Existing RL methods for language models typically use token-level entropy, which captures uncertainty at the current position rather than over future generation. We study future entropy, i.e., uncertainty over the remaining suffix conditioned on the current prefix. By the entropy chain rule, local entropy is the expected decrease in future entropy, while the realized decrease depends on the chosen token, yielding an action-dependent future-entropy gap. We evaluate this signal at inference time with MIN-F-MC, an oracle-style probe that selects the candidate token with the lowest estimated future-entropy rate at high-entropy positions. On mathematical reasoning tasks, MIN-F-MC yields higher mean accuracy than greedy and Random TopK across the tested checkpoints and datasets. We then propose FEPO (Future Entropy Policy Optimization), which uses realized future-entropy rates at a sparse set of high-entropy positions to scale update magnitudes. Across the tested model–benchmark pairs, FEPO improves Mean@32 over GRPO.

Document parsing from scanned images into structured formats remains a significant challenge due to its complexly intertwined elements such as text paragraphs, figures, formulas, and tables. Existing supervised fine-tuning methods often struggle to generalize across diverse document types, leading to poor performance, particularly on out-of-distribution data. This issue is further exacerbated by the limited availability of high-quality training data for layout-aware parsing tasks. To address these challenges, we introduce layoutRL, a reinforcement learning framework that optimizes layout understanding through composite rewards integrating normalized edit distance, paragraph count accuracy, and reading order preservation. To support this training, we construct the Infinity-Doc-400K dataset, which we use to train Infinity-Parser, a vision-language model demonstrating robust generalization across various domains. Extensive evaluations on benchmarks including OmniDocBench, olmOCR-Bench, PubTabNet, and FinTabNet show that Infinity-Parser consistently achieves state-of-the-art performance across a broad range of document types, languages, and structural complexities, substantially outperforming both specialized document parsing systems and general-purpose vision-language models. We will release our code, dataset, and model to facilitate reproducible research in document parsing.

Although LLM-based agents are proven to master tool orchestration in scientific fields, particularly chemistry, their single-task performance remains limited by underlying tool constraints. To this end, we propose tool amplification, a novel paradigm that enhances the collective capabilities of specialized tools through optimized, dynamic coordination within individual tasks. Instantiating this paradigm, we introduce ChemAmp, a computationally lightweight framework that dynamically treats chemistry tools (e.g., UniMol2, Chemformer) as composable building-block agents. It constructs task-specialized super-agents that transcend atomic tool constraints with limited data (≤10 samples). Our evaluations across four core chemistry tasks molecular design, molecule captioning, reaction prediction, and property prediction demonstrate that ChemAmp outperforms chemistry-specialized models, generalist LLMs, and agent systems with tool orchestration. Critically, this bottom-up construction strategy enables 94% inference token cost reductions versus vanilla multi-agent systems.

Large language model (LLM) agents have demonstrated strong capability in sequential decision-making, yet they remain fundamentally reactive in long-horizon tasks. Unlike humans who employ "what-if" reasoning to evaluate potential plans before commitment, standard agents lack an internal world model to simulate future outcomes. Therefore, we propose to internalize future-aware planning by training a single autoregressive model to verbalize both a prospective state rollout and a plan-conditioned success estimate—a textual analogue of the Q-value. Crucially, we identify a format-capability gap: simply fine-tuning agents on look-ahead traces during post-training leads to superficial mimicry of foresight without genuine predictive grounding. To bridge this gap, we introduce a three-stage training paradigm: (i) World Model Agentic Mid-Training (WM-AMT) to inject latent predictive capabilities into the policy; (ii) Format-Eliciting SFT (FE-SFT) to structure this injected capability; and (iii) Foresight-Conditioned Reinforcement Learning (FC-RL) to refine the calibration and utility of the generated simulations. Evaluated on search and mathematical reasoning tasks, our approach consistently outperforms other training baselines. Our results demonstrate that effective internal world modeling in LLM agents requires a capability-first training pipeline to achieve grounded and calibrated foresight.

