On-policy distillation (OPD) has emerged as a promising paradigm for reasoning-oriented post-training of large language models (LLMs), especially when combined with reinforcement learning from verifiable rewards (RLVR). Existing OPD methods rely on reverse KL (RKL)-based teacher supervision over trajectories sampled from the student policy. However, we identify a critical limitation: under large teacher–student policy divergence, RL-driven exploration often produces trajectories that the teacher assigns low probability, resulting in uninformative negative feedback. To address this, we propose Teacher-Guided Policy Optimization (TGPO). Unlike existing OPD methods that evaluate student-sampled actions using the teacher, TGPO directly guides next-token generation using the teacher’s predicted distribution on student-generated contexts, yielding an on-policy forward-KL objective. Combined with RLVR, this guidance steers exploration toward improved continuations. Experiments on reasoning benchmarks show that TGPO outperforms existing RKL-based OPD methods and remains robust across different teacher models. Furthermore, we find that teacher top-1 predictions suffice for effective guidance, without requiring the full teacher distribution.

Reinforcement learning (RL) has emerged as a promising paradigm for training reasoning-oriented models by leveraging rule-based reward signals. However, RL training typically tends to improve single-sample success rates (i.e., Pass@1) while offering limited exploration of diverse reasoning trajectories, which is crucial for multi-sample performance (i.e., Pass@k). Our preliminary analysis reveals that this limitation stems from a fundamental squeezing effect, whereby probability mass is excessively concentrated on a narrow subset of high-reward trajectories, restricting genuine exploration and constraining attainable performance under RL training. To address this issue, in this work, we propose Steering Probability Squeezing (SPS), a training paradigm that interleaves conventional RL with inverse reinforcement learning (IRL). SPS treats on-policy rollouts as demonstrations and employs IRL to explicitly reshape the induced trajectory distribution, thereby enhancing exploration without introducing external supervision. Experiments on five commonly used reasoning benchmarks demonstrate that SPS can enable better exploration and improve Pass@k. Beyond algorithmic contributions, we provide an analysis of RL learning dynamics and identify an empirical upper bound on Pass@k, shedding light on intrinsic exploration limits in RL-based reasoning models. Our findings suggest that alternating between RL and IRL offers an effective pathway toward extending the exploration capacity of reasoning-oriented large language models.

Due to the dynamically evolving nature of real-world query streams, relevance models struggle to generalize to practical search scenarios. A sophisticated solution is self-evolution techniques. However, in large-scale industrial settings with massive query streams, this technique faces two challenges: (1) informative samples are often sparse and difficult to identify, and (2) pseudo-labels generated by the current model could be unreliable. To address these challenges, in this work, we propose a Self-Evolving Relevance Model approach (SERM), which comprises two complementary multi-agent modules: a multi-agent sample miner, designed to detect distributional shifts and identify informative training samples, and a multi-agent relevance annotator, which provides reliable labels through a two-level agreement framework. We evaluated SERM on a large-scale industrial platform, which serves billions of user requests daily. Experimental results demonstrate that SERM can achieve significant performance gains through iterative self-evolution, as validated by extensive offline multilingual evaluations and online testing.

Test-Time Scaling has shown notable efficacy in addressing complex problems through scaling inference compute. However, within Large Audio-Language Models (LALMs), an unintuitive phenomenon exists: post-training models for structured reasoning trajectories results in marginal or even negative gains compared to training for direct answering. To investigate it, we introduce CAFE, an evaluation framework designed to quantify audio reasoning errors. Evaluation results reveal that LALMs encounter a critical bottleneck: Audio Perception Decay, which compromises reasoning performance. To address it, we propose MPAR², a strategy that encourages a step-by-step reasoning process with rich perceptual awareness. With reinforcement learning, MPAR² improves performance on CAFE from 31.74% to 63.51% and effectively mitigates audio perception decay, concurrently enhancing reasoning ability to achieve a strong 74.59% accuracy on the MMAU benchmark. Further analysis shows that MPAR² optimizes audio attention and enhances reasoning token budget efficiency.

