Reinforcement learning for code generation relies on verifiable rewards from unit test pass rates. Yet high-quality test suites are scarce, existing datasets offer limited coverage, and static rewards fail to adapt as models improve. Recent self-play methods unify code and test generation in a single model, but face an inherent dilemma: white-box access leads to self-collusion where the model produces trivial tests for easy rewards, yet black-box restriction yields generic tests that miss implementation-specific bugs. We introduce Code-A1, an adversarial co-evolution framework that jointly optimizes a Code LLM and a Test LLM with opposing objectives. The Code LLM is rewarded for passing more tests, while the Test LLM is rewarded for exposing more defects. This architectural separation eliminates self-collusion risks and safely enables white-box test generation, where the Test LLM can inspect candidate code to craft targeted adversarial tests. We further introduce a Mistake Book mechanism for experience replay and a composite reward balancing test validity with adversarial difficulty. Experiments on Qwen2.5-Coder models demonstrate that Code-A1 achieves code generation performance matching or exceeding models trained on human-annotated tests, while significantly improving test generation capability.

Agent skills, structured packages of procedural knowledge and executable resources that agents dynamically load at inference time, have become a reliable mechanism for augmenting LLM agents. Yet inference-time skill augmentation is fundamentally limited: retrieval noise introduces irrelevant guidance, injected skill content imposes substantial token overhead, and the model never truly acquires the knowledge it merely follows. We ask whether skills can instead be internalized into model parameters, enabling zero-shot autonomous behavior without any runtime skill retrieval. We introduce SKILL0, an in-context reinforcement learning framework designed for skill internalization. SKILL0 introduces a training-time curriculum that begins with full skill context and progressively withdraws it. Skills are grouped offline by category and rendered with interaction history into a compact visual context, teaching the model tool invocation and multi-turn task completion. A Dynamic Curriculum then evaluates each skill file’s on-policy helpfulness, retaining only those from which the current policy still benefits within a linearly decaying budget, until the agent operates in a fully zero-shot setting. Extensive agentic experiments demonstrate that SKILL0 achieves substantial improvements over the standard RL baseline (+9.7% for ALFWorld, +6.6% for Search-QA and +10.1% for WebShop), while maintaining a highly efficient context of fewer than 0.5k tokens per step.

Large reasoning models have achieved remarkable performance through extended chain-of-thought sequences, yet this computational freedom leads to excessive token generation even for simple problems. We present Length-Adaptive Policy Optimization (LAPO), a novel framework that transforms reasoning length control from an external constraint into an intrinsic model capability. Unlike existing approaches that impose rigid limits or rely on post-hoc interventions, LAPO enables models to internalize an understanding of appropriate reasoning depth through a two-stage reinforcement learning process. In the first stage, models learn natural reasoning patterns by discovering the statistical distribution of successful solution lengths. In the second stage, these learned patterns are embedded as in-context, self-declarative guidance, teaching the model to proactively plan its reasoning budget. Experiments on mathematical reasoning benchmarks demonstrate that LAPO reduces token usage by up to 38.5% while improving accuracy by 2.3%. Our analysis reveals that models trained with LAPO develop emergent abilities to allocate computational resources based on problem complexity, achieving efficient reasoning without sacrificing quality.

Theory of Mind (ToM), the ability to infer others’ mental states from behavior, is pivotal for developing machines with human-level social intelligence. Existing methods endowing LLMs with ToM fall into two paradigms: training-free methods and those repurposing ToM evaluation benchmarks as training data for RL-based fine-tuning. However, training-free methods fail to internalize the augmented ToM into the LLMs. Meanwhile, using evaluation benchmarks as training sources is conceptually problematic and, in practice, results in narrow in-domain overfitting rather than robust ToM. To address the lack of training resources within the ToM community and to empower LLMs with robust ToM, we introduce ToM-Synth, a factorial combinatorial synthesis framework of 6912 social units. This framework enables the systematic synthesis of ToM data, yielding a training dataset of 27,648 instances, termed ToM-Synth-27K. Utilizing ToM-Synth-27K for RL fine-tuning, experimental results demonstrate consistent and significant improvements across models of varying families and scales on ToM, Emotional Intelligence, and Social Commonsense benchmarks. Furthermore, we observe concurrent enhancements in IQ-related tasks (math, science, logic) and effective performance scaling with increasing data scale.

Large language models have achieved remarkable progress on complex reasoning tasks. However, they often implicitly fabricate information when inputs are incomplete, producing confident but unreliable conclusions—a failure mode we term ungrounded reasoning. We argue that this issue arises not from insufficient reasoning capability, but from the lack of inferential boundary awareness—the ability to recognize when the necessary premises for valid inference are missing. To address this issue, we propose Grounded Reasoning via Interactive Reinforcement Learning (GRIL), a multi-turn reinforcement learning framework for grounded reasoning under incomplete information. GRIL decomposes the reasoning process into two stages: clarify and pause, which identifies whether the available information is sufficient, and grounded reasoning, which performs task solving once the necessary premises are established. We design stage-specific rewards to penalize hallucinations, enabling models to detect gaps, stop proactively, and resume reasoning after clarification. Experiments on GSM8K-Insufficient and MetaMATH-Insufficient show that GRIL significantly improves premise detection (up to 45%), leading to a 30% increase in task success while reducing average response length by over 20%. Additional analyses confirm robustness to noisy user responses and generalization to out-of-distribution tasks.

Data-analytic agents must iteratively generate and execute code while interpreting real-time feedback to refine their analytical strategies. Existing training paradigms decouple the agent’s policy from the underlying environment dynamics, producing agents that learn to recover from errors rather than to avoid them: failure rates rise alongside recovery rates, and interaction turns inflate as agents resort to trial-and-error exploration. We introduce DataSight, a framework that internalizes environment dynamics by making the model’s anticipation of execution outcomes an explicit, structured output accompanying every generated action. DataSight first enriches expert trajectories with foresight prediction fields grounded in real execution feedback, then applies a dual-stage RL curriculum that calibrates prediction fidelity before optimizing long-horizon task performance, and at inference time intercepts anticipated failures before physical execution. Trained on only 7,273 trajectories, DataSight-4B surpasses both DataMind-7B and DeepAnalyse-8B in average accuracy, while DataSight-8B achieves 64.13% Avg@3 across four benchmarks with Pass@3 of 74.81%, approaching DeepSeek-V3.2 at 75.56%. Analysis confirms that foresight training generalizes across non-reasoning backbones and SQL-based tasks, improving both single-step code quality and multi-turn interaction efficiency.

