arXiv subjects

Huaisheng Zhu

Publications and source records attributed to Huaisheng Zhu.

At least 19 recordsLinked to original sources

arXiv2026

EvalBuilder: Can LLMs Build Efficient Benchmarks?

Large language model (LLM) agents are increasingly used for long-horizon research and engineering, yet evaluating these capabilities remains expensive, time-consuming, and heavily dependent on human expertise. This motivates a natural question: can LLM agents themselves construct efficient benchmarks that preserve the conclusions of costly reference evaluations? We study this capability by formulating automatic benchmark development as a meta-evaluation problem, where an agent is given a task and an expensive reference benchmark and requested to construct a lower-cost proxy benchmark that faithfully distinguishes among candidate models and solutions. Based on this formulation, we introduce EvalBuilder, a framework for evaluating such benchmark-generation capabilities along three key dimensions: fidelity, efficiency, and robustness. Specifically, we assess whether generated benchmarks preserve model rankings, generalize to held-out models and freshly collected responses, and provide well-calibrated and discriminative scores while substantially reducing evaluation cost. We further provide an iterative environment in which agents can refine their benchmarks under a fixed time budget. Our framework provides a systematic testbed and benchmark for understanding when current LLM agents can develop useful evaluators and how benchmark construction itself can be increasingly automated.

cs.AI↗

arXiv2026

Rethinking the Evaluation of Harness Evolution for Agents

Harness evolution is an iterative search procedure that repeatedly evaluates and revises candidate harnesses used for LLM agents using task feedback. We revisit the evaluation of such automatic harness evolution procedures and identify two fundamental issues in the protocol. First, prior work does not compare these approaches with simple task-level search baselines under matched feedback and inference budgets. Second, prior work searches for harness configurations using verification signals (e.g., unit test cases) drawn from the same benchmarks on which it reports the final performance of the evolved harnesses, violating the standard separation between training and test data. To address this, we compare automatic harness evolution with simple test-time scaling and discovery baselines under comparable feedback and inference budgets, and evaluate evolved harnesses on held-out tasks to assess generalization. Following prior work, we experiment on Terminal-Bench 2.1 and find that automatic harness evolution fails to outperform simple test-time scaling methods both with and without test cases, and exhibits limited generalization. However, we find that long-horizon games are a promising setting for automatic harness evolution, as they are difficult enough to leave headroom, rely heavily on adaptation to out-of-distribution dynamics, and provide granular feedback by design. In these settings, task-specific harness evolution improves over the search baseline by 80% on ARC-AGI-3 and by 11% on EdgeBench games under matched budgets. Together, these findings highlight the need for matched-budget baselines and held-out evaluation to distinguish genuine harness improvements from benchmark-specific search and overfitting, and point to a more careful characterization of when automatic harness evolution is actually useful. Our code is available at https://github.com/rethinking-harness-evolution.

cs.AI↗

arXiv2026

Long-Horizon Scaling: How Model Capabilities Shape the Returns to Computation

Long-horizon agents improve solutions through sustained interaction, execution, and task feedback. Scaling studies relate performance to resources and capabilities, yet how existing capabilities shape returns to extended interaction remains less understood. To address this gap, we analyze AutoLab and EdgeBench, two long-horizon benchmarks. We find that starting performance and subsequent growth are associated with different capabilities: within a task category, similar early scores can precede different later gains. To formalize this finding, we model capability-time scaling with category-specific logistic power laws shared across models. Fitted to early trajectories, these curves extrapolate the observed models' category-average scores to later computation. However, rising average scores mask narrowing improvement opportunities: later gains concentrate among fewer improving models. High final scores and continued improvement also have distinct capability profiles. Predicted mean gains estimate each model's fraction of improving tasks; averaging these estimates forecasts the average share of improving models. These uneven returns motivate deciding whether a specific run should continue. We therefore derive a continuation policy to save time and compute with limited score loss. The policy conditions growth predictions on the run's observed progress and weighs immediate and delayed gains against computation costs. In replay with training and price calibration based on other models' histories, the policy saves roughly one-third of full-run time, with relative score losses of 2.4% on AutoLab individual runs and 3.3% on EdgeBench published mean curves. Our repository is available at https://github.com/Chihaya-Anon-chan/long-horizon-scaling.

