–option pairs-the most extensive Bengali evaluation suite to date. The dataset preserves mathematical content via MathML, and includes BnMMLU-HARD, a compact subset constructed from questions most frequently missed by top systems to stress difficult cases. We benchmark 24 model variants across 11 LLM families, spanning open-weights general/multilingual, Bengali-centric open-weights, and proprietary models, covering multiple parameter scales and instruction-tuned settings. We evaluate models under standardized protocols covering two prompting styles (Direct vs. Chain-of-Thought) and two context regimes (0-shot vs. 5-shot), reporting accuracy consistently across families. Our analysis highlights persistent gaps in reasoning and application skills and indicates sublinear returns to scale across model sizes. We release the dataset and evaluation templates to support rigorous, reproducible assessment of Bengali language understanding and to catalyze progress in multilingual NLP.
"I Don’t Know What to Say": A Fact-Filling Questionnaire Method to Help Non-Experts Talk to LegalAI Assistant
PDF ↗Artificial intelligence has become increasingly prevalent in the legal domain. However, LegalAI systems often struggle with vague user queries that lack essential legal details, leading to suboptimal performance in practical applications. To address this challenge, we propose FactFiller, a novel approach that dynamically generates questionnaires to help users refine their input queries. Our method leverages an iterative training process that collects valuable questionnaires, eliminating the need for human annotation. Additionally, we introduce a "case-law-quiz” cascading retrieval process, ensuring that the generated questions and answer options are directly linked to specific legal provisions. Through the user study and the downstream task experiments, we demonstrate that FactFiller, while remaining easy for non-experts to understand, not only improves the completeness of queries but also ensures the performance of various domain-specific models in downstream legal tasks.
Can LLMs Understand the Impact of Trauma? Costs and Benefits of LLMs Coding the Interviews of Firearm Violence Survivors
PDF ↗Firearm violence is a pressing public health issue, yet research into survivors’ lived experiences remains underfunded and difficult to scale. Qualitative research, including in-depth interviews, is a valuable tool for understanding the personal and societal consequences of community firearm violence and designing effective interventions. However, manually analyzing these narratives through thematic analysis and inductive coding is time-consuming and labor-intensive. Recent advancements in large language models (LLMs) have opened the door to automating this process, though concerns remain about whether these models can accurately and ethically capture the experiences of vulnerable populations. In this study, we assess the use of open-source LLMs to inductively code interviews with 21 Black men who have survived community firearm violence. Our results demonstrate that while some configurations of LLMs can identify important codes, overall relevance remains low and is highly sensitive to data processing. Furthermore, LLM guardrails lead to substantial narrative erasure. These findings highlight both the potential and limitations of LLM-assisted qualitative coding and underscore the ethical challenges of applying AI in research involving marginalized communities.
Large Language Models (LLMs) face information overload when handling long contexts, particularly in Retrieval-Augmented Generation (RAG) where extensive supporting documents introduce redundant content that interferes with reasoning. Context engineering has emerged to address these challenges, yet existing methods rely on lexical or token-level features that fragment semantic units and fail to capture conceptually essential content. We propose an unsupervised context compression framework leveraging Abstract Meaning Representation (AMR) to preserve semantically essential information while filtering irrelevant text. By quantifying node-level entropy within AMR graphs, our method estimates the conceptual importance of each node, enabling retention of core semantics. Specifically, we construct AMR graphs from retrieved contexts, compute the conceptual entropy of each node, and identify statistically significant concepts to form a condensed, semantically focused context. Experiments on the PopQA and EntityQuestions datasets demonstrate that our method outperforms vanilla RAG and existing baselines, achieving superior accuracy while substantially reducing context length. To the best of our knowledge, this is the first work introducing AMR-based conceptual entropy for context compression, demonstrating the potential of structured linguistic representations in context engineering.
We propose Data Swarms, an algorithm to optimize the generation of synthetic evaluation data and advance quantitative desiderata of LLM evaluation. We first train a swarm of initial data generators using existing data, and define various evaluation objectives to reflect the desired properties of evaluation (e.g., generate more difficult problems for the evaluated models) and quantitatively evaluate data generators. We then employ particle swarm optimization to optimize the swarm of data generators, where they collaboratively search through the model parameter space to find new generators that advance these objectives. We further extend it to Adversarial Swarms, where the data generator swarm generates harder data while the test taker model swarm learns from such data, co-evolving dynamically for better data and models simultaneously. Extensive experiments demonstrate that Data Swarms outperforms eight data generation baselines across five evaluation objectives, while Adversarial Swarms produce more robust learning of synthetic data and stronger generalization. Further analysis reveals that Data Swarms successfully optimizes compositions of multiple evaluation objectives and generalizes to new off-the-shelf LLMs, unseen at optimization time.
