arXiv subjects

Leitian Tao

Publications and source records attributed to Leitian Tao.

13 recordsLinked to original sources

arXiv2026

ExecCritic: Learn to Test, Test to Improve for Coding Agents

Execution feedback can guide coding agents toward correct repository repairs, but only when the tests capture the behavior requested by the issue. Agent-generated tests can encode incomplete or incorrect behavioral targets; when the same trajectory writes both the patch and the test, their errors can agree and create false confidence. We introduce ExecCritic, combining a test--verify--revise scaffold with a role-specific reinforcement learning recipe for training agents within it. The scaffold separates test construction from source-code repair: a Test agent independently generates repository-native tests, a fail-closed harness qualifies and freezes them, and a Repair agent revises source code from their execution feedback without changing the tests. Both roles use Qwen-3.5-35B-A3B as the backbone and are trained separately. In Learn to Test, the Test agent learns to produce behaviorally valid tests that distinguish correct from incorrect patches. In Test to Improve, the Repair agent learns both direct task resolution and feedback-guided revision. On SWE-bench Verified, test quality determines whether feedback helps: holding the base Repair agent fixed, tests from the base Test agent reduce resolved rate from a no-test baseline of 61.2% to 57.3%, whereas tests from GPT-5.6-sol raise it to 65.3%. Role-specific post-training raises the Qwen Test agent's Base-to-Gold success from 22.2% to 62.2%; composing the two post-trained Qwen agents reaches 72.6%, an 11.4-point gain over the original no-test baseline without stronger-model or Oracle feedback at evaluation time. Code is publicly available at https://github.com/MSR-Orchard/execcritic.

cs.AI↗

arXiv2026

TRACE: Turn-level Reward Assignment via Credit Estimation for Long-Horizon Agents

Multi-turn agents solve complex tasks through extended sequences of tool interactions before producing a final answer, making credit assignment a fundamental challenge during post-training. Outcome rewards provide reliable supervision for short-horizon reasoning, but become sparse and high-variance as trajectories grow to tens or hundreds of tool calls. They can also be misleading: a failed rollout may contain many useful actions that move the agent closer to the goal, yet outcome-only training assigns them the same negative advantage as the eventual mistake. We propose TRACE (Turn-level Reward Assignment via Credit Estimation), a dense credit-assignment method for agentic reinforcement learning. TRACE represents rollouts as state transitions at tool-call boundaries, obtains gold-answer log-probabilities from a frozen reference model, transforms them into log-ratio state values, and derives per-action rewards as Temporal-Difference changes in those values. This requires no additional critic or process-label training, and its one-step log-ratio TD component telescopes across redundant tool calls. On long-horizon complex search, TRACE substantially improves base-model tool-use ability using pure RL, without a cold-start supervised fine-tuning stage, an agentic mid-training stage, or training on live-web data. On the closed-web BrowseComp-Plus benchmark, it raises Qwen3-4B from $7.2$ to $35.6$ and Qwen3-30B-A3B from $8.4$ to $42.6$. The learned search behavior also transfers to open-web benchmarks, and the learning curves show earlier improvement and faster convergence during RL training.

cs.LG↗

arXiv2025

Hybrid Reinforcement: When Reward Is Sparse, It's Better to Be Dense

Post-training for reasoning of large language models (LLMs) increasingly relies on verifiable rewards: deterministic checkers that provide 0-1 correctness signals. While reliable, such binary feedback is brittle--many tasks admit partially correct or alternative answers that verifiers under-credit, and the resulting all-or-nothing supervision limits learning. Reward models offer richer, continuous feedback, which can serve as a complementary supervisory signal to verifiers. We introduce HERO (Hybrid Ensemble Reward Optimization), a reinforcement learning framework that integrates verifier signals with reward-model scores in a structured way. HERO employs stratified normalization to bound reward-model scores within verifier-defined groups, preserving correctness while refining quality distinctions, and variance-aware weighting to emphasize challenging prompts where dense signals matter most. Across diverse mathematical reasoning benchmarks, HERO consistently outperforms RM-only and verifier-only baselines, with strong gains on both verifiable and hard-to-verify tasks. Our results show that hybrid reward design retains the stability of verifiers while leveraging the nuance of reward models to advance reasoning.

cs.CL↗

arXiv2025

RESTRAIN: From Spurious Votes to Signals -- Self-Driven RL with Self-Penalization

Reinforcement learning with human-annotated data has boosted chain-of-thought reasoning in large reasoning models, but these gains come at high costs in labeled data while faltering on harder tasks. A natural next step is experience-driven learning, where models improve without curated labels by adapting to unlabeled data. We introduce RESTRAIN (REinforcement learning with Self-restraint), a self-penalizing RL framework that converts the absence of gold labels into a useful learning signal. Instead of overcommitting to spurious majority votes, RESTRAIN exploits signals from the model's entire answer distribution: penalizing overconfident rollouts and low-consistency examples while preserving promising reasoning chains. The self-penalization mechanism integrates seamlessly into policy optimization methods such as GRPO, enabling continual self-improvement without supervision. On challenging reasoning benchmarks, RESTRAIN delivers large gains using only unlabeled data. With Qwen3-4B-Base and OctoThinker Hybrid-8B-Base, it improves Pass@1 by up to +140.7 percent on AIME25, +36.2 percent on MMLU_STEM, and +19.6 percent on GPQA-Diamond, nearly matching gold-label training while using no gold labels. These results demonstrate that RESTRAIN establishes a scalable path toward stronger reasoning without gold labels.