cs.LG↗

arXiv2026

Beyond Final Scores: A Systematic Evaluation of Agents for Long-Horizon AI Research and Development

Autonomous agents are increasingly capable of improving models, systems, and other technical artifacts through long-horizon experimentation. To understand the current state of this capability, however, evaluation must go beyond final scores, which neither reveal where progress is gained or lost nor indicate whether accumulated experience improves later decisions. We therefore present a systematic evaluation of seven frontier models on 36 long-horizon tasks based on a new framework that uses rule-based metrics to characterize within-run behavior through Solution Framing, Execution, and Feedback Control and controlled comparisons to assess experience reuse within and across tasks. The results show that current agents operate more like engineering optimizers than fully autonomous researchers: they can formulate and implement practical solutions, but their performance varies substantially across runs, their strongest solutions mainly adapt or combine established techniques, and genuine methodological novelty remains rare. Detailed analysis reveals that observed performance is shaped by multiple factors, including distinct process bottlenecks behind similar final outcomes, experience reuse that can help or mislead subsequent decisions, and harness designs that affect performance stability. These findings suggest concrete directions for improving model training, inference-time strategies, experience management, and harness design.

cs.AI↗

arXiv2026

Distilling Temporal Search and Reasoning: Evolving LLMs for Future Prediction via Harness-Assisted Efficient Data Synthesis

Future event prediction carries broad social impact yet remains challenging. SOTA approaches augment LLMs with external agent frameworks whose predictive capability vanishes once the harness is removed. While recent Tool-Integrated Reasoning (TIR) internalizes deep search for multi-hop retrieval of facts, forecasting further demands temporal search and reasoning over historical trends and dynamic shifts. The key obstacle is data: historical queries induce temporal leakage that degrades forecasting into retrieval. Prior works either freeze information gathering with static observations, or rely on rejection sampling or unresolved fresh queries that discard vast amounts of data, degrading synthesis efficiency. We propose a time-truncation harness that enforces a temporal cut-off at every turn, enabling TIR-style sampling from historical events, reducing temporal leakage and reliance of rejection sampling or unsolved queries, increasing the sampling efficiency. We further build a large-scale corpus and a process-based metric and show that our harness naturally induces a broader temporal breadth of search and raises the proportion of high-quality data, further increasing the efficiency and reducing the reliance on complex rubrics. Distillation experiments show that students trained on harness-intervened data achieve the best performance, demonstrating harness-assisted model evolving that turns higher quality temporal search and reasoning data into a parametric advancement of the students.

cs.AI↗

arXiv2026

Co-Harness: Co-Evolving Harnesses and Model Weights for LLM Agents

Post-training agents for automated AI research requires optimizing not only model parameters, but also the runtime harness that shapes how research trajectories are generated, evaluated, and learned from. Existing pipelines typically train models under a fixed harness, including prompts, tools, skills, middleware, and memory, while leaving the data-generating process outside the optimization objective. This creates a mismatch between model updates and the static scaffolding that determines trajectory quality. We introduce Co-Harness, a framework that jointly optimizes the agent harness and model parameters during post-training. Co-Harness alternates between harness optimization and model optimization. An LLM-based HarnessCritic analyzes failed trajectories, identifies harness-level failure modes, and proposes validated local updates. The model is then fine-tuned on high-quality trajectories generated by the improved harness, distilling effective scaffolding into model parameters. A 200+ hour autonomous case study further shows that Co-Harness can recover from system crashes, improve inference efficiency, and discover ensemble strategies without human intervention. These results suggest that joint harness and model optimization is an effective way to improve agents beyond fixed-harness post-training.