Extended chain-of-thought (CoT) traces improve LLM reasoning but incur substantial computational and memory costs. While existing CoT compression methods mitigate this by condensing thought steps into compact representations via memory tokens and retaining only these representations at inference time, the loss of fine-grained information makes subsequent steps more error-prone. To alleviate this, we propose HybridThinker, where in addition to preserving these representations, thought steps are also temporarily retained to provide fine-grained details. However, we observe that naively keeping thought steps accessible to subsequent steps during training lets the model bypass memory tokens by retrieving information directly from these steps, leaving the model’s ability to compress and retrieve information through memory tokens insufficiently trained. We therefore introduce a hybrid training scheme, in which only some thought steps are directly accessible through attention to subsequent steps, while the other thought steps are masked, forcing the model to use memory tokens for compression and retrieval. Across four reasoning benchmarks, HybridThinker matches Vanilla on Qwen2.5-7B and approaches it on Llama3.1-8B, improving over LightThinker by 5.8 and 2.7 points, respectively, with similar inference time. Ablation studies confirm that both temporary thought-step retention and the hybrid training scheme contribute to these gains.

Agentic coding tools present new opportunities to transform research workflows. The performance of agent systems built depends on both large language models (LLMs) and the harness around LLMs, which is the orchestration code that determines an agent’s behavior. We present ToFu, an agentic harness for researchers that reads your codebase, edits files, runs commands, and integrates with your development tools.

ToFu plays a dual role in research. As a research assistant, it supports practical research workflows with superior token efficiency, lower cost, and multilingual capability compared with existing agentic harnesses. Its release under the MIT License further enables local deployment for privacy-sensitive users. As a research object, ToFu provides a white-box agentic harness that allows researchers to inspect, modify, and evaluate its orchestration logic, tool-use behavior, and harness design, while retaining strong benchmark performance and an application-level user experience.

Translating LaTeX documents is challenging because natural-language content is tightly interleaved with executable structures such as commands, equations, environments, and cross-references. Direct translation may therefore corrupt document structure and lead to formatting errors or compilation failures. We present LaTeXTrans, a structure-aware framework that separates semantic translation from executable LaTeX structure through four stages: structure-aware parsing, protected translation, structural validation, and document reconstruction. LaTeXTrans isolates non-translatable structures with protected placeholders, validates structural constraints after translation, and reconstructs the translated content into the original multi-file project structure. We evaluate LaTeXTrans on 100 real-world LaTeX projects across multiple domains and translation directions. Experimental results show consistent improvements in translation quality over direct large language model (LLM) translation while maintaining strong structural reliability. On English–Chinese translation, LaTeXTrans achieves a 97% compilation success rate with only 0.43 formatting errors per document on average. Controlled experiments further show that its structural performance remains stable when previously unseen user-defined command and environment identifiers are introduced.

Emotion is a core paralinguistic feature in voice interaction. It is widely believed that emotion understanding models learn fundamental representations that transfer to synthesized speech, making emotion understanding results a plausible reward or evaluation metric for assessing emotional expressiveness in speech synthesis. In this work, we critically examine this assumption by systematically evaluating Speech Emotion Recognition (SER) on synthesized speech across datasets, discriminative and generative SER models, and diverse synthesis models. We find that current SER models can not generalize to synthesized speech, largely because speech token prediction during synthesis induces a representation mismatch between synthesized and human speech. Moreover, generative Speech Language Models (SLMs) tend to infer emotion from textual semantics while ignoring paralinguistic cues. Overall, our findings suggest that existing SER models often exploit non-robust shortcuts rather than capturing fundamental features, and paralinguistic understanding in SLMs remains challenging.

Large Language Models (LLMs) have achieved remarkable performance in Machine Translation (MT), but deploying them at scale remains prohibitively expensive. A widely adopted remedy is the hybrid system paradigm, which balances cost and quality by serving most requests with a small model and selectively routing a fraction to a large model. However, existing routing strategies often rely on heuristics, external predictors, or absolute quality estimation, which fail to capture whether the large model actually provides a worthwhile improvement over the small one. In this paper, we formulate routing as a budget allocation problem and identify marginal gain, i.e., the large model’s improvement over the small model, as the optimal signal for budgeted decisions. Building on this, we propose RouteLMT (routing for LLM-based MT), an efficient in-model router that predicts this expected gain by probing the small translator’s prompt-token representation, without requiring external models or hypothesis decoding. Extensive experiments demonstrate that our RouteLMT outperforms heuristics, quality/difficulty estimation baselines, achieving a superior quality–budget Pareto frontier. Furthermore, we analyze regression risks and show that a simple guarded variant can mitigate severe quality losses.

