Zisu Huang

Changze Lv, Jie Zhou, Wentao Zhao et al.

Nowadays, training and evaluating DeepResearch-generated reports remain challenging due to the lack of verifiable reward signals. Accordingly, rubric-based evaluation has become a common practice. However, existing approaches either rely on coarse, pre-defined rubrics that lack sufficient granularity, or depend on manually constructed query-specific rubrics that are costly and difficult to scale. In this paper, we propose a pipeline to train human-preference-aligned query-specific rubric generators tailored for DeepResearch report generation. We first construct a dataset of DeepResearch-style queries annotated with human preferences over paired reports, and train rubric generators via reinforcement learning with a hybrid reward combining human preference supervision and LLM-based rubric evaluation. To better handle long-horizon reasoning, we further introduce a Multi-agent Markov-state (MaMs) workflow for report generation. We empirically show that our proposed rubric generators deliver more discriminative and better human-aligned supervision than existing rubric design strategies. Moreover, when integrated into the MaMs training framework, DeepResearch systems equipped with our rubric generators consistently outperform all open-source baselines on the DeepResearch Bench and achieve performance comparable to that of leading closed-source models.

Xiaohua Wang, Zisu Huang, Feiran Zhang et al.

The capacity of large language models (LLMs) to generate honest, harmless, and helpful responses heavily relies on the quality of user prompts. However, these prompts often tend to be brief and vague, thereby significantly limiting the full potential of LLMs. Moreover, harmful prompts can be meticulously crafted and manipulated by adversaries to jailbreak LLMs, inducing them to produce potentially toxic content. To enhance the capabilities of LLMs while maintaining strong robustness against harmful jailbreak inputs, this study proposes a transferable and pluggable framework that refines user prompts before they are input into LLMs. This strategy improves the quality of the queries, empowering LLMs to generate more truthful, benign and useful responses. Specifically, a lightweight query refinement model is introduced and trained using a specially designed reinforcement learning approach that incorporates multiple objectives to enhance particular capabilities of LLMs. Extensive experiments demonstrate that the refinement model not only improves the quality of responses but also strengthens their robustness against jailbreak attacks. Code is available at: https://github.com/Huangzisu/query-refinement .

Xiaohua Wang, Muzhao Tian, Yuqi Zeng et al.

Reinforcement Learning from Human Feedback (RLHF) and related alignment paradigms have become central to steering large language models (LLMs) and multimodal large language models (MLLMs) toward human-preferred behaviors. However, these approaches introduce a systemic vulnerability: reward hacking, where models exploit imperfections in learned reward signals to maximize proxy objectives without fulfilling true task intent. As models scale and optimization intensifies, such exploitation manifests as verbosity bias, sycophancy, hallucinated justification, benchmark overfitting, and, in multimodal settings, perception--reasoning decoupling and evaluator manipulation. Recent evidence further suggests that seemingly benign shortcut behaviors can generalize into broader forms of misalignment, including deception and strategic gaming of oversight mechanisms. In this survey, we propose the Proxy Compression Hypothesis (PCH) as a unifying framework for understanding reward hacking. We formalize reward hacking as an emergent consequence of optimizing expressive policies against compressed reward representations of high-dimensional human objectives. Under this view, reward hacking arises from the interaction of objective compression, optimization amplification, and evaluator--policy co-adaptation. This perspective unifies empirical phenomena across RLHF, RLAIF, and RLVR regimes, and explains how local shortcut learning can generalize into broader forms of misalignment, including deception and strategic manipulation of oversight mechanisms. We further organize detection and mitigation strategies according to how they intervene on compression, amplification, or co-adaptation dynamics. By framing reward hacking as a structural instability of proxy-based alignment under scale, we highlight open challenges in scalable oversight, multimodal grounding, and agentic autonomy.

Xiaohua Wang, Jiakang Yuan, Zisu Huang et al.

A long-standing goal of the research community is to develop highly interactive LLM-based dialogue agents. Recent research focuses on optimizing policies based on fixed offline logs (Static Context RL) or using a prompt-based simulator (Interactive RL). In this work, we theoretically show that both paradigms are fundamentally limited by context distribution shift--a mismatch between dialogue histories observed during training and those encountered in real conversations. This shift compounds quadratically over turns and severely degrades dialogue quality. Specifically, we attribute this shift to two distinct sources: (i) policy-induced shift, arising from training on static histories rather than self-generated trajectories; and (ii) simulator-induced shift, stemming from discrepancies between simulated and real human behaviors. To address these challenges, we propose Calibrated Interactive RL, a unified framework that couples interactive RL with simulator alignment. By aligning the simulator with human interaction patterns, our approach reduces the sim-to-real gap and mitigates compounding distribution shifts. Experiments across multiple dialogue tasks confirm our theoretical analysis: (i) Interactive RL significantly outperforms the Static Context baseline by mitigating policy distribution shift; and (ii) calibrating simulators with our alignment method further bridges the sim-to-real gap, yielding state-of-the-art downstream performance.