Group-Relative Policy Optimization (GRPO) is widely used to post-train LLM agents, but its trajectory-level advantage is ill-suited to long-horizon credit assignment. Reward-homogeneous rollout groups yield zero advantage; otherwise, one scalar uniformly scales every token, obscuring which decisions determine the outcome. We propose **TRACE-GRPO** (**T**eacher-**R**efined **A**dvantage **C**redit **E**stimation for GRPO), which converts teacher-LLM feedback into hierarchical credit. Teacher scores provide a sample-level fallback for zero-variance groups, reweight trajectory advantage at the turn level, and redistribute turn credit at the token level using the policy’s log-probability shift under a teacher critique. This yields dense training-time credit without token-level annotations or learned critics. We also introduce **CAPA** (Critique-Aware Paged Attention), which reuses trajectory KV caches and reduces the isolated critique-conditioned forward from roughly 2× a vanilla GRPO step to approximately 5% of a training step. On ALFWorld and ScienceWorld with Qwen2.5-7B-Instruct, TRACE-GRPO achieves 96.3% and 73.4% success, respectively, outperforming strong agent RL baselines. Ablations show cumulative gains from the three credit levels, and comparison with supervised fine-tuning on teacher-generated successes confirms that critique-guided policy optimization provides benefits beyond imitation. Code is available at [https://github.com/Lau-Jonathan/TRACE-GRPO](https://github.com/Lau-Jonathan/TRACE-GRPO).

Recent advances in model distillation show that data from advanced reasoning models can effectively train smaller student models. However, standard practices discard incorrect reasoning traces—valuable, yet underutilized data. This paper addresses the critical question: How can both positive and negative distilled reasoning traces be effectively leveraged to maximize LLM reasoning performance in an offline setting? We employ a two-stage training recipe: first, Supervised Fine-Tuning (SFT) on positive traces, followed by a refinement stage using both positive and negative traces. We find that a simple REINFORCE-style objective, which we term the Reinforcement Distillation (REDI) objective, outperforms established preference optimization methods like DPO and SimPO in this distillation context. Our empirical evaluations demonstrate the effectiveness of this approach. Notably, our Qwen-REDI-1.5B model, trained on just 131k traces from the open Open-R1 dataset, achieves an 83.1% score on MATH-500. Its performance matches that of DeepSeek-R1-Distill-Qwen-1.5B, a model trained on 800k proprietary data. This result showcases the remarkable data efficiency of utilizing previously discarded negative traces.

Code LLMs have been widely used in various domains, including code generation, logical reasoning, and agent systems. However, open-access code LLMs mostly only release weights, lacking key features such as reproducible data pipelines and transparent training protocols, which are crucial for advancing deeper, more reliable investigations. To address the gap, we introduce OpenCoder, a top-tier code LLM that not only achieves performance comparable to leading models but also serves as an “open cookbook” for the research community. Unlike most prior efforts, we release not only model weights and inference code, but also the reproducible training data, complete data processing pipeline, rigorous experimental ablation results, and detailed training protocols for open scientific research. Our work identifies the key ingredients for building a top-tier code LLM: optimized heuristic rules for data cleaning and deduplication, effective recall of code-related text corpus, and high-quality synthetic data for both annealing and supervised fine-tuning stages. By offering this level of openness, we aim to broaden access to all aspects of a top-tier code LLM, with OpenCoder serving as both a powerful model and an open foundation to accelerate research and enable reproducible advancements in code intelligence. The released resource is available at https://opencoder-llm.github.io.

With the growing deployment of pre-trained models like Transformers on cloud platforms, privacy concerns about model parameters and inference data are intensifying. Existing Privacy-Preserving Transformer Inference (PPTI) frameworks face the “impossible trinity” of balancing privacy, efficiency, and performance: Secure Multi-Party Computation (SMPC)-based approaches ensure strong privacy but suffer from high computational overhead and performance losses; Conversely, permutation-based methods achieve near-plaintext efficiency and accuracy but compromise privacy by exposing sensitive model parameters and intermediate results. Bridging this gap with a single approach presents substantial challenges, motivating the introduction of CENTAUR, a groundbreaking PPTI framework that seamlessly integrates random permutations and SMPC to address the “impossible trinity”. By designing efficient PPTI algorithms tailored to the structural properties of Transformer models, CENTAUR achieves an unprecedented balance among privacy, efficiency, and performance. Our experiments demonstrate CENTAUR’s ability to resist diverse data reconstruction attacks, achieve plaintext-level inference accuracy, and boost inference speed by 5.0~30.4 times, unlocking new possibilities for secure and efficient AI deployment.

In recent years, large language models (LLMs) have attracted significant interest from the research community due to their broad applicability in many language-oriented tasks, and are now widely used in numerous areas of production and daily life. One source of the powerful capabilities of LLMs is the massive scale of their pre-training dataset. However, these pre-training datasets contain many outdated, harmful, and personally sensitive information, which inevitably becomes memorized by LLM during the pre-training process. Eliminating this undesirable data is crucial for ensuring the model’s safety and enhancing the user experience. However, the cost of extensively cleaning the pre-training dataset and retraining the model from scratch is very high. In this work, we propose ULMR , a unlearning framework for LLMs , which first uses carefully designed prompts to rewrite the instructions in the specified dataset, and generate corresponding negative responses. Subsequently, to ensure that the model does not excessively deviate post-training, we perform model parameter averaging to preserve the performance of the original LLM. We conducted experiments on two public datasets, TOFU and RWKU, demonstrating that our method can effectively forget specified information while retaining the capabilities of the original LLM.

