Searcharxiv⌕ Search

arXiv subjects

Haibo Shi

Publications and source records attributed to Haibo Shi.

At least 19 recordsLinked to original sources
arXiv2026

Beyond Polarization: The Generative Constraint of Chain-of-Thought in Pointwise Reranking

In pointwise document reranking, Chain-of-Thought models typically underperform direct scoring models. While existing diagnostics attribute this to inferior classification, score polarization, or calibration breakdown, whether targeted training can bridge this gap remains unclear. Our empirical study first confirms that this gap is stable across scales up to 32B parameters, ruling out model and data capacity confounders. We then apply stress tests utilizing reinforcement learning, fine-grained supervision, and architectural decoupling to explicitly repair these deviations. Although these interventions improve classification accuracy and absolute scores, the relative ranking gap persists. These findings suggest that, within the pointwise scoring paradigm, routing continuous relevance semantics through discrete text constrains ranking signal resolution, revealing a bottleneck that is stable and difficult to overcome under current standard methods, rather than an easily resolvable training bias.

cs.CL↗
arXiv2026

ForestBench: A Unified Graph Framework for Evaluating Multi-Agent Collaboration

Multi-agent systems (MAS) built on Large Language Models (LLMs) are proliferating rapidly, but their heterogeneous execution traces provide no common basis for evaluation across methods. Outcome-only benchmarks discard collaborations, whereas LLM-as-Judge evaluation requires additional, model-dependent inference and can vary with the LLM and rubric. We introduce a generalizable evaluation framework that maps native MAS traces into a shared space of unified collaboration graphs, enabling different methods to be evaluated under the same representation, reference set, and metric panel. Candidate graphs are compared with a query-specific reference forest. Each forest is a benchmark-provided collection of verified-success graphs: it records diverse ways in which representative MAS methods can complete the task, rather than prescribing a unique optimal process. Instantiating the framework as ForestBench, we filter $844$ collaboration-necessary queries from seven public datasets, precompute ten successful target-conditioned reference graphs per query, and evaluate six representative MAS frameworks. Controlled backbone, reference-construction, and perturbation studies test the stability and scope of evaluation. Once the benchmark forests are built, ForestBench scores a trace in milliseconds without further LLM inference, providing a reusable structural basis for comparing diverse MAS collaboration traces.

cs.AI↗
arXiv2026

Training Documents Reranker with Search Rubrics for Deep Research Agent

Retrieval systems help deep research agents generate high-quality answers by providing relevant documents. However, existing retrievers typically select documents through relevance matching, while individually well-matched top-$k$ documents may not form a \textit{set} that satisfies the complex information needs of an agent query (\eg, diverse, concise and authoritative documents). In this paper, we propose search-oriented rubrics that \textit{explicitly} define the requirements that high-quality document sets should satisfy for each agent query. Our search rubrics are organized into a hierarchical structure and synthesized using a powerful LLM. Based on these search rubrics, we further train a document reranker \textbf{RubricRanker} to select a high-quality subset from retrieved documents. We design a two-stage training framework that consists of rubrics-guided supervised fine-tuning and rubric-based reinforcement learning. Extensive experiments demonstrate that RubricRanker outperforms the strongest baseline by 2.6 points on four deep research benchmarks and generalizes well to five RAG benchmarks.

cs.IR↗
arXiv2026

Search-G1: Grounded Search Agents via Representation-Based Intrinsic Rewards

Search-augmented language agents should retrieve external information only when necessary and ground their answers in retrieved evidence. Existing external rewards provide either sparse outcome supervision or richer feedback from process annotations and LLM judges. Outcome rewards scale readily but cannot distinguish grounded retrieval from redundant search, whereas richer signals require costly annotation or inference during training. Internal rewards based on policy-side signals such as entropy, likelihood, or information gain are graded and inexpensive to evaluate, yet mainly reflect model confidence rather than evidence grounding. We propose Search-G1, a representation-based intrinsic reward framework that measures the operational grounding of an agent's answers through two intervention-calibrated readouts. A prompt-state readout predicts closed-book sufficiency, whose complement defines policy-relative retrieval necessity; an answer-commit readout estimates evidence reliance from answer-stage sensitivity to evidence deletion. Together, they provide additional credit to correct searched trajectories when retrieval is estimated necessary and the answer is evidence-sensitive, favor correct direct answers when closed-book knowledge suffices, and penalize repeated search. After calibration, reward scoring requires neither process annotations nor LLM-as-judge inference during policy optimization. Because reinforcement learning changes policy representations, Search-G1 periodically refits both readouts on trajectories from the latest checkpoint, allowing the reward to co-evolve with the policy. Experiments across multiple search-based question-answering benchmarks and two model scales show that Search-G1 improves the grounding--search-cost trade-off, producing shorter response-side trajectories at competitive task accuracy. Code is available at https://github.com/Rosy0912/Search-G1.