cs.CL↗

arXiv2025

Limited Preference Data? Learning Better Reward Model with Latent Space Synthesis

Reward modeling, crucial for aligning large language models (LLMs) with human preferences, is often bottlenecked by the high cost of preference data. Existing textual data synthesis methods are computationally expensive. We propose a novel framework LENS for synthesizing preference data directly in the LLM's latent embedding space. Our method employs a Variational Autoencoder (VAE) to learn a structured latent representation of response embeddings. By performing controlled perturbations in this latent space and decoding back to the embedding space, we efficiently generate diverse, semantically consistent synthetic preference pairs, bypassing costly text generation and annotation. We provide theoretical guarantees that our synthesized pairs approximately preserve original preference ordering and improve reward model generalization. Empirically, our latent-space synthesis significantly outperforms text-based augmentation on standard benchmarks, achieving superior results while being 18x faster in generation and using a 16,000x smaller model. Our work offers a scalable and effective alternative for enhancing reward modeling through efficient data augmentation. Code is publicly available at https://github.com/deeplearning-wisc/lens

cs.CL↗

arXiv2025

The Era of Real-World Human Interaction: RL from User Conversations

We posit that to achieve continual model improvement and multifaceted alignment, future models must learn from natural human interaction. Current conversational models are aligned using pre-annotated, expert-generated human feedback. In this work, we introduce Reinforcement Learning from Human Interaction (RLHI), a paradigm that learns directly from in-the-wild user conversations. We develop two complementary methods: (1) RLHI with User-Guided Rewrites, which revises unsatisfactory model outputs based on users' natural-language follow-up responses, (2) RLHI with User-Based Rewards, which learns via a reward model conditioned on knowledge of the user's long-term interaction history (termed persona). Together, these methods link long-term user personas to turn-level preferences via persona-conditioned preference optimization. Trained on conversations derived from WildChat, both RLHI variants outperform strong baselines in personalization and instruction-following, and similar feedback enhances performance on reasoning benchmarks. These results suggest organic human interaction offers scalable, effective supervision for personalized alignment.

cs.AI↗

arXiv2024

CodeLutra: Boosting LLM Code Generation via Preference-Guided Refinement

Large Language Models (LLMs) have revolutionized code generation but require significant resources and often over-generalize, limiting their task-specific efficiency. Fine-tuning smaller, open-source LLMs provides a cost-effective alternative. However, standard supervised approaches rely only on correct examples, missing valuable insights from failures. We introduce CodeLutra, a framework that leverages both correct and incorrect code attempts. Instead of using only correct solutions, CodeLutra applies iterative preference-based refinement, comparing successful and failed outputs to better approximate desired results. This approach narrows the performance gap with state-of-the-art larger models without requiring massive datasets or auxiliary models. For instance, on a challenging data science coding task, using only 500 samples improved Llama-3-8B's accuracy from 28.2% to 48.6%, approaching GPT-4's level. By learning from both successes and mistakes, CodeLutra provides a scalable and efficient path to high-quality code generation, making smaller open-source models more competitive with leading closed-source alternatives.

cs.CL↗

arXiv2024

Challenges and Future Directions of Data-Centric AI Alignment

As AI systems become increasingly capable and influential, ensuring their alignment with human values, preferences, and goals has become a critical research focus. Current alignment methods primarily focus on designing algorithms and loss functions but often underestimate the crucial role of data. This paper advocates for a shift towards data-centric AI alignment, emphasizing the need to enhance the quality and representativeness of data used in aligning AI systems. In this position paper, we highlight key challenges associated with both human-based and AI-based feedback within the data-centric alignment framework. Through qualitative analysis, we identify multiple sources of unreliability in human feedback, as well as problems related to temporal drift, context dependence, and AI-based feedback failing to capture human values due to inherent model limitations. We propose future research directions, including improved feedback collection practices, robust data-cleaning methodologies, and rigorous feedback verification processes. We call for future research into these critical directions to ensure, addressing gaps that persist in understanding and improving data-centric alignment practices.

