While Chain-of-Thought (CoT) reasoning enhances code generation in Large Language Models (LLMs), it introduces a critical challenge in uncertainty estimation: Confidence Saturation. Existing calibration methods, such as Self-Consistency, rely on the assumption that consensus implies correctness. This assumption fails under systematic errors, where models confidently repeat flawed logic, leading to miscalibrated high-confidence predictions. To address this, we introduce NeuroSym-Cal, a hierarchical calibration framework. We posit that reliable confidence requires interrogating the model at two complementary levels: the extrinsic consensus of its symbolic outputs and the intrinsic sensitivity of its latent reasoning. Specifically, we propose Reasoning Sensitivity Analysis to measure the local curvature of the deductive process via latent perturbation, providing a fine-grained signal that persists even when output consensus saturates. These orthogonal features are fused by a Contextual Calibration Network to predict correctness. Experiments across state-of-the-art reasoning models (e.g., DeepSeek-R1) demonstrate that NeuroSym-Cal effectively de-saturates overconfident errors, achieving state-of-the-art Expected Calibration Error (ECE) and superior selective generation performance on Out-Of-Domain (OOD) benchmarks.

Traditional reinforcement learning from human feedback (RLHF) optimizes policies on fixed training inputs, limiting the diversity of learning signals. We propose JODP (Joint Optimization of Data and Policy), a framework where the evolving policy model generates improved variants of training problems to enhance its own learning. While training problems remain fixed, JODP optimizes how they are presented: the policy generates specification hints that guide rollout generation, then learns to reproduce the discovered high-reward behaviors without the hints. This “if you can solve it with a hint, learn to solve it without one” principle creates a co-evolutionary dynamic where better policies discover better specifications, which enable further policy improvement. JODP operates as a plug-and-play enhancement to existing algorithms: specifications are selected via UCB bandits for exploration-exploitation balance, used only during training rollouts, and discarded at deployment. Through evaluation on safety alignment tasks, we demonstrate consistent improvements with GRPO, RLOO, and REINFORCE++, allowing 4B models to approach 8B model performance using less than 1% additional computational overhead.

While LLM-based agents can interact with environments via invoking external tools, their expanded capabilities also amplify security risks. Monitoring step-level tool invocation behaviors in real time and proactively intervening before unsafe execution is critical for agent deployment, yet remains underexplored. In this work, we first construct TS-Bench, a novel benchmark for step-level tool invocation safety detection in LLM agents. We then develop a guardrail model, TS-Guard, using multi-task reinforcement learning. The model proactively detects unsafe tool invocation actions before execution by reasoning over the interaction history. It assesses request harmfulness and action–attack correlations, producing interpretable and generalizable safety judgments and feedback. Furthermore, We introduce TS-Flow, a guardrail-feedback-driven reasoning framework for LLM agents, which reduces harmful tool invocations of ReAct-style agents by 65% on average and improves benign task completion by approximately 10% under prompt injection attacks.

Current safety alignment techniques for large language models (LLMs) struggle to balance harmlessness and helpfulness: improving safety often comes at the cost of degraded utility. Our preliminary study shows that guiding unaligned base models with safety-aware reasoning that includes explicit self-reflection can effectively defend jailbreak attacks while preserving response quality. This observation motivates internalizing and strengthening self-reflective reasoning capabilities within LLMs to achieve a better safety–utility trade-off. We propose Safety-aware Reflective Reasoning Optimization (SaRO), a two-stage framework: (1) Reasoning-style Warmup (RW) to internalize self-reflective reasoning, and (2) Self-reflective Reasoning Process Optimization (SRPO) to encourage reflection and correction. Experiments show that SaRO outperforms existing reasoning-based alignment methods, achieving a better balance of safety and helpfulness.

Instruction-tuned language models increasingly rely on large multi-turn dialogue corpora, but these datasets are often noisy and structurally inconsistent, with topic drift, repetitive chitchat, and mismatched answer formats across turns. We address this from a data selection perspective and propose MDS (Multi-turn Dialogue Selection), a dialogue-level framework that scores whole conversations rather than isolated turns. MDS combines a global coverage stage that performs bin-wise selection in the user-query trajectory space to retain representative yet non-redundant dialogues, with a local structural stage that evaluates within-dialogue reliability through entity-grounded topic grounding and information progress, together with query-answer form consistency for functional alignment. MDS outperforms strong single-turn selectors, dialogue-level LLM scorers, and heuristic baselines on three multi-turn benchmarks and an in-domain Banking test set, achieving the best overall rank across reference-free and reference-based metrics, and is more robust on long conversations under the same training budget. Code and resources are included in the supplementary materials.