Recently, there has been a notable surge in the significance of large language models (LLMs) that engage in conversational-style interactions, such as ChatGPT and Claude, as they contribute significantly to the progress of artificial general intelligence (AGI). Typically, these models undergo a two-phase fine-tuning process: instruction fine-tuning (IF) and reinforcement learning from human feedback (RLHF). These methods aim to align the LLMs to be helpful, honest, and harmless (HHH). However, RLHF, which incorporates independent reward models trained on high-quality human feedback datasets, incurs high costs in terms of hardware resources and human efforts. Therefore, we explore the possibility of aligning LLMs with their own understanding of HHH through IF and in-context learning (ICL). In this study, we propose a novel framework called Self-Criticism, which allows LLMs to align themselves with HHH based on the definition they learned from a large-scale text corpus. We begin by employing IF on a given instruction set and learning HHH discrimination through few-shot ICL. Subsequently, the LLMs evaluate their own generated responses and learn to produce “better” responses based on self-judgment. Finally, the model is retrained based on the self-generated responses to distill the whole process. By analyzing our proposed method, we also find interesting connections between Self-Criticism and goal-conditioned reinforcement learning, and pseudo-labeling. Experimental results demonstrate that this method achieves nearly identical performance to RLHF in terms of both human evaluation and evaluation by other LLMs, with only a minimal alignment tax.

Instruction fine-tuning has conventionally been employed to adapt Large Language Models (LLMs) to a variety of diverse tasks. Nonetheless, this technique often necessitates substantial computational resources, making it impractical for deployment by individuals or small-scale entities. Recently, Low-Rank Adaptation (LoRA) has become a promising alternative, offering tuning capabilities with reduced resource overhead. However, attaining satisfactory performance through the fine-tuning of LoRA is a non-trivial challenge. In this paper, we propose PILLOW, which aims to improve LoRA’s performance by leveraging LLM’s in-context learning capability through prompt matching via reinforcement learning in resource-constrained environments. Specifically, PILLOW incorporates a matching network that selects prompts from a user-defined pool, concatenates the optimal prompts given the user instruction, and performs inference using the LoRA-fine-tuned LLMs. Compared with typical instruction fine-tuning methods, PILLOW exhibits commensurate performance on various evaluation metrics, utilizing only consumer-grade GPU resources and exhibiting a large increase in training efficiency.

Learning on noisy datasets is a challenging problem when pre-trained language models are applied to real-world text classification tasks. In numerous industrial applications, acquiring task-specific datasets with 100% accurate labels is difficult, thus many datasets are accompanied by label noise at different levels. Previous work has shown that existing noise-handling methods could not improve the peak performance of BERT on noisy datasets, and might even deteriorate it. In this paper, we propose SaFER, a robust and efficient fine-tuning framework for BERT-based text classifiers, combating label noises without access to any clean data for training or validation. Utilizing a label-agnostic early-stopping strategy and self-supervised learning, our proposed framework achieves superior performance in terms of both accuracy and speed on multiple text classification benchmarks. The trained model is finally fully deployed in several industrial biomedical literature mining tasks and demonstrates high effectiveness and efficiency.

Numerical reasoning over texts, such as addition, subtraction, sorting and counting, is a challenging machine reading comprehension task, since it requires both natural language understanding and arithmetic computation. To address this challenge, we propose a heterogeneous graph representation for the context of the passage and question needed for such reasoning, and design a question directed graph attention network to drive multi-step numerical reasoning over this context graph. Our model, which combines deep learning and graph reasoning, achieves remarkable results in benchmark datasets such as DROP.

Chinese Spelling Check (CSC) is a task to detect and correct spelling errors in Chinese natural language. Existing methods have made attempts to incorporate the similarity knowledge between Chinese characters. However, they take the similarity knowledge as either an external input resource or just heuristic rules. This paper proposes to incorporate phonological and visual similarity knowledge into language models for CSC via a specialized graph convolutional network (SpellGCN). The model builds a graph over the characters, and SpellGCN is learned to map this graph into a set of inter-dependent character classifiers. These classifiers are applied to the representations extracted by another network, such as BERT, enabling the whole network to be end-to-end trainable. Experiments are conducted on three human-annotated datasets. Our method achieves superior performance against previous models by a large margin.