Recent advances in Large Reasoning Models (LRMs) are reshaping reasoning-intensive reranking, where step-by-step reasoning enables multi-faceted relevance judgments. To scale LRMs, existing approaches distill chain-of-thought traces from a teacher LRM into a student reranker via Supervised Fine-Tuning (SFT) followed by Reinforcement Learning (RL). However, this pipeline is misaligned with the distillation objective: teacher traces are too dense for direct imitation, while RL exploration is too sparse on hard cases. To this end, we propose SARD (Self-Adaptive Reinforced Distillation), a tri-mode RL framework that recasts teacher reasoning from a fixed supervision target into an adaptive resource the student draws on only when its own reasoning falls short. Triggered by rollout outcomes, SARD organizes training into three modes: Free Exploration, where the student reasons independently; Guided Exploration, which supplies reusable in-context guidance abstracted from teacher traces when free rollouts fail; and Teacher-Guided Repair, where the teacher diagnoses failures and revises, re-selects, or appends guidance, falling back to teacher trajectories only as a last resort. Experiments on reasoning-intensive reranking benchmarks show SARD outperforms strong reranking and SFT-then-RL baselines, with clear gains on difficult instances and under-represented reasoning patterns.

Entity-fact unlearning in large language models (LLMs) aims to remove the binding between a target entity and a forgotten fact, producing a model that behaves like one retrained without the forgotten data while preserving neighboring knowledge and overall model utility. We frame this problem as distribution shaping: such a retrained model would yield a plausible distribution at forget queries, fluent, semantically appropriate, and reflective of residual knowledge, which we take as the appropriate target. Existing objectives provide only narrow token-level or semantically uninformative supervision, falling short of this target. We propose Counterfactual Self-Distillation (CSD), a self-distillation framework that approximates the retrained distribution by constructing a counterfactual teacher through type-consistent entity substitution. The resulting teacher is decoupled from the target identity while preserving type-level relational structure, enabling directed probability reallocation without an external oracle. Experiments on TOFU and RWKU show that CSD achieves a strong forget-utility trade-off, including the best overall forgetting on RWKU and strong performance on TOFU.

Large Language Models are increasingly deployed for decision-making, yet their adoption in high-stakes domains remains limited by miscalibrated probabilities, unfaithful explanations, and inability to incorporate expert knowledge precisely. We propose IDEA, a framework that extracts LLM decision knowledge into an interpretable parametric model over semantically meaningful factors. Through joint learning of verbal-to-numerical mappings and decision parameters via EM, correlated sampling that preserves factor dependencies, and direct parameter editing with mathematical guarantees, IDEA produces calibrated probabilities while enabling quantitative human-AI collaboration. Experiments across five datasets show IDEA with Qwen-3-32B (78.6%) outperforms DeepSeek R1 (68.1%) and GPT-5.2 (77.9%), achieving perfect factor exclusion and exact calibration—precision unattainable through prompting alone.

Although Large Language Models (LLMs) have made remarkable progress, current preference optimization methods still struggle to align directional consistency while preserving reasoning diversity. To address this limitation, we propose Directional-Groupwise Preference Optimization (DGPO), a lightweight framework that aggregates supervision signals at the group level and explicitly models direction-aware alignment through multi-candidate comparisons. DGPO organizes forward and reverse question-answer instances into structured sets and optimizes a margin-based likelihood objective that separates coherent reasoning paths from inconsistent alternatives. This groupwise formulation captures richer relative information than pairwise objectives and reinforces consistency across diverse reasoning pathways. Empirical results show that our constructed reverse data yields a 3.2% average improvement across five benchmarks, while DGPO further delivers consistent gains across multiple datasets and model families, achieving average accuracy improvements of up to 3.6%. Our code and data are available at https://github.com/Demi-deng2/DGPO.

Large Language Models (LLMs) have demonstrated remarkable capabilities in processing long contexts and generating long-form texts. However, existing benchmarks evaluate these capabilities by focusing on either long-input comprehension with short outputs or short-input creative generation with lengthy outputs. This critical gap overlooks real-world Long-In-Long-Out Generation tasks, where models must simultaneously process long-context inputs and produce long-form outputs while adhering to fine-grained constraints. We introduce DocFollow, a comprehensive benchmark featuring 2,436 expert-validated instances across 5 major domains and 20 subdomains, designed to evaluate LLMs’ ability to follow hierarchical constraints in Document Following tasks. Our benchmark employs five types of constraints with increasing granularity. To ensure reliable evaluation, we develop a rule-based assessment protocol that decomposes overall evaluation uncertainty into atomic-level classification uncertainties, employing LLMs solely as information extractors for well-established NER and RE tasks rather than subjective scoring. Our evaluation of 13 state-of-the-art LLMs reveals a consistent pattern: while models maintain reasonable performance on coarse-grained constraints, they exhibit substantial degradation on fine-grained requirements, particularly entity-level constraints. This finding suggests fundamental architectural limitations in maintaining detailed structural and semantic fidelity throughout extended generation. Our findings highlight the critical need for advancing long-context generation capabilities beyond surface-level coherence to achieve precise fine-grained control. The implementation is publicly available. Source code is available at https://github.com/leonbig/DocFollow

Language agents increasingly rely on reusable skills, but post-failure repair is often handled by opaque one-shot reflection: a model generates a skill patch without explicitly maintaining how failure explanations relate to candidate repairs or how unsuccessful retests should influence later edits. We introduce RESKILL, a structured repair framework that maintains an explicit repair state across repair rounds. Given a failed rollout, the framework links failure hypotheses to candidate skill patches, selects local repairs through coverage-based attribution, retests the edited skill set in the environment, and uses retest outcomes to guide subsequent repair updates. The language model supplies structured repair factors, while the repair procedure records them, compares local skill patches by how well they address active failure explanations, and carries unsuccessful retest outcomes into later repair rounds. We evaluate RESKILL on ALFWorld and TextCraft across three model sizes under fixed repair budgets. RESKILL obtains the strongest final success in all six benchmark–model settings, improving average final success by 3.7 percentage points over direct repair and 3.3 points over hypothesis-conditioned repair. These results suggest that explicit attribution alone is insufficient; durable improvement emerges when attribution is integrated with repair selection and persistent retest-conditioned update.