cs.CL↗
arXiv2026

OPOD: On-Policy Omni Distillation

Omni-modal models provide a unified interface for text, images, and audio. However, improving these abilities together remains difficult, as post-training on pooled multimodal data often fails to preserve the strengths of modality teachers. On-policy distillation (OPD) has recently become popular in model post-training. It samples responses from the current student and compares the teacher's and student's next-token distributions along those responses, yielding dense supervision while reducing the mismatch between training and inference. Despite these advantages, standard OPD does not readily extend to several modality teachers. Their guidance may favor conflicting changes to the shared model, while matching each teacher's next-token distribution can prevent the student from moving beyond that teacher. To address these challenges, we propose On-Policy Omni Distillation (OPOD), which consolidates text, image, and audio teachers into one omni model. OPOD routes each response to the corresponding teacher, controls the teachers independently, and applies guidance only when the teacher assigns a higher probability to the generated token. The selected teacher also evaluates answer confidence and whether the reasoning increases support for the answer. Extensive experiments on twelve benchmarks show that OPOD achieves the best average at three model scales, reaching 70.8, 51.7, and 46.2 and outperforming the strongest comparator by 2.1, 1.8, and 1.7 points. At 30B, it surpasses the base model and pooled RL training on all twelve benchmarks, and ranks first or second on eleven even when the teachers are included. Only the student is retained for deployment.

cs.AI↗
arXiv2026

Beyond Relevance-Centric Retrieval: Rubric-Oriented Document Set Selection and Ranking

As large language models and AI agents become the primary consumers of search results, document set quality determines the upper bound of downstream generation. Yet existing evaluation systems remain confined to scoring documents independently and aggregating via nDCG, ignoring inter-document interactions (redundancy, conflict, complementarity) and unable to answer what makes one document set better than another. To address these issues, we propose a complete evaluate-diagnose-optimize framework. We design SetwiseEvalKit, a three-level, nine-dimension document set evaluation benchmark covering both short-form and long-form scenarios, comprising approximately 28K high-quality evaluation rubrics. We systematically evaluate 12 rerankers: even the best method achieves no more than 45% coverage, cross-document coordination dimensions are universally weak, and no single method maintains top performance across both settings. Building on this, we propose Rubric4Setwise, a training-free method that converts rubric-based evaluation criteria into document set selection signals, achieving the best downstream generation performance with fewer documents and search rounds. It is the only method that maintains state-of-the-art results across both scenarios, validating the effectiveness of closing the loop from evaluation to optimization.

cs.CL↗
arXiv2026

Parameter Importance is Not Static: Evolving Parameter Isolation for Supervised Fine-Tuning

Supervised Fine-Tuning (SFT) of large language models often suffers from task interference and catastrophic forgetting. Recent approaches alleviate this issue by isolating task-critical parameters during training. However, these methods represent a static solution to a dynamic problem, assuming that parameter importance remains fixed once identified. In this work, we empirically demonstrate that parameter importance exhibits temporal drift over the course of training. To address this, we propose Evolving Parameter Isolation (EPI), a fine-tuning framework that adapts isolation decisions based on online estimates of parameter importance. Instead of freezing a fixed subset of parameters, EPI periodically updates isolation masks using gradient-based signals, enabling the model to protect emerging task-critical parameters while releasing outdated ones to recover plasticity. Experiments on diverse multi-task benchmarks demonstrate that EPI consistently reduces interference and forgetting compared to static isolation and standard fine-tuning, while improving overall generalization. Our analysis highlights the necessity of synchronizing isolation mechanisms with the evolving dynamics of learning diverse abilities.

cs.LG↗
arXiv2026

Reason Only When Needed: Efficient Generative Reward Modeling via Model-Internal Uncertainty