Jingwen Xu, Yiyang Lu, Zisu Huang et al.

Training LLMs for code-related tasks typically depends on high-quality code-documentation pairs, which are costly to curate and often scarce for niche programming languages. We introduce BatCoder, a self-supervised reinforcement learning framework designed to jointly optimize code generation and documentation production. BatCoder employs a back-translation strategy: a documentation is first generated from code, and then the generated documentation is used to reconstruct the original code. The semantic similarity between the original and reconstructed code serves as an implicit reward, enabling reinforcement learning to improve the model's performance both in generating code from documentation and vice versa. This approach allows models to be trained using only code, substantially increasing the available training examples. Evaluated on HumanEval and MBPP with a 7B model, BatCoder achieved 83.5% and 81.0% pass@1, outperforming strong open-source baselines. Moreover, the framework demonstrates consistent scaling with respect to both training corpus size and model capacity.

Zisu Huang, Jingwen Xu, Yifan Yang et al.

Language agents increasingly improve by reusing \emph{skills} -- structured procedural artifacts distilled from past experience. In particular, \emph{domain-level} and \emph{model-generated} skills are especially promising. They offer fast adaptation within a domain by encoding domain-specific recurring procedures, and they scale beyond labor-intensive hand-crafting. However, while extraction methods continue to proliferate, understanding remains limited, with no comprehensive study spanning the full skill lifecycle -- \textbf{experience generation}, \textbf{skill extraction}, and \textbf{skill consumption} -- to ask whether such skills actually work, when they work, and what makes them succeed or fail. To close this gap, we build a utility-grounded evaluation framework that provides systematic experimental results across extractors and target agents, covering five diverse agentic task domains. We find that model-generated skills are beneficial on average but exhibit non-trivial negative transfer, and that neither extractors nor targets behave uniformly. A model can be a strong extractor yet a weak consumer, or vice versa, with skill utility independent of model scale or baseline task strength. To explain these patterns, we then dissect each lifecycle stage in depth, analyzing how experience composition shapes skill quality, what properties characterize useful skills, and how the same skill transfers across different consumers. Finally, we translate these findings into a concrete \emph{meta-skill} that guides skill extraction toward the features tied to actual utility, which consistently improves skill quality across domains and substantially reduces negative transfer.

Yifan Yang, Ziyang Gong, Weiquan Huang et al.

Agent skills today are hand-crafted, generated one-shot, or evolved through loosely controlled self-revision, none of which behaves like a deep-learning optimizer for the skill, and none of which reliably improves over its starting point under feedback. We argue the skill should instead be trained as the external state of a frozen agent, with the same discipline that makes weight-space optimization reproducible. SkillOpt is, to our knowledge, the first systematic controllable text-space optimizer for agent skills: a separate optimizer model turns scored rollouts into bounded add/delete/replace edits on a single skill document, and an edit is accepted only when it strictly improves a held-out validation score. A textual learning-rate budget, rejected-edit buffer, and epoch-wise slow/meta update make skill training stable while adding zero inference-time model calls at deployment. Across six benchmarks, seven target models, and three execution harnesses (direct chat, Codex, Claude Code), SkillOpt is best or tied on all 52 evaluated (model, benchmark, harness) cells and beats every per-cell competitor among human, one-shot LLM, Trace2Skill, TextGrad, GEPA, and EvoSkill skills. On GPT-5.5 it lifts the average no-skill accuracy by +23.5 points in direct chat, by +24.8 inside the Codex agentic loop, and by +19.1 inside Claude Code. Transfer experiments further show that optimized skill artifacts retain value when moved across model scales, between Codex and Claude Code execution environments, and to a nearby math benchmark without further optimization.

Qi Qian, Chengsong Huang, Jingwen Xu et al.