Hint-based reinforcement learning addresses reward sparsity in long-horizon agentic tasks by retaining a prefix of an expert trajectory before each rollout, letting the policy explore from a state closer to success. Its effectiveness hinges on the guidance depth: how much of the trajectory to keep. Existing methods treat this depth as a deterministic scalar. Scheduled approaches share one value across samples and ignore per-task heterogeneity; per-sample probing estimates it separately at the cost of extra rollouts. We find that useful guidance occupies a band of depths whose informativeness profile is approximately Gaussian around the band center, rather than concentrating at a single optimal point. We propose Agent-G², a Gaussian guidance framework that draws the depth per task from a Gaussian whose center and spread are estimated online from rollouts already collected for policy optimization, requiring no probe rollouts or learned depth predictor. The center combines a global baseline with per-cluster difficulty, and the spread tracks within-cluster variance. We evaluate Agent-G² on ALFWorld and WebShop with Qwen2.5-1.5B / 7B-Instruct. Agent-G² outperforms the strongest hint-based, hint-free, and Aux-RL baselines on ALFWorld by 2.3 / 3.9 / 7.4 points at under one-third the rollout cost of per-sample probing.

Deep search agents solve knowledge-intensive tasks through long-horizon, multi-turn interactions, where assigning credit to individual search decisions is fundamentally difficult. Outcome-based reinforcement learning provides only a single terminal reward shared across an entire trajectory. Recent process-supervision methods aim to assign credit at finer granularity, but they typically anchor credit to either surface-level trajectory states or model-internal signals, both of which become unreliable in deep search where observations are heterogeneous and retrieved contexts are long and noisy. We propose KnowGPO, a reinforcement learning framework that aligns trajectories by the task-relevant knowledge they have acquired rather than by their surface observations. For each query, KnowGPO decomposes the information required to derive the answer into atomic knowledge conditions, and represents the state of a trajectory at each step by the subset of conditions it has satisfied. Trajectories sharing the same knowledge state are grouped for within-group comparison, and the incremental knowledge gain between consecutive steps serves as an environment-grounded process reward. Experiments on Xbench, BrowseComp-ZH, and GAIA show KnowGPO consistently outperforms outcome-only (GRPO) and process-supervision baselines (GiGPO, IGPO) by 3.8 to 4.2 points, with our 8B model rivaling or exceeding most 32B open-source systems. Our code is available at https://github.com/ZJU-REAL/KnowGPO.

Deep search requires LLM agents to decompose complex queries, search for evidence, and synthesize grounded answers, yet existing ReAct-style agents suffer from two limitations: role coupling, where one policy must handle planning, evidence use, and synthesis; and context accumulation, where growing search histories introduce noise and obscure useful information. To address these issues, we propose IterSynth, a role-decoupled and summary-based paradigm that alternates between a Planner for identifying information needs and a Synthesizer for integrating evidence into an evolving summary state. This design separates planning from synthesis while using the summary as the persistent state of search, reducing both capability coupling and context noise. To train IterSynth effectively, we further introduce Role-Decoupled Policy Optimization (RDPO) for reinforcement learning, which combines terminal outcome rewards with turn-level rubric evaluations and computes role-specific advantages for more precise credit assignment. Experiments on five long-horizon deep-search benchmarks such as BrowseComp and Xbench-DS show that IterSynth-8B achieves an average score of 50.7, surpassing the strongest prior ≤8B agent by +4.2%. Moreover, IterSynth serves as a model-agnostic prompting paradigm, delivering substantial zero-shot gains over ReAct and similar prompting paradigms on frontier proprietary models.

Retrieval-Augmented Generation (RAG) has advanced open-domain question answering by incorporating external information into model reasoning. However, raw retrieved content often suffers from low signal-to-noise ratio, where answer-supportive information is diluted by irrelevant material, and error accumulation, where incomplete or misleading content degrades multi-hop reasoning. To address these challenges, we propose EviNoteRAG, a framework that enhances RAG through quality-augmented evidence notes. The model is trained to generate Supportive-Evidence Notes (SENs), concise abstractions that preserve only answer-critical information and explicitly highlight key and uncertain content to improve reasoning accuracy. To further ensure note quality, we introduce an entailment-based Evidence Quality Reward (EQR), which incentivizes SENs to be logically sufficient for deriving the correct answer. Experiments on both in-domain and out-of-domain benchmarks demonstrate that EviNoteRAG achieves state-of-the-art performance, with relative F1 gains of 20% on HotpotQA (+0.093), 40% on Bamboogle (+0.151), and 91% on 2Wiki (+0.256), while also improving training stability and efficiency.

Mathematical reasoning is a key capability for vision-language models (VLMs), yet current benchmarks mainly evaluate text-based or explicitly symbolic visual inputs. It remains unclear whether VLMs can reason mathematically when information must be perceived and inferred from images rather than read from explicit symbols. We introduce GSM8K-V, a benchmark transforming GSM8K into multi-image sequences with semantic equivalence preserved. By mapping text-based problems into visual form via an automated pipeline and human verification, we curate 1,319 high-quality samples. In GSM8K-V, quantities must be extracted through visual perception, and reasoning chains must be reconstructed by integrating implicit cues across scenes. Evaluation of 34 VLMs reveals a striking modality gap: while most models exceed 90% on text, the best model achieves only 59% on GSM8K-V, far below the 91% human accuracy. Notably, models enhanced for visual math reasoning show no improvement on GSM8K-V despite large gains on existing benchmarks, confirming that it evaluates a distinct capability. Error analysis shows that the primary bottleneck lies in Implicit Visual Inference Error (IVIE), where models fail to recover visual semantics that are implied rather than explicitly stated. Our code and data are released at https://github.com/ZJU-REAL/GSM8K-V.

Large language model (LLM) steering has emerged as a promising paradigm for controlling model behavior at inference time through targeted manipulation of hidden states, offering a lightweight alternative to expensive retraining. However, existing steering frameworks suffer from critical limitations: computational inefficiency, limited extensibility, and restricted functionality that hinder both research progress and practical deployment. We present EasySteer, a unified framework for high-performance, extensible LLM steering built on vLLM. Our system features modular architecture with pluggable interfaces for both analysis-based and learning-based methods, fine-grained parameter control, pre-computed steering vectors for eight application domains, and an interactive demonstration system. Through deep integration with vLLM’s optimized inference engine, EasySteer achieves 10.8-22.3× speedup over existing frameworks. Extensive experiments demonstrate its effectiveness in overthinking mitigation, hallucination reduction, and other key applications. EasySteer transforms steering from research technique to production-ready capability, establishing critical infrastructure for deployable, controllable language models.