Recent advancements in the Generative Reward Model (GRM) have demonstrated its potential to enhance the reasoning abilities of LLMs through Chain-of-Thought (CoT) prompting. Despite these gains, existing implementations of GRM suffer from two critical limitations. First, CoT prompting is applied indiscriminately to all inputs regardless of their inherent complexity. This introduces unnecessary computational costs for tasks amenable to fast, direct inference. Second, existing approaches primarily rely on voting-based mechanisms to evaluate CoT outputs, which often lack granularity and precision in assessing reasoning quality. In this paper, we propose E-GRM, an efficient generative reward modeling framework grounded in model-internal uncertainty. E-GRM leverages the convergence behavior of parallel model generations to estimate uncertainty and selectively trigger CoT reasoning only when needed, without relying on handcrafted features or task-dependent signals. To improve reward fidelity, we introduce a lightweight discriminative scorer trained with a hybrid regression--ranking objective to provide fine-grained evaluation of reasoning paths. Experiments on multiple reasoning benchmarks show that E-GRM substantially reduces inference cost while consistently improving answer accuracy, demonstrating that model-internal uncertainty is an effective and general signal for efficient reasoning-aware reward modeling.

cs.CL↗
arXiv2026

Why Supervised Fine-Tuning Fails to Learn: A Systematic Study of Incomplete Learning in Large Language Models

Supervised Fine-Tuning (SFT) is the standard approach for adapting large language models (LLMs) to downstream tasks. However, we observe a persistent failure mode: even after convergence, models often fail to correctly reproduce a subset of their own supervised training data. We refer to this behavior as the Incomplete Learning Phenomenon(ILP). This paper presents the first systematic study of ILP in LLM fine-tuning. We formalize ILP as post-training failure to internalize supervised instances and demonstrate its prevalence across multiple model families, domains, and datasets. Through controlled analyses, we identify five recurrent sources of incomplete learning: (1) missing prerequisite knowledge in the pre-trained model, (2) conflicts between SFT supervision and pre-training knowledge, (3) internal inconsistencies within SFT data, (4) left-side forgetting during sequential fine-tuning, and (5) insufficient optimization for rare or complex patterns. We introduce a diagnostic-first framework that maps unlearned samples to these causes using observable training and inference signals, and study several targeted mitigation strategies as causal interventions. Experiments on Qwen, LLaMA, and OLMo2 show that incomplete learning is widespread and heterogeneous, and that improvements in aggregate metrics can mask persistent unlearned subsets. The findings highlight the need for fine-grained diagnosis of what supervised fine-tuning fails to learn, and why.

cs.CL↗
arXiv2026

When Safety Becomes a Vulnerability: Exploiting LLM Alignment Homogeneity for Transferable Blocking in RAG

Retrieval-Augmented Generation (RAG) systems are vulnerable to blocking attacks, in which poisoned documents cause large language models (LLMs) to refuse benign queries. Existing attacks rely on adversarial suffixes or explicit instructions, which are increasingly ineffective against modern LLMs, susceptible to prompt injection filtering, or require feedback from the target system. We observe overlapping risk categories and refusal criteria across safety-aligned LLMs, a phenomenon we term alignment homogeneity. This shared attack surface makes refusal-inducing context transferable across models. Accordingly, we propose TabooRAG, which optimizes one document per query for retrieval and refusal induction in a surrogate RAG environment, then transfers it to an unknown target system. Rather than injecting instructions, TabooRAG constructs query-relevant risk context to trigger alignment-driven refusal. To reduce optimization cost, it reuses validated strategies through a query-aware strategy library. Across nine LLMs and three datasets, TabooRAG achieves state-of-the-art ASR after filtering, with a 67.3% relative gain over the average per-setting best baseline. Further experiments show that TabooRAG remains effective with diverse surrogate models, against unseen target models, and under stronger RAG pipelines and existing defenses.

cs.CR↗
arXiv2025

FinEval-KR: A Financial Domain Evaluation Framework for Large Language Models' Knowledge and Reasoning