Monte Carlo Tree Search (MCTS) has been widely used for automated reasoning data exploration, but current supervision extraction methods remain inefficient. Standard approaches retain only the single highest-reward trajectory, discarding the comparative signals present in the many explored paths. Here we introduce Contrastive Reasoning Path Synthesis (CRPS), a framework that transforms supervision extraction from a filtering process into a synthesis procedure. CRPS uses a structured reflective process to analyze the differences between high- and low-quality search trajectories, extracting explicit information about strategic pivots and local failure modes. These insights guide the synthesis of reasoning chains that incorporate success patterns while avoiding identified pitfalls. We show empirically that models fine-tuned on just 60K CRPS-synthesized examples match or exceed the performance of baselines trained on 590K examples derived from standard rejection sampling, a 20× reduction in dataset size. Furthermore, CRPS improves generalization on out-of-domain benchmarks, demonstrating that learning from the contrast between success and failure produces more transferable reasoning capabilities than learning from success alone.

Instruction tuning relies on large instruction–response corpora whose quality and composition strongly affect downstream performance. We propose Answer Divergence-Guided Selection (ADG), which selects instruction data based on the geometric structure of multi-sample outputs. ADG draws several high-temperature generations per instruction, maps responses into an embedding space, and computes an output divergence score that jointly encodes dispersion magnitude and shape anisotropy. High scores correspond to instructions whose answers are both far apart and multi-modal, rather than clustered paraphrases along a single direction. Across two backbones and three public instruction pools, fine-tuning on only 10K ADG-selected examples consistently outperforms strong selectors on six benchmarks spanning reasoning, knowledge, and coding. Analyses further show that both dispersion magnitude and shape anisotropy are necessary, supporting answer divergence as a practical signal for instruction data selection. Code and appendix are included in the supplementary materials.

We revisit retrieval-augmented generation (RAG) by embedding retrieval control directly into generation. Instead of treating retrieval as an external intervention, we express retrieval decisions within token-level decoding, enabling end-to-end coordination without additional controllers or classifiers. Under the paradigm of Retrieval as Generation, we propose GRIP (Generation-guided Retrieval with Information Planning), a unified framework in which the model regulates retrieval behavior through control-token emission. Central to GRIP is Self-Triggered Information Planning, which allows the model to decide when to retrieve, how to reformulate queries, and when to terminate, all within a single autoregressive trajectory. This design tightly couples retrieval and reasoning and supports dynamic multi-step inference with on-the-fly evidence integration. To supervise these behaviors, we construct a structured training set covering answerable, partially answerable, and multi-hop queries, each aligned with specific token patterns. Experiments on five QA benchmarks show that GRIP surpasses strong RAG baselines and is competitive with GPT-4o while using substantially fewer parameters. Code and resources are provided in the supplementary materials.

Fine-tuning large language models (LLMs) is an effective approach to enhancing their performance on specialized downstream tasks. Among the various techniques, low-rank adaptation has garnered significant attention due to its ability to maintain the full performance of fine-tuning while enhancing computational efficiency. However, existing approaches often rely on manually specified and fixed hyperparameters to identify the trainable components within weight matrices, resulting in suboptimal performance and low parameter efficiency. This paper presents a novel Learnable Low-Rank Adaptation (LeLoRA) framework that utilizes dynamically learned fine-tuning strategies to facilitate the effective adaptation of LLMs. Our framework integrates an LLM with a policy network that automatically and adaptively generates matrix-specific adaptation strategies to identify the trainable components of each weight matrix, taking into account their unique characteristics, such as singular values and matrix norms. A reinforcement learning-based optimization algorithm is then employed to iteratively update the LLM and the policy network, ensuring that the generated strategies adapt in real time to the evolving states of the LLM. Extensive experiments have been conducted across various natural language processing and multimodal tasks. The results across ten different LLMs, ranging from 125M to 70B parameters, provide compelling evidence that LeLoRA consistently outperforms existing baselines in adapting LLMs. Moreover, analytical experiments provide valuable insights into the effectiveness of the generated strategies.

