Reinforcement learning with verifiable rewards (RLVR) has delivered impressive gains in mathematical and multimodal reasoning and has become a standard post-training paradigm for contemporary language and vision-language models. However, the RLVR recipe introduces a significant risk of capability regression, where models forget foundational skills after prolonged training without employing regularization strategies. We empirically confirm this concern, observing that open-source reasoning models suffer performance degradation on core capabilities such as perception and faithfulness. While imposing regularization terms like KL divergence can help prevent deviation from the base model, these terms are calculated on the current task, thus they do not guarantee broader knowledge. Meanwhile, commonly used experience replay across heterogeneous domains makes it nontrivial to decide how much training focus each objective should receive. To address this, we propose RECAP—a replay strategy with dynamic objective reweighting for general knowledge preservation. Our reweighting mechanism adapts in an online manner using short-horizon signals of convergence and instability, shifting the post-training focus away from saturated objectives and toward underperforming or volatile ones. Our method is end-to-end and readily applicable to existing RLVR pipelines without training additional models or heavy tuning. Extensive experiments on benchmarks based on Qwen2.5-VL-3B and Qwen2.5-VL-7B demonstrate the effectiveness of our method, which not only preserves general capabilities but also improves reasoning by enabling more flexible trade-offs among in-task rewards.
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
FAMA: Failure-Aware Meta-Agentic Framework for Open-Source LLMs in Interactive Tool Use Environments
PDF ↗Large Language Models are being increasingly deployed as the decision-making core of autonomous agents capable of effecting change in external environments. Yet, in conversational benchmarks, which simulate real-world customer-centric issue resolution scenarios, these agents frequently fail due to the cascading effects of incorrect decision-making. These challenges are particularly pronounced for open-source LLMs with smaller parameter sizes, limited context windows, and constrained inference budgets, which contribute to increased error accumulation in agentic settings. To tackle these challenges, we present the **Failure-Aware Meta-Agentic (FAMA)** framework. FAMA operates in two stages: first, it analyzes failure trajectories from baseline agents to identify the most prevalent errors; second, it employs an orchestration mechanism that activates a minimal subset of specialized agents tailored to address these failures by injecting a targeted context for the tool-use agent before the decision-making step. Experiments across open-source LLMs demonstrate performance gains up to **27%** across evaluation modes over standard baselines. These results highlight that targeted curation of context through specialized agents to address common failures is a valuable design principle for building reliable, multi-turn tool-use LLM agents that simulate real-world conversational scenarios.
Probing the Plasticity and Correlation of LLM Value Systems: LLM Value Rankings are Not Stable
PDF ↗The value alignment of Large Language Models (LLMs) is critical because value is the foundation of LLM decision-making and behavior. Some recent work show that LLMs have similar value rankings. However, little is known about how susceptible LLM value rankings are to external influence and how different values are correlated with each other. In this work, we investigate the plasticity of LLM value systems by examining how their value rankings are influenced by different prompting strategies and exploring the intrinsic relationships between values. To this end, we design 6 different value transformation prompting methods including direct instruction, rubrics, in-context learning, scenario, persuasion, and persona, and benchmark the effectiveness of these methods on 3 different families and totally 8 LLMs. Our main findings include that the value rankings in large LLMs are much more susceptible to external influence than small LLMs, and there are intrinsic correlations between certain values (e.g., Privacy and Respect). Besides, through detailed correlation analysis, we find that the value correlations are more similar between large LLMs of different families than small LLMs of the same family. We also identify that scenario method is the strongest persuader and can help entrench the value rankings.