MLLM-based GUI agents have demonstrated strong capabilities in complex user interface interaction tasks. However, long-horizon scenarios remain challenging, as these agents are burdened with tasks beyond their intrinsic capabilities, suffering from memory degradation, progress confusion, and math hallucination. To address these challenges, we present UI-Copilot, a collaborative framework where the GUI agent focuses on task execution while a lightweight copilot provides on-demand assistance for memory retrieval and numerical computation. We introduce memory decoupling to separate persistent observations from transient execution context, and train the policy agent to selectively invoke the copilot as Retriever or Calculator based on task demands. To enable effective tool invocation learning, we propose  ̲Tool- ̲Integrated  ̲Policy  ̲Optimization (TIPO), which separately optimizes tool selection through single-turn prediction and task execution through on-policy multi-turn rollouts. Experimental results show that UI-Copilot-7B achieves state-of-the-art performance on challenging MemGUI-Bench, outperforming strong 7B-scale GUI agents such as GUI-Owl-7B and UI-TARS-1.5-7B. Moreover, UI-Copilot-7B delivers a 17.1% absolute improvement on AndroidWorld over the base Qwen model, highlighting UI-Copilot’s strong generalization to real-world GUI tasks. Code website: https://anonymous.4open.science/r/UI-Copilot-0535.

GUI agents have demonstrated remarkable progress in automating complex user interface interactions. However, training such agents for long-horizon tasks remains challenging. Single-turn reinforcement learning conditions on expert histories during training but self-generated histories during deployment, causing distribution mismatch. Online multi-turn methods eliminate this gap via environment interaction but suffer from sparse rewards and prohibitive costs. We propose  ̲Experience-driven  ̲Multi-turn  ̲Policy  ̲Optimization (EMPO), which leverages expert trajectories as environment experiences for on-policy multi-turn training. The agent constructs self-generated history throughout rollouts; when actions match expert experiences, the trajectory provides valid state transitions, and a Patch Module recovers mismatched steps to maintain on-policy rollouts. EMPO further incorporates discounted future rewards and dual-level advantage estimation to capture long-horizon dependencies. We also propose AndroidControl-Real, an evaluation metric strongly correlated with real-world performance (R²=0.934). With only 1K public trajectories as RL experiences, our method achieves substantial gains over the base model (e.g., +12.0% on AndroidWorld and +23.8% on AITW) and achieves competitive performance against strong baselines such as UI-TARS-7B and GPT-4o, demonstrating better generalization than prior single-turn RL approaches. Code available: https://anonymous.4open.science/r/UI-S1-0DAF.

Despite the rapid progress of LLMs, their evaluation remains hindered by static, manually curated benchmarks with limited task coverage and poor adaptability to emerging domains. Existing automated approaches typically operate within fixed task schemas and often fail to autonomously discover new evaluation dimensions, limiting both scalability and effectiveness. To address these gaps, we propose AutoTaskEval, an automated framework that constructs domain-specific benchmarks directly from unstructured corpora. Using a refined Bloom’s Taxonomy, the framework systematically discovers tasks, enriches contextual grounding via iterative Socratic prompting, and generates diverse, progressively challenging evaluation instances. Applied to the complex and knowledge-intensive legal domain, AutoTaskEval uncovers a broader and more fine-grained task space than expert-curated benchmarks while producing high-quality instances that preserve established model-level evaluation trends. We further validate its robustness in a low-structure e-commerce review domain. Together, these results show that AutoTaskEval enables scalable, adaptive, and high-fidelity LLM assessment across domains and model families, advancing autonomous and capability-sensitive evaluation.

Recent advances in reasoning models have demonstrated remarkable capabilities on mathematical and coding tasks. However, their effectiveness in embodied domains, where the agent must continuously interact with environments and process observation-action interleaved trajectories, remains largely unexplored. We present Embodied-Reasoner, a reasoning model for interactive embodied tasks. Unlike mathematical reasoning that relies primarily on logical deduction, embodied scenarios demand spatial understanding, temporal reasoning, and ongoing self-reflection based on interaction history. To address these challenges, we synthesize 9.3k coherent Observation-Thought-Action trajectories containing 64k ego-centric images and 90k diverse reasoning processes (analysis, spatial reasoning, reflection, planning, and verification). We develop a three-stage training recipe that progressively enhances the model’s capabilities through imitation learning, rejection sampling tuning on self-exploration trajectories, and reflection tuning. The evaluation shows that our model significantly outperforms advanced visual reasoning models, e.g., exceeds OpenAI o1, o3-mini, and Claude-3.7 by +9%, 24%, and +13%. Analysis reveals that our model exhibits fewer repeated searches and logical inconsistencies, with particular advantages in complex long-horizon tasks. Real-world testing further validates the effectiveness of our approach.

Long-form outline generation requires satisfying multiple competing objectives simultaneously: outlines must be engaging, well-organized, topically relevant, and comprehensive while maintaining logical consistency across hierarchical structures. Current approaches either rely on expensive multi-turn interactions with large language models or employ procedural refinement pipelines that cannot systematically learn from critique. We present Logic-RL, a framework that transforms critique-guided outline refinement into a learnable policy through reinforcement learning. Our approach constructs refinement trajectories from teacher demonstrations, synthesizes explicit reasoning chains that decompose the critique-revision process, and optimizes a refinement policy using group relative policy optimization with structure-aware rewards. Experiments on FreshWiki and WikiOutline demonstrate that Logic-RL achieves substantial improvements over strong baselines, with the 0.6B model obtaining 79.17% relative gain and the 1.7B model achieving 8.67% improvement in average rubric scores compared to the best existing methods. Further analysis reveals that learned refinement policies generalize across domains and can be iteratively applied, with quality continuing to improve through three refinement rounds before diminishing returns.