Existing work has shown that o1-level performance can be achieved with limited data distillation, but most existing methods focus on unidirectional supervised fine-tuning (SFT), overlooking the intricate interplay between diverse reasoning patterns. In this paper, we construct r1k, a high-quality reverse reasoning dataset derived by inverting 1,000 forward examples from s1k, and examine how SFT and Direct Preference Optimization (DPO) affect alignment under bidirectional reasoning objectives. SFT on r1k yields a 1.6%–6.8% accuracy improvement over s1k across evaluated benchmarks. However, naively mixing forward and reverse data during SFT weakens the directional distinction. Although DPO can partially recover this distinction, it also suppresses less preferred reasoning paths by shifting the probability mass toward irrelevant outputs. These findings suggest that mixed reasoning data introduce conflicting supervision signals, underscoring the need for robust and direction-aware alignment strategies. Our code and data are available at: https://github.com/16demi/ReasonAlign-analysis.

Entity resolution is a fundamental problem in data management that aims to identify all duplicate entries within collections of multi-attribute tuples. Most existing works focus on supervised learning, relying on large amounts of high-quality labeled data, including both positive and negative tuple pairs that are meticulously prepared. However, in reality, the manual annotation process is labor-intensive; in particular, selecting high-quality negative data for labeling is both important and challenging. In this paper, we propose an end-to-end ER solution, PUER, to address low-resource entity resolution (ER) by leveraging Large Language Models (LLMs) in a Positive-Unlabeled (PU) learning setting, where only a small number of positively labeled examples, e.g., 50, and unlabeled data are provided. Unlike directly fine-tuning LLMs in a supervised manner, we solve the entity matching task using reinforcement learning and propose a self-adaptive reward function in the process of RL. To enhance performance, we design an iterative workflow based on the co-training mechanism that fully utilizes entity blocking component to assist the entity matching. This workflow aims to improve the robustness and quality of pseudo-labels so that the performance of entity matching improves. Comprehensive experimental results on various benchmark datasets demonstrate the superiority of PUER. Full version and code are available.

The improvement of LLMs’ instruction-following capabilities depends critically on the availability of high-quality instruction-response pairs. While existing automatic data synthetic methods alleviate the burden of manual curation, they often rely heavily on either the quality of seed data or strong assumptions about the structure and content of web documents. To tackle these challenges, we propose Web Reconstruction (WebR), a fully automated framework for synthesizing high-quality instruction-tuning (IT) data directly from raw web documents with minimal assumptions. Leveraging the inherent diversity of raw web content, we conceptualize web reconstruction as an instruction-tuning data synthesis task via a novel dual-perspective paradigm—Web as Instruction and Web as Response—where each web document is designated as either the input or output role to trigger the reconstruction process. Comprehensive experiments show that datasets generated by WebR outperform state-of-the-art baselines by up to 16.65% across four instruction-following benchmarks. Notably, WebR demonstrates superior compatibility, data efficiency, and scalability, enabling enhanced domain adaptation with minimal effort.

Eliminating toxicity from Large Language Models (LLMs) is crucial for ensuring user safety. However, current methods have limitations in the analysis and utilization of toxic samples, failing to fully harness their potential. Through comparative analysis of toxic and safe samples, we discover that toxic samples exhibit diversity and, within this diversity, there lies specificity. These findings suggest that leveraging these characteristics of toxic samples could enhance the performance of algorithms in detoxifying LLMs. To this end, we propose a novel diverse detoxification framework, DivDetox, which comprises two innovative components: a Multi-Category-Induced Personalized Sample Generation (MPSG) strategy and a Scaled Contrastive DPO (SC-DPO) approach. The former is designed to elicit a variety of personalized toxic responses from the LLM, while the latter is constructed to precisely and fully utilize these toxic responses. Experiments on benchmark datasets across different model scales and different detoxification tasks verify the effectiveness of our architecture.

Keyphrase extraction (KPE) automatically extracts phrases in a document that provide a concise summary of the core content, which benefits downstream information retrieval and NLP tasks. Previous state-of-the-art methods select candidate keyphrases based on the similarity between learned representations of the candidates and the document. They suffer performance degradation on long documents due to discrepancy between sequence lengths which causes mismatch between representations of keyphrase candidates and the document. In this work, we propose a novel unsupervised embedding-based KPE approach, Masked Document Embedding Rank (MDERank), to address this problem by leveraging a mask strategy and ranking candidates by the similarity between embeddings of the source document and the masked document. We further develop a KPE-oriented BERT (KPEBERT) model by proposing a novel self-supervised contrastive learning method, which is more compatible to MDERank than vanilla BERT. Comprehensive evaluations on six KPE benchmarks demonstrate that the proposed MDERank outperforms state-of-the-art unsupervised KPE approach by average 1.80 F1@15 improvement. MDERank further benefits from KPEBERT and overall achieves average 3.53 F1@15 improvement over SIFRank.