Identifying the Achilles’ Heel: An Iterative Method for Uncovering Factual Errors in Large Language Models
PDF ↗Large Language Models (LLMs) like ChatGPT are foundational in various applications due to their extensive knowledge from pre-training and fine-tuning. Despite this, they are prone to generating factual and commonsense errors, raising concerns in critical areas like healthcare, journalism, and education to mislead users. Current methods for evaluating LLMs’ veracity are limited by the need for extensive human labor, test data contamination, or limited scope, hindering efficient and effective exposure of errors. To address these challenges, we propose HalluHunter, a novel, fully automated framework for systematically uncovering factual inaccuracies in LLMs. HalluHunter employs a knowledge-graph-based approach, extracting fact triplets to generate diverse question types for single- and multi-hop reasoning using rule-based Natural Language Processing (NLP) techniques. Its iterative process starts with random triplet selection for question generation, followed by adaptive selection in subsequent iterations, targeting triplets where LLMs frequently err based on their performance analysis. Our extensive tests on nine prominent LLMs reveal that HalluHunter can trigger factual errors in up to 55% of questions in these models. Moreover, we demonstrate that HalluHunter’s test cases, particularly in adaptive selection, could further expose the weaknesses in benchmarking the factuality in LLMs meanwhile maintaining the coverage of questions. All code, data, and results will be released for future research.
We annotate PubMed Central paragraphs for document type, domain, and educational quality using a two-stage pipeline: Llama-3.1-70B labels 400K paragraphs, then a fine-tuned XLM-RoBERTa propagates annotations to the full corpus. This paragraph-level approach captures content diversity within scientific articles that document-level labels miss. The resulting Biomed-Enriched corpus contains 2M clinical case paragraphs, providing a publicly available alternative to restricted clinical datasets. For decoders, continual pretraining experiments enable targeted improvements, with clinical upsampling boosting performance by 4 points on MMLU ProfMed and educational filtering improving MedQA and MedMCQA by ~1 point. Combinations of these techniques led to faster convergence, reaching the same performance with a third of training tokens. For encoders, our best recipe matches BioClinical-ModernBERT on 11 tasks (77.3% vs 77.1% F1) while using 2.5x fewer tokens and only public data.
Generating coherent, semantically accurate text from large structured inputs remains a persistent challenge in data-to-text generation, as single-step LLM mappings from data-to-text limit control over discourse structuring and amplify hallucinations and omissions as input size grows. We introduce a new dataset of extended DBpedia triple sets (up to 199 triples per input), and a modular multi-agent framework: specialised LLM agents handle content ordering, text structuring, and surface realisation under the supervision of an orchestrator and guardrail control loop. The system generates multi-paragraph outputs in English and Irish (low-resource). We compare a three-worker multi-agent configuration against a single-worker multi-task variant and a strong end-to-end baseline. Quality is assessed via human evaluation and LLM-as-a-judge (with truncation-based sanity checks). Results show slightly superior coherence for the multi-agent approach in both languages, with statistically significant inter-rater correlation over all criteria for English and no statistically significant correlation for Irish. Human-LLM alignment is very weak overall, thus exposing key limits in scalable NLG evaluation.
TRACE: Two-Phase RL for Causal Graph Exploration and Deeper Psychological Intervention in Dynamic Counseling Scenarios
PDF ↗LLMs have shown promise in mental health counseling, but existing models are limited to surface-level empathy or predefined therapeutic procedures and lack the ability to actively explore the root causes of psychological distress. Inspired by case conceptualization, we formalize counseling as the online reconstruction of a client’s underlying causal graph through multi-turn dialogue. To this end, we propose TRACE, a two-phase reinforcement learning framework. It implements a causal-graph-driven reward scheme across two phases: an exploration phase that rewards the causal graph reconstruction following a surface-to-deep path, and an intervention phase that rewards targeted restructuring of irrational beliefs. Extensive experiments show that TRACE outperforms existing models, enabling causal-chain-aware psychological intervention beyond surface-level empathy.
Speakers of under-represented languages face both a language barrier, as most online knowledge is in a few dominant languages, and a modality barrier, since information is largely text-based while many languages are primarily oral. We address this for French-Wolof by training the first bilingual speech-text Matryoshka embedding model, enabling efficient retrieval of French text from Wolof speech queries without relying on a costly ASR-translation pipelines. We introduce large-scale data curation pipelines and new benchmarks, compare modeling strategies, and show that modality fusion within a frozen text Matryoshka model performs best. Although trained only for retrieval, the model generalizes well to other tasks, such as speech intent detection, indicating the learning of general semantic representations. Finally, we analyze cost-accuracy trade-offs across Matryoshka dimensions and ranks, showing that information is concentrated only in a few components, suggesting potential for efficiency improvements.