Large Language Models (LLMs) demonstrate significant potential but face challenges in complex financial reasoning tasks requiring both domain knowledge and sophisticated reasoning. Current evaluation benchmarks often fall short by not decoupling these capabilities indicators from single task performance and lack root cause analysis for task failure. To address this, we introduce FinEval-KR, a novel evaluation framework for decoupling and quantifying LLMs' knowledge and reasoning abilities independently, proposing distinct knowledge score and reasoning score metrics. Inspired by cognitive science, we further propose a cognitive score based on Bloom's taxonomy to analyze capabilities in reasoning tasks across different cognitive levels. We also release a new open-source Chinese financial reasoning dataset covering 22 subfields to support reproducible research and further advancements in financial reasoning. Our experimental results reveal that LLM reasoning ability and higher-order cognitive ability are the core factors influencing reasoning accuracy. We also specifically find that even top models still face a bottleneck with knowledge application. Furthermore, our analysis shows that specialized financial LLMs generally lag behind the top general large models across multiple metrics.

cs.CL↗
arXiv2025

A Dual-Task Synergy-Driven Generalization Framework for Pancreatic Cancer Segmentation in CT Scans

Pancreatic cancer, characterized by its notable prevalence and mortality rates, demands accurate lesion delineation for effective diagnosis and therapeutic interventions. The generalizability of extant methods is frequently compromised due to the pronounced variability in imaging and the heterogeneous characteristics of pancreatic lesions, which may mimic normal tissues and exhibit significant inter-patient variability. Thus, we propose a generalization framework that synergizes pixel-level classification and regression tasks, to accurately delineate lesions and improve model stability. This framework not only seeks to align segmentation contours with actual lesions but also uses regression to elucidate spatial relationships between diseased and normal tissues, thereby improving tumor localization and morphological characterization. Enhanced by the reciprocal transformation of task outputs, our approach integrates additional regression supervision within the segmentation context, bolstering the model's generalization ability from a dual-task perspective. Besides, dual self-supervised learning in feature spaces and output spaces augments the model's representational capability and stability across different imaging views. Experiments on 594 samples composed of three datasets with significant imaging differences demonstrate that our generalized pancreas segmentation results comparable to mainstream in-domain validation performance (Dice: 84.07%). More importantly, it successfully improves the results of the highly challenging cross-lesion generalized pancreatic cancer segmentation task by 9.51%. Thus, our model constitutes a resilient and efficient foundational technological support for pancreatic disease management and wider medical applications. The codes will be released at https://github.com/SJTUBME-QianLab/Dual-Task-Seg.

eess.IV↗
arXiv2024

HG2P: Hippocampus-inspired High-reward Graph and Model-Free Q-Gradient Penalty for Path Planning and Motion Control

Goal-conditioned hierarchical reinforcement learning (HRL) decomposes complex reaching tasks into a sequence of simple subgoal-conditioned tasks, showing significant promise for addressing long-horizon planning in large-scale environments. This paper bridges the goal-conditioned HRL based on graph-based planning to brain mechanisms, proposing a hippocampus-striatum-like dual-controller hypothesis. Inspired by the brain mechanisms of organisms (i.e., the high-reward preferences observed in hippocampal replay) and instance-based theory, we propose a high-return sampling strategy for constructing memory graphs, improving sample efficiency. Additionally, we derive a model-free lower-level Q-function gradient penalty to resolve the model dependency issues present in prior work, improving the generalization of Lipschitz constraints in applications. Finally, we integrate these two extensions, High-reward Graph and model-free Gradient Penalty (HG2P), into the state-of-the-art framework ACLG, proposing a novel goal-conditioned HRL framework, HG2P+ACLG. Experimentally, the results demonstrate that our method outperforms state-of-the-art goal-conditioned HRL algorithms on a variety of long-horizon navigation tasks and robotic manipulation tasks.

cs.LG↗
arXiv2024

ATM: Adversarial Tuning Multi-agent System Makes a Robust Retrieval-Augmented Generator

Large language models (LLMs) are proven to benefit a lot from retrieval-augmented generation (RAG) in alleviating hallucinations confronted with knowledge-intensive questions. RAG adopts information retrieval techniques to inject external knowledge from semantic-relevant documents as input contexts. However, since today's Internet is flooded with numerous noisy and fabricating content, it is inevitable that RAG systems are vulnerable to these noises and prone to respond incorrectly. To this end, we propose to optimize the retrieval-augmented Generator with an Adversarial Tuning Multi-agent system (ATM). The ATM steers the Generator to have a robust perspective of useful documents for question answering with the help of an auxiliary Attacker agent through adversarially tuning the agents for several iterations. After rounds of multi-agent iterative tuning, the Generator can eventually better discriminate useful documents amongst fabrications. The experimental results verify the effectiveness of ATM and we also observe that the Generator can achieve better performance compared to the state-of-the-art baselines.