Large Language Models (LLMs) are increasingly deployed in socially sensitive domains, yet their unpredictable behaviors, ranging from misaligned intent to inconsistent personality, pose significant risks. We introduce SteerEval, a hierarchical benchmark for evaluating LLM controllability across three domains: language features, sentiment, and personality. Each domain is structured into three specification levels: L1 (what to express), L2 (how to express), and L3 (how to instantiate), connecting high-level behavioral intent to concrete textual output. Using SteerBench, we systematically evaluate contemporary steering methods, revealing that control often degrades at finer-grained levels. Our benchmark offers a principled and interpretable framework for safe and controllable LLM behavior, serving as a foundation for future research.

Multimodal Mixture-of-Experts (MoE) models have achieved remarkable performance on vision-language tasks. However, we identify a puzzling phenomenon termed Seeing but Not Thinking: models accurately perceive image content yet fail in subsequent reasoning, while correctly solving identical problems presented as pure text. Through systematic analysis, we first verify that cross-modal semantic sharing exists in MoE architectures, ruling out semantic alignment failure as the sole explanation. We then reveal that visual experts and domain experts exhibit layer-wise separation, with image inputs inducing significant routing divergence from text inputs in middle layers where domain experts concentrate. Based on these findings, we propose the Routing Distraction hypothesis: when processing visual inputs, the routing mechanism fails to adequately activate task-relevant reasoning experts. To validate this hypothesis, we design a routing-guided intervention method that enhances domain expert activation. Experiments on three multimodal MoE models across six benchmarks demonstrate consistent improvements, with gains of up to 3.17% on complex visual reasoning tasks. Our analysis further reveals that domain expert identification locates cognitive functions rather than sample-specific solutions, enabling effective transfer across tasks with different information structures.

Label-free reinforcement learning enables large language models to improve reasoning capabilities without ground-truth supervision, typically by treating majority-voted answers as pseudo-labels. However, we identify a critical failure mode: as training maximizes self-consistency, output diversity collapses, causing the model to confidently reinforce systematic errors that evade detection. We term this the consensus trap. To escape it, we propose CoVerRL, a framework where a single model alternates between generator and verifier roles, with each capability bootstrapping the other. Majority voting provides noisy but informative supervision for training the verifier, while the improving verifier progressively filters self-consistent errors from pseudo-labels. This co-evolution creates a virtuous cycle that maintains high reward accuracy throughout training. Experiments across Qwen and Llama model families demonstrate that CoVerRL outperforms label-free baselines by 4.7-5.9% on mathematical reasoning benchmarks. Moreover, self-verification accuracy improves from around 55% to over 85%, confirming that both capabilities genuinely co-evolve.

Language is a vehicle for thought, intricately tied to sounds, symbols, and meaning. However, most large language model (LLM) research focuses on meaning (semantics) and symbols (spelling) while largely overlooking sounds. Existing benchmarks on LLMs’ phonological abilities are either solvable through rote memorization or intertwined with other abilities, making them inadequate to measure LLMs’ genuine ability in phonological understanding. Here, we present Phun-Bench, a purpose-built Chinese benchmark with diverse tasks and settings across three dimensions (Homophony, Rhyme, and Phonetic Similarity), designed to systematically evaluate LLMs’ phonological understanding. Our results show that while LLMs excel at recalling correct pronunciations, they generally struggle to leverage phonological knowledge in the flexible and intuitive way that human speakers do. Moreover, through detailed analyses, we propose a hypothesis regarding the underlying mechanism of LLMs’ phonological understanding and “perception”, highlighting an underexplored frontier for future research.

There has been a rising interest in utilizing tools in applications of autonomous agents based on large language models (LLMs) to address intricate real-world tasks. To develop LLMbased agents, it usually requires LLMs to understand many tool functions from different tool documentations. However, these documentations could be diverse, redundant, or incomplete, which immensely affects the capability of LLMs in using tools. Current LLMs exhibit satisfactory instruction-following capabilities based on instruction-following fine-tuning process. Motivated by this, in this paper, we introduce EASYTOOL, a framework transforming diverse and lengthy tool documentation into a unified and concise tool instruction to fully leverage instruction-following capabilities of LLMs for easier tool usage. EASYTOOL purifies essential information from extensive tool documentation of different sources, and elaborates a unified interface (i.e., tool instruction) to offer standardized tool descriptions and functionalities for LLM-based agents. Extensive experiments on multiple different tasks demonstrate that EASYTOOL can significantly reduce token consumption and improve the performance of LLM-based agents on tool utilization in real-world scenarios. Our code is available in supplemental materials. Our code is available at https://github.com/microsoft/JARVIS/tree/main/easytool.

Long-form outline generation for expository articles requires both comprehensive knowledge coverage and logical coherence, which is essential for creating detailed Wikipedia-like content. However, existing methods face critical limitations: outlines generated in the pre-writing stage often have low knowledge density and lack detail, while retrieval-augmented approaches struggle to maintain logical coherence across retrieved information. Additionally, unlike human writers who can iteratively improve through peer feedback and reference similar topics, current approaches lack effective mechanisms for systematic outline refinement. To address these challenges, we propose Logic, a Long-form Outline Generation system via Imitative and Critical self-refinement that mimics human writers’ refinement process. Logic establishes a coherent planning framework and structured knowledge base, learns from similar topic outlines through imitation, and continuously improves through model-based critique. Experiments on FreshWiki and our dataset WikiOutline show that, compared to the best baseline, Logic’s long-form outlines are more organized (with increases of 22.85% and 21.65% respectively) and more logically coherent (with increases of 16.19% and 12.24% respectively). Human evaluation further validates Logic’s effectiveness in generating comprehensive and well-structured long-form outlines.

Recent text-to-SQL systems powered by large language models (LLMs) have demonstrated remarkable performance in translating natural language queries into SQL.However, these systems often struggle with complex database structures and domain-specific queries, as they primarily focus on enhancing logical reasoning and SQL syntax while overlooking the critical need for comprehensive database understanding.To address this limitation, we propose DB-Explore, a novel framework that systematically aligns LLMs with database knowledge through automated exploration and instruction synthesis.DB-Explore constructs database graphs to capture complex relational schemas, leverages GPT-4 to systematically mine structural patterns and semantic knowledge, and synthesizes instructions to distill this knowledge for efficient fine-tuning of LLMs.Our framework enables comprehensive database understanding through diverse sampling strategies and automated instruction generation, bridging the gap between database structures and language models.Experiments conducted on the SPIDER and BIRD benchmarks validate the effectiveness of DB-Explore, achieving an execution accuracy of 67.0% on BIRD and 87.8% on SPIDER. Notably, our open‐source implementation based on Qwen2.5‐Coder‐7B achieves state‐of‐the‐art results at minimal computational cost, outperforming several GPT‐4‐driven Text‐to‐SQL systems.