The rapid proliferation of benchmarks for evaluating large language models (LLMs) has created an urgent need for systematic methods to assess benchmark quality itself. We propose Benchmark^2, a comprehensive framework comprising three complementary metrics: (1) Cross-Benchmark Ranking Consistency, measuring whether a benchmark produces model rankings aligned with peer benchmarks; (2) Discriminability Score, quantifying a benchmark's ability to differentiate between models; and (3) Capability Alignment Deviation, identifying problematic instances where stronger models fail but weaker models succeed within the same model family. We conduct extensive experiments across 15 benchmarks spanning mathematics, reasoning, and knowledge domains, evaluating 11 LLMs across four model families. Our analysis reveals significant quality variations among existing benchmarks and demonstrates that selective benchmark construction based on our metrics can achieve comparable evaluation performance with substantially reduced test sets.

Jingwen Xu, Yiyang Lu, Changze Lv et al.

Existing code similarity metrics, such as BLEU, CodeBLEU, and TSED, largely rely on surface-level string overlap or abstract syntax tree structures, and often fail to capture deeper semantic relationships between programs.We propose CSSG (Code Similarity using Semantic Graphs), a novel metric that leverages program dependence graphs to explicitly model control dependencies and variable interactions, providing a semantics-aware representation of code.Experiments on the CodeContests+ dataset show that CSSG consistently outperforms existing metrics in distinguishing more similar code from less similar code under both monolingual and cross-lingual settings, demonstrating that dependency-aware graph representations offer a more effective alternative to surface-level or syntax-based similarity measures.

Muzhao Tian, Zisu Huang, Xiaohua Wang et al.

As LLM-based agents are increasingly used in long-term interactions, cumulative memory is critical for enabling personalization and maintaining stylistic consistency. However, most existing systems adopt an ``all-or-nothing'' approach to memory usage: incorporating all relevant past information can lead to \textit{Memory Anchoring}, where the agent is trapped by past interactions, while excluding memory entirely results in under-utilization and the loss of important interaction history. We show that an agent's reliance on memory can be modeled as an explicit and user-controllable dimension. We first introduce a behavioral metric of memory dependence to quantify the influence of past interactions on current outputs. We then propose \textbf{Stee}rable \textbf{M}emory Agent, \texttt{SteeM}, a framework that allows users to dynamically regulate memory reliance, ranging from a fresh-start mode that promotes innovation to a high-fidelity mode that closely follows interaction history. Experiments across different scenarios demonstrate that our approach consistently outperforms conventional prompting and rigid memory masking strategies, yielding a more nuanced and effective control for personalized human-agent collaboration.

Yuanzhe Shen, Zisu Huang, Zhengyuan Wang et al.

As LLM-based agents are deployed in increasingly complex real-world settings, existing benchmarks underrepresent key challenges such as enforcing global constraints, coordinating multi-tool reasoning, and adapting to evolving user behavior over long, multi-turn interactions. To bridge this gap, we introduce \textbf{TRIP-Bench}, a long-horizon benchmark grounded in realistic travel-planning scenarios. TRIP-Bench leverages real-world data, offers 18 curated tools and 40+ travel requirements, and supports automated evaluation. It includes splits of varying difficulty; the hard split emphasizes long and ambiguous interactions, style shifts, feasibility changes, and iterative version revision. Dialogues span up to 15 user turns, can involve 150+ tool calls, and may exceed 200k tokens of context. Experiments show that even advanced models achieve at most 50\% success on the easy split, with performance dropping below 10\% on hard subsets. We further propose \textbf{GTPO}, an online multi-turn reinforcement learning method with specialized reward normalization and reward differencing. Applied to Qwen2.5-32B-Instruct, GTPO improves constraint satisfaction and interaction robustness, outperforming Gemini-3-Pro in our evaluation. We expect TRIP-Bench to advance practical long-horizon interactive agents, and GTPO to provide an effective online RL recipe for robust long-horizon training.

Muling Wu, Qi Qian, Wenhao Liu et al.

Large Language Models (LLMs) have achieved remarkable performance across various reasoning tasks, yet post-training is constrained by inefficient sample utilization and inflexible difficulty samples processing. To address these limitations, we propose Customized Curriculum Learning (CCL), a novel framework with two key innovations. First, we introduce model-adaptive difficulty definition that customizes curriculum datasets based on each model's individual capabilities rather than using predefined difficulty metrics. Second, we develop "Guided Prompting," which dynamically reduces sample difficulty through strategic hints, enabling effective utilization of challenging samples that would otherwise degrade performance. Comprehensive experiments on supervised fine-tuning and reinforcement learning demonstrate that CCL significantly outperforms uniform training approaches across five mathematical reasoning benchmarks, confirming its effectiveness across both paradigms in enhancing sample utilization and model performance.