Aligning Large Language Models (LLMs) and Large Multimodal Models (LMMs) to human preferences is crucial for improving their real-world behavior. A common approach is to use reward models that enable reinforcement-learning post-training. However, traditional reward modeling requires finetuning on large preference datasets, limiting adaptability to new preferences. We introduce Activation Reward Models (Activation RMs)—the first mechanistic interpretability approach that steers LLM activations to align with few-shot preference data without finetuning. Our method combines activation denoising and output token likelihood scoring, achieving state-of-the-art performance on standard reward modeling benchmarks, surpassing zero-shot, few-shot, and voting-based baselines. We further demonstrate that Activation RMs mitigate reward hacking behaviors and remain robust to noisy exemplars and spurious reward signals. To evaluate this, we propose PreferenceHack, a novel few-shot benchmark testing reward models on reward hacking in a paired preference format, where Activation RMs achieve state-of-the-art performance, surpassing GPT-4o.
Large language models (LLMs) are increasingly deployed in monetization-driven systems such as search engines, advertising platforms, and e-commerce services, where decision making is shaped by complex interactions among user intent, advertiser objectives, and platform constraints. Despite rapid progress, existing benchmarks primarily focus on shopping-centric scenarios and user-facing data, capturing only a limited subset of real-world monetization pipelines and overlooking intermediate decision stages and robustness considerations. In this work, we introduce MonBench, a high-quality multi-task benchmark designed to evaluate LLMs in realistic monetization contexts. The benchmark is constructed from large-scale production data collected from multiple search engines, including both intermediate candidate pools and user-visible outcomes, better reflecting the distributional characteristics of real monetization systems. MonBench covers key capability dimensions such as intent understanding, commercial matching, and user behavior modeling, and adopts a unified multiple-choice formulation to enable systematic comparison across models. We further propose a comprehensive evaluation protocol that measures both performance and robustness. We evaluate a diverse set of state-of-the-art LLMs and conduct detailed task-level analyses. Our results reveal monetization-specific behaviors, including gaps between relevance optimization and broader decision-making capabilities, as well as differences in robustness across model families. These findings provide new insights into the strengths and limitations of current LLMs and highlight the need for richer domain-specific supervision in monetization-oriented applications.
Diversity has been gaining interest in the NLP community in recent years. At the same time, state-of-the-art transformer models such as ModernBERT use very large pre-training datasets, which are driven by size rather than by diversity. This summons to investigate theimpact of diversity on pre-training. We do so in this study, with the express intent of reducing pre-training dataset size, while retaining atleast comparable performance. We compare diversity-driven sampling algorithms, and we use the best one to pre-train several ModernBERT models on French with a fixed compute budget. We fine-tune and evaluate them on a variety of French benchmarks. We compare them with models pre-trained on randomly sampled data of commensurate size, with the same compute budget. We find that both random and diversity-driven sampling may reduce the pre-training dataset by up to 94% and the pre-training time by up to 73% while maintaining performance. Moreover, in some tasks, the inherent quality of models, estimated via head-only fine-tuning, is up to 10 points higher with diversity sampling than with random sampling.
Personalized image aesthetics assessment (PIAA) is an important research problem with practical real-world applications. While methods based on vision-language models (VLMs) are promising candidates for PIAA, it remains unclear whether they internally encode rich, multi-level aesthetic attributes required for effective personalization. In this paper, we first analyze the internal representations of VLMs to examine the presence and distribution of such aesthetic attributes, and then leverage them for lightweight, individual-level personalization without model fine-tuning. Our analysis reveals that VLMs encode diverse aesthetic attributes that propagate into the language decoder layers. Building on these representations, we demonstrate that simple linear models can achieve effective personalized image aesthetics assessment. We further analyze how aesthetic information is transferred across layers in different VLM architectures and across image domains. Our findings provide insights into how VLMs can be utilized for modeling subjective, individual aesthetic preferences.