Humans continually engage in reasoning about others’ mental states, a capability known as Theory of Mind (ToM), is essential for social interactions. While this social reasoning capability emerges naturally in human cognitive development, how has the social reasoning capability of Large Language Models (LLMs) evolved during their development process? Various datasets have been proposed to assess LLMs’ social reasoning capabilities, but each is designed with a distinct focus, and none have explored how models’ social reasoning capabilities evolve during model size scaling or reasoning tokens scaling. In light of this, we optimize the evaluation of LLMs’ social reasoning from both data and model perspectives, constructing progressively difficult levels of social reasoning data and systematically exploring how LLMs’ social reasoning capabilities evolve. Furthermore, through an in-depth analysis of DeepSeek-R1’s reasoning trajectories, we identify notable cognitive “Aha Moment” and the reasons for its reasoning errors. Experiments reveal that long-thought logical capabilities and cognitive thinking are key to scaling LLMs’ social reasoning capabilities. By equipping the Qwen2.5-32B-Instruct model with long-thought logical capabilities and cognitive thinking, we achieve an improvement of 19.0 points, attaining social reasoning performance comparable to o1-preview model.

Large language models (LLMs) have demonstrated remarkable capabilities in tool learning. In real-world scenarios, user queries are often ambiguous and incomplete, requiring effective clarification. However, existing interactive clarification approaches face two critical limitations: reliance on manually constructed datasets, which inherently constrains training data scale and diversity, and lack of error correction mechanisms during multi-turn clarification, leading to error accumulation that compromises both accuracy and efficiency. We present AskToAct, which addresses these challenges by exploiting the structural mapping between queries and their tool invocation solutions. Our key insight is that tool parameters naturally represent explicit user intents. By systematically removing key parameters from queries while retaining them as ground truth, we enable automated construction of high-quality training data. We further enhance model robustness through error-correction pairs and selective masking, enabling dynamic error detection during clarification interactions. Comprehensive experiments demonstrate that AskToAct significantly outperforms existing approaches, achieving above 57% accuracy in recovering critical unspecified intents and enhancing clarification efficiency by an average of 10.46% while maintaining high accuracy in tool invocation. Our framework exhibits robust performance across different model architectures and successfully generalizes to entirely unseen APIs without additional training, achieving performance comparable to GPT-4o with substantially fewer computational resources.

Generating step-by-step “chain-of-thought” rationales has proven effective for improving the performance of large language models on complex reasoning tasks. However, applying such techniques to structured tasks, such as text-to-SQL, remains largely unexplored. In this paper, we introduce Self-Taught Reasoner for text-to-SQL (STaR-SQL), a novel approach that reframes SQL query generation as a reasoning-driven process. Our method prompts the LLM to produce detailed reasoning steps for SQL queries and fine-tunes it on rationales that lead to correct outcomes. Unlike traditional methods, STaR-SQL dedicates additional test-time computation to reasoning, thereby positioning LLMs as spontaneous reasoners rather than mere prompt-based agents. To further scale the inference process, we incorporate an outcome-supervised reward model (ORM) as a verifier, which enhances SQL query accuracy. Experimental results on the challenging Spider benchmark demonstrate that STaR-SQL significantly improves text-to-SQL performance, achieving an execution accuracy of 86.6%. This surpasses a few-shot baseline by 31.6% and a baseline fine-tuned to predict answers directly by 18.0%. Additionally, STaR-SQL outperforms agent-like prompting methods that leverage more powerful yet closed-source models such as GPT-4. These findings underscore the potential of reasoning-augmented training for structured tasks and open the door to extending self-improving reasoning models to text-to-SQL generation and beyond.

Understanding sentiment is arguably an advanced and important capability of AI agents in the physical world. In previous works, many efforts have been devoted to individual sentiment subtasks, without considering interrelated sentiment knowledge among these subtasks. Although some recent works model multiple sentiment subtasks in a unified manner, they merely simply combine these subtasks without deeply exploring the hierarchical relationships among subtasks. In this paper, we introduce GSA-7B, an open-source large language model specific to the sentiment domain. Specifically, we deeply explore the hierarchical relationships between sentiment subtasks, proposing progressive sentiment reasoning benchmark and progressive task instructions. Subsequently, we use Llama2-7B as the backbone model and propose parameter-efficient progressive tuning paradigm which is implemented by constructing chain of LoRA, resulting in the creation of GSA-7B. Experimental results show that GSA-7B as a unified model performs well across all datasets in the progressive sentiment reasoning benchmark. Additionally, under the few-shot setting, GSA-7B also exhibits good generalization ability for sentiment subtasks and datasets that were not encountered during its training phase.

Theory of Mind (ToM)—the cognitive ability to reason about mental states of ourselves and others, is the foundation of social interaction. Although ToM comes naturally to humans, it poses a significant challenge to even the most advanced Large Language Models (LLMs). Due to the complex logical chains in ToM reasoning, especially in higher-order ToM questions, simply utilizing reasoning methods like Chain of Thought (CoT) will not improve the ToM capabilities of LLMs. We present TimeToM, which constructs a temporal space and uses it as the foundation to improve the ToM capabilities of LLMs in multiple scenarios. Specifically, within the temporal space, we construct Temporal Belief State Chain (TBSC) for each character and inspired by the cognition perspective of the social world model, we divide TBSC into self-world beliefs and social world beliefs, aligning with first-order ToM (first-order beliefs) and higher-order ToM (higher-order beliefs) questions, respectively. Moreover, we design a novel tool-belief solver that, by considering belief communication between characters in temporal space, can transform a character’s higher-order beliefs into another character’s first-order beliefs under belief communication period.

Large Language Models (LLMs) have demonstrated remarkable potential in handling complex reasoning tasks by generating step-by-step rationales. Some methods have proven effective in boosting accuracy by introducing extra verifiers to assess these paths. However, existing verifiers, typically trained on binary-labeled reasoning paths, fail to fully utilize the relative merits of intermediate steps, thereby limiting the effectiveness of the feedback provided. To overcome this limitation, we propose Tree-based Preference Learning Verifier (Tree-PLV), a novel approach that constructs reasoning trees via a best-first search algorithm and collects step-level paired data for preference training. Compared to traditional binary classification, step-level preferences more finely capture the nuances between reasoning steps, allowing for a more precise evaluation of the complete reasoning path. We empirically evaluate Tree-PLV across a range of arithmetic and commonsense reasoning tasks, where it significantly outperforms existing benchmarks. For instance, Tree-PLV achieved substantial performance gains over the Mistral-7B self-consistency baseline on GSM8K (67.55% → 82.79%), MATH (17.00% → 26.80%), CSQA (68.14% → 72.97%), and StrategyQA (82.86% → 83.25%). Additionally, our study explores the appropriate granularity for applying preference learning, revealing that step-level guidance provides feedback that better aligns with the evaluation of the reasoning process.

