Searcharxiv⌕ Search

arXiv subjects

Yutao Xie

Publications and source records attributed to Yutao Xie.

At least 19 recordsLinked to original sources
arXiv2026

RMSWeb: Reflection, Failure-Mode Mining, and Salvage-DS for Web Agent Reinforcement Learning

Compact web agents can reduce deployment cost, but training them poses challenges in both data collection and post-SFT reinforcement learning (RL). Successful trajectories are expensive to collect and often contain inefficient detours. After supervised fine-tuning (SFT), full trajectory corpora are dominated by routine states; moreover, when group-relative RL is applied to web actions, inadequately designed action-level rewards can yield weak or misleading relative updates, while groups rejected as unsuitable for such updates receive no fallback learning signal. We present RMSWeb, a three-part recipe for Qwen3-VL-Instruct at 8B and 32B. Reflection-conditioned retries increase collection yield and shorten successful trajectories; failure-mode mining concentrates offline RL on critical states exposed by the SFT policy; and Salvage-DS combines an action-semantic polarized reward, contrast-and-competence-gated dynamic sampling, and an action-only anchor for rejected groups. Policies trained with reflection-collected data use up to 19.7% fewer action steps on solved tasks. On WebVoyager, Online-Mind2Web, and WebTailBench, RMSWeb improves over SFT by 2.4-7.0 points at 8B and 1.2-7.7 points at 32B. Our 8B model also achieves the strongest reported Online-Mind2Web result among similarly sized open-weight models in our comparison and a leading reported accuracy-cost trade-off on WebVoyager and WebTailBench, with the caveat that external evaluation protocols differ.

cs.AI↗
arXiv2026

ConPress: Learning Efficient Reasoning from Multi-Question Contextual Pressure

Large reasoning models (LRMs) typically solve reasoning-intensive tasks by generating long chain-of-thought (CoT) traces, leading to substantial inference overhead. We identify a reproducible inference-time phenomenon, termed Self-Compression: when multiple independent and answerable questions are presented within a single prompt, the model spontaneously produces shorter reasoning traces for each question. This phenomenon arises from multi-question contextual pressure during generation and consistently manifests across models and benchmarks. Building on this observation, we propose ConPress (Learning from Contextual Pressure), a lightweight self-supervised fine-tuning approach. ConPress constructs multi-question prompts to induce self-compression, samples the resulting model outputs, and parses and filters per-question traces to obtain concise yet correct reasoning trajectories. These trajectories are directly used for supervised fine-tuning, internalizing compressed reasoning behavior in single-question settings without external teachers, manual pruning, or reinforcement learning. With only 8k fine-tuning examples, ConPress reduces reasoning token usage by 59% on MATH500 and 33% on AIME25, while maintaining competitive accuracy.

cs.CL↗
arXiv2026

PIKA: Expert-Level Synthetic Datasets for Post-Training Alignment from Scratch

High-quality instruction data is critical for LLM alignment, yet existing open-source datasets often lack efficiency, requiring hundreds of thousands of examples to approach proprietary performance. In this work, we find that beyond the widely recognized importance of prompt-response quality, prompt difficulty itself plays a critical role in driving alignment gains. Motivated by this observation, we introduce PiKa, a data-efficient family of expert-level alignment datasets that concentrates supervision on high-difficulty instructions. The PiKa-SFT dataset contains only 30k examples, an order of magnitude fewer than state-of-the-art open datasets like Magpie-Pro. Despite its small size, fine-tuning Llama-3-8B-Base on PiKa-SFT even outperforms the official Llama-3-8B-Instruct model trained on over 10M proprietary examples on widely used benchmarks such as AlpacaEval 2.0 and Arena-Hard. We also validate the generalizability of PiKa across the Qwen2.5 series (0.5B-7B), consistently surpassing their official instruction-tuned counterparts. Additionally, we provide 30k high-quality preference optimization examples to further enhance alignment. Our results demonstrate that promising alignment is achievable with significantly reduced data, democratizing access for resource-constrained research. Our code and data will be available at https://github.com/SJY8460/PiKa.

cs.CL↗
arXiv2026

IsoCompute Playbook: Optimally Scaling Sampling Compute for LLM RL

While scaling laws guide compute allocation for LLM pre-training, analogous prescriptions for reinforcement learning (RL) post-training of large language models (LLMs) remain poorly understood. We study the compute-optimal allocation of sampling compute for on-policy RL methods in LLMs, framing scaling as a compute-constrained optimization over three resources: parallel rollouts per problem, number of problems per batch, and number of update steps. We find that the compute-optimal number of parallel rollouts per problem increases predictably with compute budget and then saturates. This trend holds across both easy and hard problems, though driven by different mechanisms: solution sharpening on easy problems and coverage expansion on hard problems. We further show that increasing the number of parallel rollouts mitigates interference across problems, while the number of problems per batch primarily affects training stability and can be chosen within a broad range. Validated across base models and data distributions, our results recast RL scaling laws as prescriptive allocation rules and provide practical guidance for compute-efficient LLM RL post-training.

cs.LG↗
arXiv2026

TIPS: Turn-Level Information-Potential Reward Shaping for Search-Augmented LLMs

Search-augmented large language models (LLMs) trained with reinforcement learning (RL) have achieved strong results on open-domain question answering (QA), but training still remains a significant challenge. The optimization is often unstable due to sparse rewards and difficult credit assignments across reasoning and tool calls. To address this, we introduce Turn-Level Information Potential Reward Shaping (TIPS), a simple framework that assigns dense, turn-level rewards to each reasoning + tool-call segment based on the increased likelihood of the correct answer under a teacher model. By leveraging the potential-based reward shaping, TIPS offers fine-grained and policy-invariant guidance that overcomes the limitations of outcome-only optimization. Evaluated on seven QA benchmarks, TIPS consistently outperforms GRPO/PPO baselines and substantially improves training stability. For instance, with a Qwen-2.5 7B Instruct model, TIPS improves the average Exact Match score by 11.8% and F1 by 13.6% relative to PPO. Our results demonstrate that turn-level information-potential reward shaping provides an effective and general solution to sparse-reward credit assignment for multi-turn LLM reasoning.