While Large Language Models (LLMs) have emerged as powerful foundational models to solve a variety of tasks, they have also been shown to be prone to hallucinations, i.e., generating responses that sound confident but are actually incorrect or even nonsensical. Existing hallucination detectors propose a wide range of empirical scoring rules, but their performance varies across models and datasets, and it is hard to determine which ones to rely on in practice or to treat as a reliable detector. In this work, we formulate the problem of detecting hallucinations as a hypothesis testing problem and draw parallels with the problem of out-of-distribution detection in machine learning models. We then propose a multiple-testing-inspired method that systematically aggregates multiple evaluation scores via conformal p-values, enabling calibrated detection with controlled false alarm rate. Extensive experiments across diverse models and datasets validate the robustness of our approach against state-of-the-art methods.
Existing hate speech detection models are often opaque and rely on surface-level lexical cues, which makes them vulnerable to spurious correlations and limits robustness, interpretability and cultural contextualization. We propose Supervised Moral Rationale Attention (SMRA), the first self-explaining hate speech detection framework to incorporate moral rationales as direct supervision for attention alignment. Based on Moral Foundations Theory, SMRA aligns token-level attention with expert-annotated moral rationales, guiding models to attend to morally salient spans. Unlike prior rationale-supervised or post-hoc approaches, SMRA integrates moral rationale supervision directly into the training objective, producing inherently interpretable and contextualized explanations. To support our framework, we also introduce HateBRMoralXplain, a Brazilian Portuguese benchmark dataset annotated with hate labels, moral categories, token-level moral rationales, and socio-political metadata. Across binary hate speech detection and multi-label moral sentiment classification, SMRA consistently improves performance while enhancing both faithful and plausible explanations. Although explanations become more concise, sufficiency decreases, indicating more compact and informative rationales. Fairness remains stable, suggesting that improvements in explanation quality do not introduce significant bias trade-offs.
Retrieval-augmented generation (RAG) has become a core technique for improving the factuality and reasoning ability of large language models. Recent efforts extend RAG with graph-structured knowledge, enhancing retrieval to capture relational context beyond isolated text chunks. However, many graph-based RAG systems rely on a two-stage pipeline: (i) classical approximate nearest neighbor (ANN) search to identify top-k entities in the embedding space, (ii) heuristic neighbor expansion which augments the retrieved set by traversing immediate neighbors. This design underutilizes graph topology during retrieval and often introduces noisy or high-degree neighbors, leading to suboptimal evidence selection. In this paper, we propose TopoRAG, a retrieval framework that directly integrates structural constraints into ANN search via a diameter-constrained formulation. By selecting entities whose induced subgraph satisfies a diameter bound, TopoRAG enables topology-aware and noise-controlled graph retrieval. Experiments show that our approach consistently improves precision and significantly reduces context redundancy compared to existing methods.
Investigating Links between Illicit Massage Businesses through Natural Language Processing and Graph Machine Learning
PDF ↗Human trafficking exploits vulnerable individuals through forced sex or labor. Illicit massage businesses offer a clandestine front to illicit activities by disguising themselves as legitimate businesses. This makes it challenging for law enforcement agencies and anti-trafficking organizations to detect these enterprises and their associated entities, disrupt the network, and save victims. We adopt a multi-stream data integration approach primarily focusing on consumer-generated business reviews on Yelp.com, enriched with features from contextual data sources, such as the U.S. Census and business license records. We propose a novel decision support framework that extends the traditional link prediction methods by defining a higher-order neighborhood to detect links between pairs of massage businesses and the exposure of businesses to illicit activities related to human trafficking. We achieve this by introducing a bespoke subgraph extraction strategy in GNNs where the node features are derived using NLP techniques. Comprehensive experimental results demonstrate the competitive performance of our approach over the baseline methods.