Although most current large multimodal models (LMMs) can already understand photos of natural scenes and portraits, their understanding of abstract images, e.g., charts, maps, or layouts, and visual reasoning capabilities remains quite rudimentary. They often struggle with simple daily tasks, such as reading time from a clock, understanding a flowchart, or planning a route using a road map. In light of this, we design a multi-modal self-instruct, utilizing large language models and their code capabilities to synthesize massive abstract images and visual reasoning instructions across daily scenarios. Our strategy effortlessly creates a multimodal benchmark with 11,193 instructions for eight visual scenarios: charts, tables, simulated maps, dashboards, flowcharts, relation graphs, floor plans, and visual puzzles. This benchmark, constructed with simple lines and geometric elements, exposes the shortcomings of most advanced LMMs like GPT-4V and Llava in abstract image understanding, spatial relations reasoning, and visual element induction. Besides, to verify the quality of our synthetic data, we fine-tune an LMM using 62,476 synthetic chart, table and road map instructions. The results demonstrate improved chart understanding and map navigation performance, and also demonstrate potential benefits for other visual reasoning tasks.

Large Language Models (LLMs) exhibit robust problem-solving capabilities for diverse tasks. However, most LLM-based agents are designed as specific task solvers with sophisticated prompt engineering, rather than agents capable of learning and evolving through interactions. These task solvers necessitate manually crafted prompts to inform task rules and regulate LLM behaviors, inherently incapacitating to address complex dynamic scenarios e.g., large interactive games. In light of this, we propose Agent-Pro: an LLM-based Agent with Policy-level Reflection and Optimization that can learn a wealth of expertise from interactive experiences and progressively elevate its behavioral policy. Specifically, it involves a dynamic belief generation and reflection process for policy evolution. Rather than action-level reflection, Agent-Pro iteratively reflects on past trajectories and beliefs, “fine-tuning” its irrational beliefs for a better policy. Moreover, a depth-first search is employed for policy optimization, ensuring continual enhancement in policy payoffs. Agent-Pro is evaluated across two games: Blackjack and Texas Hold’em, outperforming vanilla LLM and specialized models. Our results show Agent-Pro can learn and evolve in complex and dynamic scenes, which also benefits numerous LLM-based applications.

While large language models (LLMs) have showcased remarkable prowess in various natural language processing tasks, their training costs are exorbitant. Consequently, a plethora of parameter-efficient fine-tuning methods have emerged to tailor large models for downstream tasks, including low-rank training. Recent approaches either amalgamate existing fine-tuning methods or dynamically adjust rank allocation. Nonetheless, these methods continue to grapple with issues like local optimization, inability to train with full rank and lack of focus on specific tasks. In this paper, we introduce an innovative parameter-efficient method for exploring optimal solutions within latent space. More specifically, we introduce a set of latent units designed to iteratively extract input representations from LLMs, continuously refining informative features that enhance downstream task performance. Due to the small and independent nature of the latent units in relation to input size, this significantly reduces training memory requirements. Additionally, we employ an asymmetric attention mechanism to facilitate bidirectional interaction between latent units and freezed LLM representations, thereby mitigating issues associated with non-full-rank training. Furthermore, we apply distillation over hidden states during the interaction, which guarantees a trimmed number of trainable parameters.Experimental results demonstrate that our approach achieves state-of-the-art performance on a range of natural language understanding, generation and reasoning tasks.

Taxonomy completion, enriching existing taxonomies by inserting new concepts as parents or attaching them as children, has gained significant interest. Previous approaches embed concepts as vectors in Euclidean space, which makes it difficult to model asymmetric relations in taxonomy. In addition, they introduce pseudo-leaves to convert attachment cases into insertion cases, leading to an incorrect bias in network learning dominated by numerous pseudo-leaves. Addressing these, our framework, TaxBox, leverages box containment and center closeness to design two specialized geometric scorers within the box embedding space. These scorers are tailored for insertion and attachment operations and can effectively capture intrinsic relationships between concepts by optimizing on a granular box constraint loss. We employ a dynamic ranking loss mechanism to balance the scores from these scorers, allowing adaptive adjustments of insertion and attachment scores. Experiments on four real-world datasets show that TaxBox significantly outperforms previous methods, yielding substantial improvements over prior methods in real-world datasets, with average performance boosts of 6.7%, 34.9%, and 51.4% in MRR, Hit@1, and Prec@1, respectively.

The reflection capacity of Large Language Model (LLM) has garnered extensive attention. A post-hoc prompting strategy, e.g., reflexion and self-refine, refines LLM’s response based on self-evaluated or external feedback. However, recent research indicates without external feedback, LLM’s intrinsic reflection is unstable. Our investigation unveils that the key bottleneck is the quality of the self-evaluated feedback. We find LLMs often exhibit overconfidence or high randomness when self-evaluate, offering stubborn or inconsistent feedback, which causes poor reflection. To remedy this, we advocate Self-Contrast: It adaptively explores diverse solving perspectives tailored to the request, contrasts the differences, and summarizes these discrepancies into a checklist which could be used to re-examine and eliminate discrepancies. Our method endows LLM with diverse perspectives to alleviate stubborn biases. Moreover, their discrepancies indicate potential errors or inherent uncertainties that LLM often overlooks. Reflecting upon these can catalyze more accurate and stable reflection. Experiments conducted on a series of reasoning and translation tasks with different LLMs serve to underscore the effectiveness and generality of our strategy.