cs.AI↗

arXiv2026

Optimizing Token Choice for Code Watermarking: An RL Approach

Protecting intellectual property on LLM-generated code necessitates effective watermarking systems that can operate within code's highly structured, syntactically constrained nature. In this work, we introduce CodeTracer, an innovative adaptive code watermarking framework underpinned by a novel reinforcement learning training paradigm. At its core, CodeTracer features a policy-driven approach that utilizes a parameterized model to intelligently bias token choices during next-token prediction. This strategy ensures that embedded watermarks maintain code functionality while exhibiting subtle yet statistically detectable deviations from typical token distributions. To facilitate policy learning, we devise a comprehensive reward system that seamlessly integrates execution feedback with watermark embedding signals, balancing process-level and outcome-level rewards. Additionally, we employ Gumbel Top-k reparameterization to enable gradient-based optimization of discrete watermarking decisions. Extensive comparative evaluations demonstrate CodeTracer's significant superiority over state-of-the-art baselines in both watermark detectability and the preservation of generated code's functionality. Our code is available at https://github.com/TimeLovercc/CodeTracer.

cs.CR↗

arXiv2026

Meta-Reinforcement Learning with Self-Reflection for Agentic Search

This paper introduces MR-Search, an in-context meta reinforcement learning (RL) formulation for agentic search with self-reflection. Instead of optimizing a policy within a single independent episode with sparse rewards, MR-Search trains a policy that conditions on past episodes and adapts its search strategy across episodes. MR-Search learns to learn a search strategy with self-reflection, allowing search agents to improve in-context exploration at test-time. Specifically, MR-Search performs cross-episode exploration by generating explicit self-reflections after each episode and leveraging them as additional context to guide subsequent attempts, thereby promoting more effective exploration during test-time. We further introduce a multi-turn RL algorithm that estimates a dense relative advantage at the turn level, enabling fine-grained credit assignment on each episode. Empirical results across various benchmarks demonstrate the advantages of MR-Search over baselines based RL, showing strong generalization and relative improvements of 9.2% to 19.3% across eight benchmarks. Our code and data are available at https://github.com/tengxiao1/MR-Search.

cs.LG↗

arXiv2026

Simple Denoising Diffusion Language Models

Recent Uniform State Diffusion Models (USDMs), initialized from a uniform prior, offer the promise of fast text generation due to their inherent self-correction ability compared to masked diffusion models. However, they still rely on complex loss formulations with additional computational overhead, which hinders scalability. In this work, we explore a simplified denoising-based loss for USDMs that optimizes only noise-replaced tokens, stabilizing training while matching the performance of prior methods with more complex objectives. In addition, we introduce an efficient regularization term to mitigate corruption toward uniform output distributions, which further improves performance. We demonstrate the effectiveness and efficiency of our simple and improved loss formulations by pretraining models on widely used text datasets for USDMs. More importantly, our conclusions scale to larger models, showing strong potential for large-scale training.

cs.LG↗

arXiv2025

The Path of Self-Evolving Large Language Models: Achieving Data-Efficient Learning via Intrinsic Feedback

Reinforcement learning (RL) has demonstrated potential in enhancing the reasoning capabilities of large language models (LLMs), but such training typically demands substantial efforts in creating and annotating data. In this work, we explore improving LLMs through RL with minimal data. Our approach alternates between the LLM proposing a task and then attempting to solve it. To minimize data dependency, we introduce two novel mechanisms grounded in self-awareness: (1) self-aware difficulty prediction, where the model learns to assess task difficulty relative to its own abilities and prioritize challenging yet solvable tasks, and (2) self-aware limit breaking, where the model recognizes when a task is beyond its capability boundary and proactively requests external data to break through that limit. Extensive experiments on nine benchmarks showing a 53.8% relative improvement with less than 1.2% extra data demonstrate the efficacy of self-aware RL and underscore the promise of self-evolving agent training.