Yuanzhe Shen, Zisu Huang, Zhengkang Guo et al.

The rapid advancement of large language models (LLMs) has driven their adoption across diverse domains, yet their ability to generate harmful content poses significant safety challenges. While extensive research has focused on mitigating harmful outputs, such efforts often come at the cost of excessively rejecting harmless prompts. Striking a balance among safety, over-refusal, and utility remains a critical challenge. In this work, we introduce IntentionReasoner, a novel safeguard mechanism that leverages a dedicated guard model to perform intent reasoning, multi-level safety classification, and query rewriting to neutralize potentially harmful intent in edge-case queries. Specifically, we first construct a comprehensive dataset comprising approximately 163,000 queries, each annotated with intent reasoning, safety labels, and rewritten versions. Supervised fine-tuning is then applied to equip the guard model with foundational capabilities in format adherence, intent analysis, and safe rewriting. Finally, we apply a tailored multi-reward optimization strategy that integrates rule-based heuristics and reward model signals within a reinforcement learning framework to further enhance performance. Extensive experiments show that IntentionReasoner excels in multiple safeguard benchmarks, generation quality evaluations, and jailbreak attack scenarios, significantly enhancing safety while effectively reducing over-refusal rates and improving the quality of responses.

Zhengkang Guo, Wenhao Liu, Mingchen Xie et al.

Large language models (LLMs) are increasingly expected to tackle complex tasks, driven by their expanding applications and users' growing proficiency in crafting sophisticated prompts. However, as the number of explicitly stated requirements increases (particularly more than 10 constraints), LLMs often struggle to accurately follow such complex instructions, which limits their applicability in complex real-world scenarios. To the best of our knowledge, existing datasets do not exceed 10 constraints per instance. To address this challenge, we propose RECAST, an efficient and scalable framework for synthesizing datasets where each example incorporates far more constraints than those in existing benchmarks, aiming to challenge and extend the boundaries of models' ability to follow complex instructions. These constraints are extracted from real-world prompt-response pairs to ensure practical relevance. Using this framework, we construct RECAST-30K, a large-scale, high-quality dataset comprising 30k instances spanning 19 constraint types. Experimental results demonstrate that models finetuned on RECAST-30K substantially improve in following complex instructions while maintaining their general capabilities without degradation. Moreover, RECAST enables automatic verification of constraint satisfaction via rule-based validators for quantitative constraints and LLM-based validators for qualitative ones; the verifiability provided by RECAST enables the design of reward functions for reinforcement learning, which further boosts model performance on complex and challenging tasks.

Yuanzhe Shen, Yide Liu, Zisu Huang et al.

Large language models (LLMs) demonstrate remarkable performance across diverse tasks, yet their effectiveness frequently depends on costly commercial APIs or cloud services. Model selection thus entails a critical trade-off between performance and cost: high-performing LLMs typically incur substantial expenses, whereas budget-friendly small language models (SLMs) are constrained by limited capabilities. Current research primarily proposes two routing strategies: pre-generation routing and cascade routing. Both approaches have distinct characteristics, with cascade routing typically offering superior cost-effectiveness and accuracy despite its higher latency. To further address the limitations of both approaches, we introduce SATER, a dual-mode compatible approach that fine-tunes models through shortest-response preference optimization and a confidence-aware rejection mechanism. SATER significantly reduces redundant outputs and response times, while improving both the performance of pre-generation routing and the efficiency of cascade routing. Experiments across three SLMs and six datasets, varying in type and complexity, demonstrate that SATER achieves comparable performance while consistently reducing computational costs by over 50\% and cascade latency by over 80\%.

Zhibo Xu, Jianhao Zhu, Jingwen Xu et al.

The primary goal of traditional federated learning is to protect data privacy by enabling distributed edge devices to collaboratively train a shared global model while keeping raw data decentralized at local clients. The rise of large language models (LLMs) has introduced new challenges in distributed systems, as their substantial computational requirements and the need for specialized expertise raise critical concerns about protecting intellectual property (IP). This highlights the need for a federated learning approach that can safeguard both sensitive data and proprietary models. To tackle this challenge, we propose FedQSN, a federated learning approach that leverages random masking to obscure a subnetwork of model parameters and applies quantization to the remaining parameters. Consequently, the server transmits only a privacy-preserving proxy of the global model to clients during each communication round, thus enhancing the model's confidentiality. Experimental results across various models and tasks demonstrate that our approach not only maintains strong model performance in federated learning settings but also achieves enhanced protection of model parameters compared to baseline methods.