Emotion recognition in conversation (ERC) has attracted increasing attention in natural language processing community. Previous work commonly first extract semantic-view features via fine-tuning PLMs, then models context-view features based on the obtained semantic-view features by various graph neural networks. However, it is difficult to fully model interaction between utterances simply through a graph neural network and the features at semantic-view and context-view are not well aligned. Moreover, the previous parametric learning paradigm struggle to learn the patterns of tail class given fewer instances. To this end, we treat the pre-trained conversation model as a prior knowledge base and from which we elicit correlations between utterances by a probing procedure. And we adopt supervised contrastive learning to align semantic-view and context-view features, these two views of features work together in a complementary manner, contributing to ERC from distinct perspectives. Meanwhile, we propose a new semi-parametric paradigm of inferencing through memorization to solve the recognition problem of tail class samples. We consistently achieve state-of-the-art results on four widely used benchmarks. Extensive experiments demonstrate the effectiveness of our proposed multi-view feature alignment and memorization.

Extractive summarization focuses on extracting salient sentences from the source document and incorporating them in the summary without changing their wording or structure. The naive approach for extractive summarization is sentence classification, which makes independent binary decisions for each sentence, resulting in the model cannot detect the dependencies between sentences in the summary. Recent approaches introduce an autoregressive decoder to detect redundancy relationship between sentences by step-by-step sentence selection, but bring train-inference gap. To address these issues, we formulate extractive summarization as a salient sentence set recognition task. To solve the sentence set recognition task, we propose a set prediction network (SetSum), which sets up a fixed set of learnable queries to extract the entire sentence set of the summary, while capturing the dependencies between them.Different from previous methods with an auto-regressive decoder, we employ a non-autoregressive decoder to predict the sentences within the summary in parallel during both the training and inference process, which eliminates the train-inference gap. Experimental results on both single-document and multi-document extracted summary datasets show that our approach outperforms previous state-of-the-art models.

Generating mathematical equations from natural language requires an accurate understanding of the relations among math expressions. Existing approaches can be broadly categorized into token-level and expression-level generation. The former treats equations as a mathematical language, sequentially generating math tokens. Expression-level methods generate each expression one by one. However, each expression represents a solving step, and there naturally exist parallel or dependent relations between these steps, which are ignored by current sequential methods. Therefore, we integrate tree structure into the expression-level generation and advocate an expression tree decoding strategy. To generate a tree with expression as its node, we employ a layer-wise parallel decoding strategy: we decode multiple independent expressions (leaf nodes) in parallel at each layer and repeat parallel decoding layer by layer to sequentially generate these parent node expressions that depend on others. Besides, a bipartite matching algorithm is adopted to align multiple predictions with annotations for each layer. Experiments show our method outperforms other baselines, especially for these equations with complex structures.

Distantly supervised named entity recognition (DS-NER) aims to locate entity mentions and classify their types with only knowledge bases or gazetteers and unlabeled corpus. However, distant annotations are noisy and degrade the performance of NER models. In this paper, we propose a noise-robust prototype network named MProto for the DS-NER task. Different from previous prototype-based NER methods, MProto represents each entity type with multiple prototypes to characterize the intra-class variance among entity representations. To optimize the classifier, each token should be assigned an appropriate ground-truth prototype and we consider such token-prototype assignment as an optimal transport (OT) problem. Furthermore, to mitigate the noise from incomplete labeling, we propose a novel denoised optimal transport (DOT) algorithm. Specifically, we utilize the assignment result between Other class tokens and all prototypes to distinguish unlabeled entity tokens from true negatives. Experiments on several DS-NER benchmarks demonstrate that our MProto achieves state-of-the-art performance. The source code is now available on Github.

Prompt learning is a new paradigm for utilizing pre-trained language models and has achieved great success in many tasks. To adopt prompt learning in the NER task, two kinds of methods have been explored from a pair of symmetric perspectives, populating the template by enumerating spans to predict their entity types or constructing type-specific prompts to locate entities. However, these methods not only require a multi-round prompting manner with a high time overhead and computational cost, but also require elaborate prompt templates, that are difficult to apply in practical scenarios. In this paper, we unify entity locating and entity typing into prompt learning, and design a dual-slot multi-prompt template with the position slot and type slot to prompt locating and typing respectively. Multiple prompts can be input to the model simultaneously, and then the model extracts all entities by parallel predictions on the slots. To assign labels for the slots during training, we design a dynamic template filling mechanism that uses the extended bipartite graph matching between prompts and the ground-truth entities. We conduct experiments in various settings, including resource-rich flat and nested NER datasets and low-resource in-domain and cross-domain datasets. Experimental results show that the proposed model achieves a significant performance improvement, especially in the cross-domain few-shot setting, which outperforms the state-of-the-art model by +7.7% on average.

In this paper, we propose DiffusionNER, which formulates the named entity recognition task as a boundary-denoising diffusion process and thus generates named entities from noisy spans. During training, DiffusionNER gradually adds noises to the golden entity boundaries by a fixed forward diffusion process and learns a reverse diffusion process to recover the entity boundaries. In inference, DiffusionNER first randomly samples some noisy spans from a standard Gaussian distribution and then generates the named entities by denoising them with the learned reverse diffusion process. The proposed boundary-denoising diffusion process allows progressive refinement and dynamic sampling of entities, empowering DiffusionNER with efficient and flexible entity generation capability. Experiments on multiple flat and nested NER datasets demonstrate that DiffusionNER achieves comparable or even better performance than previous state-of-the-art models.

The MultiCoNER shared task aims at detecting semantically ambiguous and complex named entities in short and low-context settings for multiple languages. The lack of contexts makes the recognition of ambiguous named entities challenging. To alleviate this issue, our team DAMO-NLP proposes a knowledge-based system, where we build a multilingual knowledge base based on Wikipedia to provide related context information to the named entity recognition (NER) model. Given an input sentence, our system effectively retrieves related contexts from the knowledge base. The original input sentences are then augmented with such context information, allowing significantly better contextualized token representations to be captured. Our system wins 10 out of 13 tracks in the MultiCoNER shared task.

Relation Induction is a very practical task in Natural Language Processing (NLP) area. In practical application scenarios, people want to induce more entity pairs having the same relation from only a few seed entity pairs. Thus, instead of the laborious supervised setting, in this paper, we focus on the minimally-supervised setting where only a couple of seed entity pairs per relation are provided. Although the conventional relation induction methods have made some success, their performance depends heavily on the quality of word embeddings. The great success of Pre-trained Language Models, such as BERT, changes the NLP area a lot, and they are proven to be able to better capture relation knowledge. In this paper, we propose a novel method to induce relation with BERT under the minimally-supervised setting. Specifically, we firstly extract proper templates from the corpus by using the mask-prediction task in BERT to build pseudo-sentences as the context of entity pairs. Then we use BERT attention weights to better represent the pseudo-sentences. In addition, We also use the IntegratedGradient of entity pairs to iteratively select better templates further. Finally, with the high-quality pseudo-sentences, we can train a better classifier for relation induction. Experiments onGoogle Analogy Test Sets (GATS), Bigger Analogy TestSet (BATS) and DiffVec demonstrate that our proposed method achieves state-of-the-art performance.