Retrieval-augmented language models (RALMs) have recently shown great potential in mitigating the limitations of implicit knowledge in LLMs, such as untimely updating of the latest expertise and unreliable retention of long-tail knowledge. However, since the external knowledge base, as well as the retriever, can not guarantee reliability, potentially leading to the knowledge retrieved not being helpful or even misleading for LLM generation. In this paper, we introduce Supportiveness-based Knowledge Rewriting (SKR), a robust and pluggable knowledge rewriter inherently optimized for LLM generation. Specifically, we introduce the novel concept of “supportiveness”—which represents how effectively a knowledge piece facilitates downstream tasks. Based on supportiveness, we first design a training data curation strategy for our rewriter model, effectively identifying and filtering out poor or irrelevant rewrites to improve data efficacy. We then introduce the direct preference optimization (DPO) algorithm to align the generated rewrites to optimal supportiveness, guiding the rewriter model to summarize augmented content that better improves the final response. Comprehensive evaluations across six popular knowledge-intensive tasks and four LLMs have demonstrated the effectiveness and superiority of SKR. With only 7B parameters, SKR has shown better knowledge rewriting capability over GPT-4.

Existing pretraining data mixing methods for large language models (LLMs) typically follow a domain-wise methodology, a top-down process that first determines domain weights and then performs uniform data sampling across each domain. However, these approaches neglect significant inter-domain overlaps and commonalities, failing to control the global diversity of the constructed training dataset. Further, uniform sampling within domains ignores fine-grained sample-specific features, potentially leading to suboptimal data distribution. To address these shortcomings, we propose a novel sample-wise data mixture approach based on a bottom-up paradigm. This method performs global cross-domain sampling by systematically evaluating the quality and diversity of each sample, thereby dynamically determining the optimal domain distribution. Comprehensive experiments across multiple downstream tasks and perplexity assessments demonstrate that SampleMix surpasses existing domain-based methods. Meanwhile, SampleMix requires 1.4x to 2.1x fewer training steps to achieve the baselines’ performance, highlighting the substantial potential of SampleMix to optimize pre-training data.

Recent research in information extraction (IE) focuses on utilizing code-style inputs to enhance structured output generation. The intuition behind this is that the programming languages (PLs) inherently exhibit greater structural organization than natural languages (NLs). This structural advantage makes PLs particularly suited for IE tasks. Nevertheless, existing research primarily focuses on Python for code-style simulation, overlooking the potential of other widely-used PLs (e.g., C++ and Java) during the supervised fine-tuning (SFT) phase. In this research, we propose Multiple Programming Languages with large language models for information extraction (abbreviated as MPL), a novel framework that explores the potential of incorporating different PLs in the SFT phase. Additionally, we introduce function-prompt with virtual running to simulate code-style inputs more effectively and efficiently. Experimental results on a wide range of datasets demonstrate the effectiveness of MPL. Furthermore, we conduct extensive experiments to provide a comprehensive analysis. Our code and additional files are in the supplementary materials.

Code security and usability are both essential for various coding assistant applications driven by large language models (LLMs). Current code security benchmarks focus solely on single evaluation task and paradigm, such as code completion and generation, lacking comprehensive assessment across dimensions like secure code generation, vulnerability repair and discrimination. In this paper, we first propose CoV-Eval, a multi-task benchmark covering various tasks such as code completion, vulnerability repair, vulnerability detection and classification, for comprehensive evaluation of LLM code security. Besides, we developed VC-Judge, an improved judgment model that aligns closely with human experts and can review LLM-generated programs for vulnerabilities in a more efficient and reliable way. We conduct a comprehensive evaluation of 20 proprietary and open-source LLMs. Overall, while most LLMs identify vulnerable codes well, they still tend to generate insecure codes and struggle with recognizing specific vulnerability types and performing repairs. Extensive experiments and qualitative analyses reveal key challenges and optimization directions, offering insights for future research in LLM code security.

Aligning large language models (LLMs) with human values and preferences is a significant challenge. Training-based methods, such as reinforcement learning from human feedback (RLHF) and direct preference optimization (DPO), require substantial resources and are impractical for API-based LLMs. Post-processing methods decouple alignment from training but may incur high multiple-time inference costs or rely on less knowledgeable lightweight models for response refinement. In this paper, we propose a new LLM alignment paradigm from the perspective of pre-processing. By reformulating risky queries into highly relevant yet harmless ones before feeding them into LLMs, our method eliminates the high costs of training base LLMs, efficiently applies to both open-source and proprietary LLMs, and achieves a promising balance of harmlessness and helpfulness. For example, with Vicuna-7B as the LLM to align, it enhances helpfulness by 28.52% over DPO while maintaining comparable harmlessness levels. When applied to Gemini-1.5-pro, it increased harmlessness and helpfulness by 7.04% and 29.37%, respectively.