cs.CL↗
arXiv2024

GOVERN: Gradient Orientation Vote Ensemble for Multi-Teacher Reinforced Distillation

Pre-trained language models have become an integral component of question-answering systems, achieving remarkable performance. However, for practical deployment, it is crucial to perform knowledge distillation to maintain high performance while operating under computational constraints. In this paper, we address a key question: given the importance of unsupervised distillation for student model performance, how can knowledge from multiple teacher models be effectively ensemble during this stage without the guidance of labels? We propose a novel algorithm, GOVERN, to tackle this issue. GOVERN has demonstrated significant improvements in both offline and online experiments, enabling the student model to achieve results comparable to that of teacher ensembles. Our experiments show that GOVERN remarkably requires a mere 1\% of the ensemble method's inference budget to achieve 99.5\% of performance. The proposed algorithm has been successfully deployed in a real-world commercial question-answering system, demonstrating its real-world applicability.

cs.CL↗
arXiv2024

The Real, the Better: Aligning Large Language Models with Online Human Behaviors

Large language model alignment is widely used and studied to avoid LLM producing unhelpful and harmful responses. However, the lengthy training process and predefined preference bias hinder adaptation to online diverse human preferences. To this end, this paper proposes an alignment framework, called Reinforcement Learning with Human Behavior (RLHB), to align LLMs by directly leveraging real online human behaviors. By taking the generative adversarial framework, the generator is trained to respond following expected human behavior; while the discriminator tries to verify whether the triplets of query, response, and human behavior come from real online environments. Behavior modeling in natural-language form and the multi-model joint training mechanism enable an active and sustainable online alignment. Experimental results confirm the effectiveness of our proposed methods by both human and automatic evaluations.

cs.CL↗
arXiv2024

Learning to Use Tools via Cooperative and Interactive Agents

Tool learning empowers large language models (LLMs) as agents to use external tools and extend their utility. Existing methods employ one single LLM-based agent to iteratively select and execute tools, thereafter incorporating execution results into the next action prediction. Despite their progress, these methods suffer from performance degradation when addressing practical tasks due to: (1) the pre-defined pipeline with restricted flexibility to calibrate incorrect actions, and (2) the struggle to adapt a general LLM-based agent to perform a variety of specialized actions. To mitigate these problems, we propose ConAgents, a Cooperative and interactive Agents framework, which coordinates three specialized agents for tool selection, tool execution, and action calibration separately. ConAgents introduces two communication protocols to enable the flexible cooperation of agents. To effectively generalize the ConAgents into open-source models, we also propose specialized action distillation, enhancing their ability to perform specialized actions in our framework. Our extensive experiments on three datasets show that the LLMs, when equipped with the ConAgents, outperform baselines with substantial improvement (i.e., up to 14% higher success rate).

cs.CL↗
arXiv2024

KnowTuning: Knowledge-aware Fine-tuning for Large Language Models

Despite their success at many natural language processing (NLP) tasks, large language models still struggle to effectively leverage knowledge for knowledge-intensive tasks, manifesting limitations such as generating incomplete, non-factual, or illogical answers. These limitations stem from inadequate knowledge awareness of LLMs during vanilla fine-tuning. To address these problems, we propose a knowledge-aware fine-tuning (KnowTuning) method to improve fine-grained and coarse-grained knowledge awareness of LLMs. We devise a fine-grained knowledge augmentation stage to train LLMs to identify difficult fine-grained knowledge in answers. We also propose a coarse-grained knowledge comparison stage to train LLMs to distinguish between reliable and unreliable knowledge, in three aspects: completeness, factuality, and logicality. Extensive experiments on both generic and medical question answering (QA) datasets confirm the effectiveness of KnowTuning, through automatic and human evaluations, across various sizes of LLMs. We further verify that KnowTuning generates more facts with less factual error rate under fine-grained facts evaluation.

cs.CL↗