cs.CL↗

arXiv2025

Multimodal LLMs as Customized Reward Models for Text-to-Image Generation

We introduce LLaVA-Reward, an efficient reward model designed to automatically evaluate text-to-image (T2I) generations across multiple perspectives, leveraging pretrained multimodal large language models (MLLMs). Existing MLLM-based approaches require instruction-following data for supervised fine-tuning and evaluate generation quality on analyzing text response, which is time-consuming and difficult to train. To address this problem, we propose LLaVA-Reward, which directly utilizes the hidden states of MLLMs given text-image pairs. To enhance the bidirectional interaction between visual and textual representations in decoder-only MLLMs, we further propose adding a Skip-connection Cross Attention (SkipCA) module. This design enhances text-image correlation reasoning by connecting early-layer visual features with later-layer hidden representations. In addition, LLaVA-Reward supports different types of preference data for efficient fine-tuning, including paired preference data and unpaired data. We train LLaVA-Reward on four evaluation perspectives: text-image alignment, fidelity/artifact, safety, and overall ranking. Empirical results demonstrate that LLaVA-Reward outperforms conventional and MLLM-based methods in generating human-aligned scores for automatic evaluations and inference-time scaling in text-to-image generations.

cs.CV↗

arXiv2025

Enhancing Diffusion Posterior Sampling for Inverse Problems by Integrating Crafted Measurements

Diffusion models have emerged as a powerful foundation model for visual generations. With an appropriate sampling process, it can effectively serve as a generative prior for solving general inverse problems. Current posterior sampling-based methods take the measurement (i.e., degraded image sample) into the posterior sampling to infer the distribution of the target data (i.e., clean image sample). However, in this manner, we show that high-frequency information can be prematurely introduced during the early stages, which could induce larger posterior estimate errors during restoration sampling. To address this observation, we first reveal that forming the log-posterior gradient with the noisy measurement ( i.e., noisy measurement from a diffusion forward process) instead of the clean one can benefit the early posterior sampling. Consequently, we propose a novel diffusion posterior sampling method DPS-CM, which incorporates a Crafted Measurement (i.e., noisy measurement crafted by a reverse denoising process, rather than constructed from the diffusion forward process) to form the posterior estimate. This integration aims to mitigate the misalignment with the diffusion prior caused by cumulative posterior estimate errors. Experimental results demonstrate that our approach significantly improves the overall capacity to solve general and noisy inverse problems, such as Gaussian deblurring, super-resolution, inpainting, nonlinear deblurring, and tasks with Poisson noise, relative to existing approaches. Code is available at: https://github.com/sjz5202/DPS-CM.

cs.CV↗

arXiv2024

Cal-DPO: Calibrated Direct Preference Optimization for Language Model Alignment

We study the problem of aligning large language models (LLMs) with human preference data. Contrastive preference optimization has shown promising results in aligning LLMs with available preference data by optimizing the implicit reward associated with the policy. However, the contrastive objective focuses mainly on the relative values of implicit rewards associated with two responses while ignoring their actual values, resulting in suboptimal alignment with human preferences. To address this limitation, we propose calibrated direct preference optimization (Cal-DPO), a simple yet effective algorithm. We show that substantial improvement in alignment with the given preferences can be achieved simply by calibrating the implicit reward to ensure that the learned implicit rewards are comparable in scale to the ground-truth rewards. We demonstrate the theoretical advantages of Cal-DPO over existing approaches. The results of our experiments on a variety of standard benchmarks show that Cal-DPO remarkably improves off-the-shelf methods.

cs.LG↗

arXiv2024

GeomCLIP: Contrastive Geometry-Text Pre-training for Molecules

Pretraining molecular representations is crucial for drug and material discovery. Recent methods focus on learning representations from geometric structures, effectively capturing 3D position information. Yet, they overlook the rich information in biomedical texts, which detail molecules' properties and substructures. With this in mind, we set up a data collection effort for 200K pairs of ground-state geometric structures and biomedical texts, resulting in a PubChem3D dataset. Based on this dataset, we propose the GeomCLIP framework to enhance for multi-modal representation learning from molecular structures and biomedical text. During pre-training, we design two types of tasks, i.e., multimodal representation alignment and unimodal denoising pretraining, to align the 3D geometric encoder with textual information and, at the same time, preserve its original representation power. Experimental results show the effectiveness of GeomCLIP in various tasks such as molecular property prediction, zero-shot text-molecule retrieval, and 3D molecule captioning. Our code and collected dataset are available at \url{https://github.com/xiaocui3737/GeomCLIP}

cs.LG↗

arXiv2024

How to Leverage Demonstration Data in Alignment for Large Language Model? A Self-Imitation Learning Perspective