Chinese Spelling Correction (CSC) commonly lacks large-scale high-quality corpora, due to the labor-intensive labeling of spelling errors in real-life human writing or typing scenarios. Two data augmentation methods are widely adopted: (1) Random Replacement with the guidance of confusion sets and (2) OCR/ASR-based Generation that simulates character misusing. However, both methods inevitably introduce noisy data (e.g., false spelling errors), potentially leading to over-correction. By carefully analyzing the two types of corpora, we find that though the latter achieves more robust generalization performance, the former yields better-calibrated CSC models. We then provide a theoretical analysis of this empirical observation, based on which a corpus refining strategy is proposed. Specifically, OCR/ASR-based data samples are fed into a well-calibrated CSC model trained on random replacement-based corpora and then filtered based on prediction confidence. By learning a simple BERT-based model on the refined OCR/ASR-based corpus, we set up impressive state-of-the-art performance on three widely-used benchmarks, while significantly alleviating over-correction (e.g., lowering false positive predictions).

Although large language models (LLMs) have demonstrated impressive few-shot learning capabilities via in-context learning (ICL), ICL performance is known to be highly sensitive to the order of examples provided. To identify appropriate orders, recent studies propose heuristic methods to evaluate order performance using a set of unlabeled data. However, the requirement of in-domain data limits their utility in real-world scenarios where additional annotated data is challenging to acquire. Additionally, these dataset-based approaches are prone to being sub-optimal for a lack of consideration for individual differences. To address the problems, we first analyze the properties of performant example orders at both corpus level and instance level. Based on the analysis we propose DEmO to adaptively identify performant example order for each instance without extra data. DEmO works by filtering out a subset of orders featuring label fairness, then selecting the most influential order for each test instance. The employment of a content-free metric makes DEmO independent of in-domain data. Extensive experiments indicate the superiority of DEmO over a wide range of strong baselines. Further analysis validates the generalizability across various settings.

Large Language Models (LLMs) demonstrate human-level capabilities in dialogue, reasoning, and knowledge retention. However, even the most advanced LLMs face challenges such as hallucinations and real-time updating of their knowledge. Current research addresses this bottleneck by equipping LLMs with external knowledge, a technique known as Retrieval Augmented Generation (RAG). However, two key issues constrained the development of RAG. First, there is a growing lack of comprehensive and fair comparisons between novel RAG algorithms. Second, open-source tools such as LlamaIndex and LangChain employ high-level abstractions, which results in a lack of transparency and limits the ability to develop novel algorithms and evaluation metrics. To close this gap, we introduce RAGLAB, a modular and research-oriented open-source library. RAGLAB reproduces 6 existing algorithms and provides a comprehensive ecosystem for investigating RAG algorithms. Leveraging RAGLAB, we conduct a fair comparison of 6 RAG algorithms across 10 benchmarks. With RAGLAB, researchers can efficiently compare the performance of various algorithms and develop novel algorithms.

The rapid growth of evaluation methodologies and datasets for large language models (LLMs) has created a pressing need for their unified integration. Meanwhile, concerns about data contamination and bias compromise the trustworthiness of evaluation findings, while the efficiency of evaluation processes remains a bottleneck due to the significant computational costs associated with LLM inference.In response to these challenges, we introduce FreeEval, a modular framework not only for conducting trustworthy and efficient automatic evaluations of LLMs but also serving as a platform to develop and validate new evaluation methodologies. FreeEval addresses key challenges through: (1) unified abstractions that simplify the integration of diverse evaluation methods, including dynamic evaluations requiring complex LLM interactions; (2) built-in meta-evaluation techniques such as data contamination detection and human evaluation to enhance result fairness; (3) a high-performance infrastructure with distributed computation and caching strategies for efficient large-scale evaluations; and (4) an interactive Visualizer for result analysis and interpretation to support innovation of evaluation techniques. We open-source all our code at https://github.com/WisdomShell/FreeEval and our demostration video, live demo, installation guides are available at: https://freeeval.zhuohao.me/.