Before Forgetting, Learn to Remember: Revisiting Foundational Learning Failures in LVLM Unlearning Benchmarks
PDF ↗While Large Vision-Language Models (LVLMs) offer powerful capabilities, they pose privacy risks by unintentionally memorizing sensitive personal information. Current unlearning benchmarks attempt to mitigate this using fictitious identities but overlook a critical stage 1 failure: models fail to effectively memorize target information initially, rendering subsequent unlearning evaluations unreliable. Diagnosing under-memorization and the multi-hop curse as root causes, we introduce ReMem, a Reliable Multi-hop and Multi-image Memorization Benchmark. ReMem ensures robust foundational learning through principled data scaling, reasoning-aware QA pairs, and diverse visual contexts. Additionally, we propose a novel Exposure metric to quantify the depth of information erasure from the model’s internal probability distribution. Extensive experiments demonstrate that ReMem provides a rigorous and trustworthy framework for diagnosing both learning and unlearning behaviors in LVLMs.
Large Multimodal Models (LMMs) have achieved remarkable success in visual understanding, yet they struggle with knowledge-intensive queries involving long-tail entities or evolving information due to static parametric knowledge. Recent search-augmented approaches attempt to address this limitation, but existing methods rely on indiscriminate whole-image retrieval that introduces substantial visual redundancy and noise, and lack deep iterative reflection, limiting their effectiveness on complex visual queries. To overcome these challenges, we propose Glance-or-Gaze (GoG), a fully autonomous framework that shifts from passive perception to active visual planning. GoG introduces a Selective Gaze mechanism that dynamically chooses whether to glance at global context or gaze into high-value regions, filtering irrelevant information before retrieval. We design a dual-stage training strategy: Reflective GoG Behavior Alignment via supervised fine-tuning instills the fundamental GoG paradigm, while Complexity-Adaptive Reinforcement Learning further enhances the model’s capability to handle complex queries through iterative reasoning. Experiments across six benchmarks demonstrate state-of-the-art performance. Ablation studies confirm that both Selective Gaze and complexity-aware RL are essential for effective visual search. We will release our data and models for further exploration soon.
Scientific rigour tends to be sidelined in favour of bold statements, leading authors to overstate claims beyond what their results support. We present RIGOURATE, a two-stage multimodal framework that retrieves supporting evidence from a paper’s body and assigns each claim an overstatement score. The framework consists of a dataset of over 10K claim–evidence sets from ICLR and NeurIPS papers, annotated using eight LLMs, with overstatement scores calibrated using peer-review comments and validated through human evaluation. It employes a fine-tuned reranker for evidence retrieval and a fine-tuned model to predict overstatement scores with justification. Compared to strong baselines, RIGOURATE enables improved evidence retrieval and overstatement detection. Overall, our work operationalises evidential proportionality and supports clearer, more transparent scientific communication.
When 20 Agents Fail to Sort: The Distributed Sorting Benchmark for Scalable Multi-Agent Systems
PDF ↗Current LLM-based multi-agent systems remain fragile under scaling, even on algorithmically trivial tasks. We introduce MAS-BENCH, a distributed-sorting benchmark that isolates coordination under explicit communication constraints: each agent observes only a local segment and must collectively produce a globally consistent order via broadcasting, peer-to-peer messaging, or a shared key-value store. Across LLM-based agents, success drops sharply as the number of agents grows, exposing persistent failures in shared state, convention alignment, and consistent termination. To mitigate these breakdowns, we propose CAMOC, a lightweight, drop-in proof-of-concept built on collaboration-aware information sharing, early global metadata exchange, and single-commit verification. CAMOC substantially improves coordination success and efficiency across backends, with the largest gains under shared-state interaction. Overall, MAS-BENCH provides a diagnostic benchmark and CAMOC offers a practical step toward more reliable large-scale LLM collaboration, highlighting a gap between individual reasoning and collective correctness.