Math word problem solver requires both precise relation reasoning about quantities in the text and reliable generation for the diverse equation. Current sequence-to-tree or relation extraction methods regard this only from a fixed view, struggling to simultaneously handle complex semantics and diverse equations. However, human solving naturally involves two consistent reasoning views: top-down and bottom-up, just as math equations also can be expressed in multiple equivalent forms: pre-order and post-order. We propose a multi-view consistent contrastive learning for a more complete semantics-to-equation mapping. The entire process is decoupled into two independent but consistent views: top-down decomposition and bottom-up construction, and the two reasoning views are aligned in multi-granularity for consistency, enhancing global generation and precise reasoning. Experiments on multiple datasets across two languages show our approach significantly outperforms the existing baselines, especially on complex problems. We also show after consistent alignment, multi-view can absorb the merits of both views and generate more diverse results consistent with the mathematical laws.

Joint entity and relation extraction has been a core task in the field of information extraction. Recent approaches usually consider the extraction of relational triples from a stereoscopic perspective, either learning a relation-specific tagger or separate classifiers for each relation type. However, they still suffer from error propagation, relation redundancy and lack of high-level connections between triples. To address these issues, we propose a novel query-based approach to construct instance-level representations for relational triples. By metric-based comparison between query embeddings and token embeddings, we can extract all types of triples in one step, thus eliminating the error propagation problem. In addition, we learn the instance-level representation of relational triples via contrastive learning. In this way, relational triples can not only enclose rich class-level semantics but also access to high-order global connections. Experimental results show that our proposed method achieves the state of the art on five widely used benchmarks.

Named entity recognition (NER) is a fundamental task in natural language processing. Recent works treat named entity recognition as a reading comprehension task, constructing type-specific queries manually to extract entities. This paradigm suffers from three issues. First, type-specific queries can only extract one type of entities per inference, which is inefficient. Second, the extraction for different types of entities is isolated, ignoring the dependencies between them. Third, query construction relies on external knowledge and is difficult to apply to realistic scenarios with hundreds of entity types. To deal with them, we propose Parallel Instance Query Network (PIQN), which sets up global and learnable instance queries to extract entities from a sentence in a parallel manner. Each instance query predicts one entity, and by feeding all instance queries simultaneously, we can query all entities in parallel. Instead of being constructed from external knowledge, instance queries can learn their different query semantics during training. For training the model, we treat label assignment as a one-to-many Linear Assignment Problem (LAP) and dynamically assign gold entities to instance queries with minimal assignment cost. Experiments on both nested and flat NER datasets demonstrate that our proposed method outperforms previous state-of-the-art models.

Named entity recognition (NER) is a fundamental task to recognize specific types of entities from a given sentence. Depending on how the entities appear in the sentence, it can be divided into three subtasks, namely, Flat NER, Nested NER, and Discontinuous NER. Among the existing approaches, only the generative model can be uniformly adapted to these three subtasks. However, when the generative model is applied to NER, its optimization objective is not consistent with the task, which makes the model vulnerable to the incorrect biases. In this paper, we analyze the incorrect biases in the generation process from a causality perspective and attribute them to two confounders: pre-context confounder and entity-order confounder. Furthermore, we design Intra- and Inter-entity Deconfounding Data Augmentation methods to eliminate the above confounders according to the theory of backdoor adjustment. Experiments show that our method can improve the performance of the generative NER model in various datasets.

Prerequisite relations among concepts are crucial for educational applications, such as curriculum planning and intelligent tutoring. In this paper, we propose a novel concept prerequisite relation learning approach, named CPRL, which combines both concept representation learned from a heterogeneous graph and concept pairwise features. Furthermore, we extend CPRL under weakly supervised settings to make our method more practical, including learning prerequisite relations from learning object dependencies and generating training data with data programming. Our experiments on four datasets show that the proposed approach achieves the state-of-the-art results comparing with existing methods.

Named entity recognition (NER) is a well-studied task in natural language processing. Traditional NER research only deals with flat entities and ignores nested entities. The span-based methods treat entity recognition as a span classification task. Although these methods have the innate ability to handle nested NER, they suffer from high computational cost, ignorance of boundary information, under-utilization of the spans that partially match with entities, and difficulties in long entity recognition. To tackle these issues, we propose a two-stage entity identifier. First we generate span proposals by filtering and boundary regression on the seed spans to locate the entities, and then label the boundary-adjusted span proposals with the corresponding categories. Our method effectively utilizes the boundary information of entities and partially matched spans during training. Through boundary regression, entities of any length can be covered theoretically, which improves the ability to recognize long entities. In addition, many low-quality seed spans are filtered out in the first stage, which reduces the time complexity of inference. Experiments on nested NER datasets demonstrate that our proposed method outperforms previous state-of-the-art models.

In this paper, we study a new task of synonym expansion using transitivity, and propose a novel approach named SynET, which considers both the contexts of two given synonym pairs. It introduces an auxiliary task to reduce the impact of noisy sentences, and proposes a Multi-Perspective Entity Matching Network to match entities from multiple perspectives. Extensive experiments on a real-world dataset show the effectiveness of our approach.

Large pre-trained transformer-based language models have achieved impressive results on a wide range of NLP tasks. In the past few years, Knowledge Distillation(KD) has become a popular paradigm to compress a computationally expensive model to a resource-efficient lightweight model. However, most KD algorithms, especially in NLP, rely on the accessibility of the original training dataset, which may be unavailable due to privacy issues. To tackle this problem, we propose a novel two-stage data-free distillation method, named Adversarial self-Supervised Data-Free Distillation (AS-DFD), which is designed for compressing large-scale transformer-based models (e.g., BERT). To avoid text generation in discrete space, we introduce a Plug & Play Embedding Guessing method to craft pseudo embeddings from the teacher’s hidden knowledge. Meanwhile, with a self-supervised module to quantify the student’s ability, we adapt the difficulty of pseudo embeddings in an adversarial training manner. To the best of our knowledge, our framework is the first data-free distillation framework designed for NLP tasks. We verify the effectiveness of our method on several text classification datasets.