Preference optimization methods like DPO have achieved remarkable performance in LLM alignment. However, the evaluation for these methods relies on a single response and overlooks other potential outputs, which could also be generated in real-world applications within this hypothetical space. To address this issue, this paper presents a Hypothesis-based PrEference-aware AnaLysis Framework (HEAL), a novel evaluation paradigm that formulates preference alignment as a re-ranking process within hypothesis spaces. The framework incorporates two complementary metrics: ranking accuracy for evaluating ordinal consistency and preference strength correlation for assessing continuous alignment. To facilitate this framework, we develop UniHypoBench, a unified hypothesis benchmark constructed from diverse instruction-response pairs. Through extensive experiments based on HEAL, with a particular focus on the intrinsic mechanisms of preference learning, we demonstrate that current preference learning methods can effectively capture preferences provided by proxy models while simultaneously suppressing negative samples. These findings contribute to preference learning research through two significant avenues. Theoretically, we introduce hypothesis space analysis as an innovative paradigm for understanding preference alignment. Practically, HEAL offers researchers robust diagnostic tools for refining preference optimization methods, while our empirical results identify promising directions for developing more advanced alignment algorithms capable of comprehensive preference capture.

Test-Time Scaling (TTS) is a promising approach to progressively elicit the model’s intelligence during inference. Recently, training-based TTS methods, such as continued reinforcement learning (RL), have further surged in popularity, while training-free TTS methods are gradually fading from prominence. However, the additional computation overhead of training amplifies the burden on test-time scaling.In this paper, we focus on training-free TTS methods for reasoning. We first design Conditional Step-level Self-refinement, a fine-grained sequential scaling method guided by process verification. On top of its effectiveness, we further combine it with other classical parallel scaling methods at the step level, to introduce a novel inference paradigm called Hybrid Test-Time Scaling. Extensive experiments on five instruction-tuned LLMs across different scales (3B-14B) and families demonstrate that hybrid strategy incorporating various training-free TTS methods at a fine granularity has considerable potential for expanding the reasoning performance boundaries of LLMs.

Alignment training is crucial for enabling large language models (LLMs) to cater to human intentions and preferences. It is typically performed based on two stages with different objectives: instruction-following alignment and human-preference alignment. However, aligning LLMs with these objectives in sequence suffers from an inherent problem: the objectives may conflict, and the LLMs cannot guarantee to simultaneously align with the instructions and human preferences well. To response to these, in this work, we propose a Hybrid Alignment Training (Hbat) approach, based on alternating alignment and modified elastic weight consolidation methods. The basic idea is to alternate between different objectives during alignment training, so that better collaboration can be achieved between the two alignment tasks. We experiment with Hbat on summarization and dialogue tasks. Experimental results show that the proposed Hbat can significantly outperform all baselines. Notably, Hbat yields consistent performance gains over the traditional two-stage alignment training when using both proximal policy optimization and direct preference optimization.

Large language models (LLMs) can handle multilingual and cross-lingual text within a single input; however, previous works leveraging multilingualism in LLMs primarily focus on using English as the pivot language to enhance language understanding and reasoning. Given that multiple languages are a compensation for the losses caused by a single language’s limitations, it’s a natural next step to enrich the model’s learning context through the integration of the original input with its multiple translations. In this paper, we start by revealing that LLMs learn from parallel multilingual input (PMI). Our comprehensive evaluation shows that PMI enhances the model’s comprehension of the input, achieving superior performance than conventional in-context learning (ICL). Furthermore, to explore how multilingual processing affects prediction, we examine the activated neurons in LLMs. Surprisingly, involving more languages in the input activates fewer neurons, leading to more focused and effective neural activation patterns. Also, this neural reaction coincidently mirrors the neuroscience insight about synaptic pruning, highlighting a similarity between artificial and biological ‘brains’.