cs.CL↗

arXiv2024

Your Weak LLM is Secretly a Strong Teacher for Alignment

The burgeoning capabilities of large language models (LLMs) have underscored the need for alignment to ensure these models act in accordance with human values and intentions. Existing alignment frameworks present constraints either in the form of expensive human effort or high computational costs. This paper explores a promising middle ground, where we employ a weak LLM that is significantly less resource-intensive than top-tier models, yet offers more automation than purely human feedback. We present a systematic study to evaluate and understand weak LLM's ability to generate feedback for alignment. Our empirical findings demonstrate that weak LLMs can provide feedback that rivals or even exceeds that of fully human-annotated data. Our study indicates a minimized impact of model size on feedback efficacy, shedding light on a scalable and sustainable alignment strategy. To deepen our understanding of alignment under weak LLM feedback, we conduct a series of qualitative and quantitative analyses, offering novel insights into the quality discrepancies between human feedback vs. weak LLM feedback.

cs.CL↗

arXiv2023

Activate and Reject: Towards Safe Domain Generalization under Category Shift

Albeit the notable performance on in-domain test points, it is non-trivial for deep neural networks to attain satisfactory accuracy when deploying in the open world, where novel domains and object classes often occur. In this paper, we study a practical problem of Domain Generalization under Category Shift (DGCS), which aims to simultaneously detect unknown-class samples and classify known-class samples in the target domains. Compared to prior DG works, we face two new challenges: 1) how to learn the concept of ``unknown'' during training with only source known-class samples, and 2) how to adapt the source-trained model to unseen environments for safe model deployment. To this end, we propose a novel Activate and Reject (ART) framework to reshape the model's decision boundary to accommodate unknown classes and conduct post hoc modification to further discriminate known and unknown classes using unlabeled test data. Specifically, during training, we promote the response to the unknown by optimizing the unknown probability and then smoothing the overall output to mitigate the overconfidence issue. At test time, we introduce a step-wise online adaptation method that predicts the label by virtue of the cross-domain nearest neighbor and class prototype information without updating the network's parameters or using threshold-based mechanisms. Experiments reveal that ART consistently improves the generalization capability of deep networks on different vision tasks. For image classification, ART improves the H-score by 6.1% on average compared to the previous best method. For object detection and semantic segmentation, we establish new benchmarks and achieve competitive performance.

cs.CV↗

arXiv2023

Non-Parametric Outlier Synthesis

Out-of-distribution (OOD) detection is indispensable for safely deploying machine learning models in the wild. One of the key challenges is that models lack supervision signals from unknown data, and as a result, can produce overconfident predictions on OOD data. Recent work on outlier synthesis modeled the feature space as parametric Gaussian distribution, a strong and restrictive assumption that might not hold in reality. In this paper, we propose a novel framework, Non-Parametric Outlier Synthesis (NPOS), which generates artificial OOD training data and facilitates learning a reliable decision boundary between ID and OOD data. Importantly, our proposed synthesis approach does not make any distributional assumption on the ID embeddings, thereby offering strong flexibility and generality. We show that our synthesis approach can be mathematically interpreted as a rejection sampling framework. Extensive experiments show that NPOS can achieve superior OOD detection performance, outperforming the competitive rivals by a significant margin. Code is publicly available at https://github.com/deeplearning-wisc/npos.

cs.LG↗

arXiv2022

Pyramid Feature Alignment Network for Video Deblurring

Video deblurring remains a challenging task due to various causes of blurring. Traditional methods have considered how to utilize neighboring frames by the single-scale alignment for restoration. However, they typically suffer from misalignment caused by severe blur. In this work, we aim to better utilize neighboring frames with high efficient feature alignment. We propose a Pyramid Feature Alignment Network (PFAN) for video deblurring. First, the multi-scale feature of blurry frames is extracted with the strategy of Structure-to-Detail Downsampling (SDD) before alignment. This downsampling strategy makes the edges sharper, which is helpful for alignment. Then we align the feature at each scale and reconstruct the image at the corresponding scale. This strategy effectively supervises the alignment at each scale, overcoming the problem of propagated errors from the above scales at the alignment stage. To better handle the challenges of complex and large motions, instead of aligning features at each scale separately, lower-scale motion information is used to guide the higher-scale motion estimation. Accordingly, a Cascade Guided Deformable Alignment (CGDA) is proposed to integrate coarse motion into deformable convolution for finer and more accurate alignment. As demonstrated in extensive experiments, our proposed PFAN achieves superior performance with competitive speed compared to the state-of-the-art methods.

cs.CV↗

arXiv2021

Predicate correlation learning for scene graph generation

For a typical Scene Graph Generation (SGG) method, there is often a large gap in the performance of the predicates' head classes and tail classes. This phenomenon is mainly caused by the semantic overlap between different predicates as well as the long-tailed data distribution. In this paper, a Predicate Correlation Learning (PCL) method for SGG is proposed to address the above two problems by taking the correlation between predicates into consideration. To describe the semantic overlap between strong-correlated predicate classes, a Predicate Correlation Matrix (PCM) is defined to quantify the relationship between predicate pairs, which is dynamically updated to remove the matrix's long-tailed bias. In addition, PCM is integrated into a Predicate Correlation Loss function ($L_{PC}$) to reduce discouraging gradients of unannotated classes. The proposed method is evaluated on Visual Genome benchmark, where the performance of the tail classes is significantly improved when built on the existing methods.

cs.CV↗