This paper introduces a novel generalized self-imitation learning ($\textbf{GSIL}$) framework, which effectively and efficiently aligns large language models with offline demonstration data. We develop $\textbf{GSIL}$ by deriving a surrogate objective of imitation learning with density ratio estimates, facilitating the use of self-generated data and optimizing the imitation learning objective with simple classification losses. $\textbf{GSIL}$ eliminates the need for complex adversarial training in standard imitation learning, achieving lightweight and efficient fine-tuning for large language models. In addition, $\textbf{GSIL}$ encompasses a family of offline losses parameterized by a general class of convex functions for density ratio estimation and enables a unified view for alignment with demonstration data. Extensive experiments show that $\textbf{GSIL}$ consistently and significantly outperforms baselines in many challenging benchmarks, such as coding (HuamnEval), mathematical reasoning (GSM8K) and instruction-following benchmark (MT-Bench).

cs.CL↗

arXiv2024

Learning Fair Models without Sensitive Attributes: A Generative Approach

Most existing fair classifiers rely on sensitive attributes to achieve fairness. However, for many scenarios, we cannot obtain sensitive attributes due to privacy and legal issues. The lack of sensitive attributes challenges many existing fair classifiers. Though we lack sensitive attributes, for many applications, there usually exists features or information of various formats that are relevant to sensitive attributes. For example, purchase history of a person can reflect his or her race, which would help for learning fair classifiers on race. However, the work on exploring relevant features for learning fair models without sensitive attributes is rather limited. Therefore, in this paper, we study a novel problem of learning fair models without sensitive attributes by exploring relevant features. We propose a probabilistic generative framework to effectively estimate the sensitive attribute from the training data with relevant features in various formats and utilize the estimated sensitive attribute information to learn fair models. Experimental results on real-world datasets show the effectiveness of our framework in terms of both accuracy and fairness.

cs.LG↗

arXiv2024

3M-Diffusion: Latent Multi-Modal Diffusion for Language-Guided Molecular Structure Generation

Generating molecular structures with desired properties is a critical task with broad applications in drug discovery and materials design. We propose 3M-Diffusion, a novel multi-modal molecular graph generation method, to generate diverse, ideally novel molecular structures with desired properties. 3M-Diffusion encodes molecular graphs into a graph latent space which it then aligns with the text space learned by encoder-based LLMs from textual descriptions. It then reconstructs the molecular structure and atomic attributes based on the given text descriptions using the molecule decoder. It then learns a probabilistic mapping from the text space to the latent molecular graph space using a diffusion model. The results of our extensive experiments on several datasets demonstrate that 3M-Diffusion can generate high-quality, novel and diverse molecular graphs that semantically match the textual description provided.

cs.LG↗

arXiv2024

MolBind: Multimodal Alignment of Language, Molecules, and Proteins

Recent advancements in biology and chemistry have leveraged multi-modal learning, integrating molecules and their natural language descriptions to enhance drug discovery. However, current pre-training frameworks are limited to two modalities, and designing a unified network to process different modalities (e.g., natural language, 2D molecular graphs, 3D molecular conformations, and 3D proteins) remains challenging due to inherent gaps among them. In this work, we propose MolBind, a framework that trains encoders for multiple modalities through contrastive learning, mapping all modalities to a shared feature space for multi-modal semantic alignment. To facilitate effective pre-training of MolBind on multiple modalities, we also build and collect a high-quality dataset with four modalities, MolBind-M4, including graph-language, conformation-language, graph-conformation, and conformation-protein paired data. MolBind shows superior zero-shot learning performance across a wide range of tasks, demonstrating its strong capability of capturing the underlying semantics of multiple modalities.

cs.LG↗