The success of vision-language models is primarily attributed to effective cross-modal alignment between vision and language. However, modality gaps persist even in well-aligned models and may be necessary for human perception, as evidenced by modality-specific phenomena such as visual texture and linguistic tone. These observations motivate us to computationally measure and leverage modality gaps to explore their utility in downstream applications. In this paper, we introduce the Modality Dominance Score (MDS), which attributes multimodal features to specific modalities by categorizing them as vision-dominant, language-dominant, or cross-modal. We then propose automatic interpretability metrics to evaluate these modality-specific features in a scalable manner. Finally, we demonstrate how the identified modality-specific features enable training-free probing and editing methods for understanding model perception across genders, generating adversarial examples, and controlling text-to-image generation. Combined with task-agnostic interpretability tools, our work provides a systematic framework for analyzing and efficiently controlling multimodal models.
Distilling the capabilities from a large reasoning model (LRM) to a smaller student model often involves training on substantial amounts of reasoning data. However, knowledge distillation (KD) over lengthy sequences with prompt (P), chain-of-thought (CoT), and answer (A) sections makes the process computationally expensive. In this work, we investigate how the allocation of supervision across different sections (P, CoT, A) affects student performance. Our analysis shows that selective KD over only the CoT tokens can be effective when the prompt and answer information is encompassed by it. Building on this insight, we establish a truncation protocol to quantify computation-quality tradeoffs as a function of sequence length. We observe that beyond a specific length, longer training sequences provide marginal returns for downstream performance but require substantially higher memory and FLOPs. To this end, training on only the first 50% of tokens of every training sequence can retain, on average, ≈91% of full-sequence performance on math benchmarks while reducing training time, memory usage, and FLOPs by about 50% each. Codes are available at https://github.com/weiruichen01/distilling-the-essence.
Previous research has sought to enhance the graph reasoning capabilities of LLMs by supervised fine-tuning on synthetic graph data. While these led to specialized LLMs better at solving graph algorithm problems, we don’t need LLMs for shortest path: we need generalization from synthetic graph data to real-world tasks with implicit graph structures. In this work, we propose to unlock generalizable learning of graph with post-training alignment with synthetic data. We first design solution-based and process-based rewards for synthetic graph problems: instead of rigid memorizing response patterns in direct fine-tuning, we posit that post-training alignment would help LLMs grasp the essentials underlying graph reasoning and alleviate overfitting on synthetic data. We employ post-training alignment algorithms such as GRPO and DPO, aligning both off-the-shelf LLMs and LLMs fine-tuned on synthetic graph data. We then compare them against existing settings on both in-domain synthetic tasks and out-of-domain real-world tasks with implicit graph structures such as multi-hop QA, structured planning, and more. Extensive experiments demonstrate that our post-training alignment recipe leads to statistically significant improvement on 5 datasets, with an average gain of 12.9% over baseline settings. Further analysis reveals that process-based rewards consistently outperform solution-based rewards on synthetic data but not on real-world tasks, and compositionality and explainable intermediate steps remains a critical challenge even after post-training alignment.
Reasoning in humans is prone to biases due to underlying motivations like identity protection, that undermine rational decision-making and judgment. This motivated reasoning at a collective level can be detrimental to society when debating critical issues such as human-driven climate change or vaccine safety, and can further aggravate political polarization. Prior studies have reported that large language models (LLMs) are also susceptible to human-like cognitive biases, however, the extent to which LLMs selectively reason toward identity-congruent conclusions remains largely unexplored. Here, we investigate whether assigning 8 personas across 4 political and socio-demographic attributes induces motivated reasoning in LLMs. Testing 8 LLMs (open source and proprietary) across two reasoning tasks from human-subject studies — veracity discernment of misinformation headlines and evaluation of numeric scientific evidence — we find that persona-assigned LLMs have up to 9% reduced veracity discernment relative to models without personas. Political personas specifically are up to 90% more likely to correctly evaluate scientific evidence on gun control when the ground truth is congruent with their induced political identity. Prompt-based debiasing methods are largely ineffective at mitigating these effects. Taken together, our empirical findings are the first to suggest that persona-assigned LLMs exhibit human-like motivated reasoning that is hard to mitigate through conventional debiasing prompts — raising concerns of exacerbating identity-congruent reasoning in both LLMs and humans.
Deploying large language models in long-horizon, goal-oriented interactions remains challenging because similar entities and facts recur under different latent goals and con-straints, causing memory systems to retrieve context-mismatched evidence. We propose STITCH (Structured Intent Tracking in Contextual History), an agentic memory system that indexes each trajectory step with a structured retrieval cue, contextual intent, and retrieves history by matching the current step’s intent. Contextual intent provides compact signals that disambiguate repeated mentions and reduce interference: (1) the current latent goal defining a thematic segment, (2) the action type, and (3) the salient entity types anchoring which attributes matter. During inference, STITCH filters and prioritizes memory snippets by intent compatibility, suppressing semantically similar but context-incompatible history.For evaluation, we introduce CAME-Bench, a benchmark for context-aware retrieval in realistic, dynamic, goal-oriented trajectories. Across CAME-Bench and LongMemEval, STITCH achieves state-of-the-art performance, outperforming the strongest baseline by 35.6%, with the largest gains as trajectory length increases. Our analysis shows that intent indexing substantially reduces retrieval noise, supporting intent-aware memory for robust long-horizon reasoning.