cs.CL↗
arXiv2026

Lost in Stories: Consistency Bugs in Long Story Generation by LLMs

What happens when a storyteller forgets its own story? Large Language Models (LLMs) can now generate narratives spanning tens of thousands of words, but they often fail to maintain consistency throughout. When generating long-form narratives, these models can contradict their own established facts, character traits, and world rules. Existing story generation benchmarks focus mainly on plot quality and fluency, leaving consistency errors largely unexplored. To address this gap, we present ConStory-Bench, a benchmark designed to evaluate narrative consistency in long-form story generation. It contains 2,000 prompts across four task scenarios and defines a taxonomy of five error categories with 19 fine-grained subtypes. We also develop ConStory-Checker, an automated pipeline that detects contradictions and grounds each judgment in explicit textual evidence. Evaluating a range of LLMs through five research questions, we find that consistency errors show clear tendencies: they are most common in factual and temporal dimensions, tend to appear around the middle of narratives, occur in text segments with higher token-level entropy, and certain error types tend to co-occur. These findings can inform future efforts to improve consistency in long-form narrative generation. Our project page is available at https://picrew.github.io/constory-bench.github.io/.

cs.CL↗
arXiv2026

Towards Better Code Understanding in Decoder-Only Models with Contrastive Learning

Recent advances in large-scale code generation models have led to remarkable progress in producing high-quality code. These models are trained in a self-supervised manner on extensive unlabeled code corpora using a decoder-only architecture. However, despite their generative strength, decoder-only models often exhibit limited performance on code understanding tasks such as code search and clone detection, primarily due to their generation-oriented training objectives. While training large encoder-only models from scratch on massive code datasets can improve understanding ability but remains computationally expensive and time-consuming. In this paper, we explore a more efficient alternative by transferring knowledge from pre-trained decoder-only code generation models to code understanding tasks. We investigate how decoder-only architectures can be effectively adapted to learn discriminative and semantically meaningful code representations. To this end, we propose CL4D, a contrastive learning framework tailored to strengthen the representation capabilities of decoder-only models. Extensive experiments on multiple benchmark datasets demonstrate that CL4D achieves competitive or superior performance compared to existing methods on representative code understanding tasks, including code search and clone detection. Further analysis reveals that CL4D substantially improves the semantic alignment of code representations by reducing the distance between semantically similar code snippets. These findings highlight the feasibility of leveraging decoder-only models as a unified backbone for both code generation and understanding.

cs.SE↗
arXiv2026

Beyond Rejection Sampling: Trajectory Fusion for Scaling Mathematical Reasoning

Large language models (LLMs) have made impressive strides in mathematical reasoning, often fine-tuned using rejection sampling that retains only correct reasoning trajectories. While effective, this paradigm treats supervision as a binary filter that systematically excludes teacher-generated errors, leaving a gap in how reasoning failures are modeled during training. In this paper, we propose TrajFusion, a fine-tuning strategy that reframes rejection sampling as a structured supervision construction process. Specifically, TrajFusion forms fused trajectories that explicitly model trial-and-error reasoning by interleaving selected incorrect trajectories with reflection prompts and correct trajectories. The length of each fused sample is adaptively controlled based on the frequency and diversity of teacher errors, providing richer supervision for challenging problems while safely reducing to vanilla rejection sampling fine-tuning (RFT) when error signals are uninformative. TrajFusion requires no changes to the architecture or training objective. Extensive experiments across multiple math benchmarks demonstrate that TrajFusion consistently outperforms RFT, particularly on challenging and long-form reasoning problems.

cs.CL↗
arXiv2026

Recent Advances of End-to-End Video Coding Technologies for AVS Standard Development

Video coding standards are essential to enable the interoperability and widespread adoption of efficient video compression technologies. In pursuit of greater video compression efficiency, the AVS video coding working group launched the standardization exploration of end-to-end intelligent video coding, establishing the AVS End-to-End Intelligent Video Coding Exploration Model (AVS-EEM) project. A core design principle of AVS-EEM is its focus on practical deployment, featuring inherently low computational complexity and requiring strict adherence to the common test conditions of conventional video coding. This paper details the development history of AVS-EEM and provides a systematic introduction to its key technical framework, covering model architectures, training strategies, and inference optimizations. These innovations have collectively driven the project's rapid performance evolution, enabling continuous and significant gains under strict complexity constraints. Through over two years of iterative refinement and collaborative effort, the coding performance of AVS-EEM has seen substantial improvement. Experimental results demonstrate that its latest model achieves superior compression efficiency compared to the conventional AVS3 reference software, marking a significant step toward a deployable intelligent video coding standard.

eess.IV↗
arXiv2026

Does Motion Intensity Impair Cognition in HCI? The Critical Role of Physical Motion-Visual Target Directional Congruency

Human-computer interaction (HCI) increasingly occurs in motion-rich environments. The ability to accurately and rapidly respond to directional visual cues is critical in these contexts. How whole-body motion and individual differences affect human perception and reaction to these directional cues is therefore a key, yet an underexplored question for HCI. This study used a 6-DOF motion platform to measure task performance on a visual direction judgment task. We analyzed performance by decomposing the complex motion into two distinct components: a task-irrelevant lateral interference component and a task-aligned directional congruency component. Results indicate that increased motion intensity lengthened reaction times. This effect was primarily driven by the lateral interference component, and this detrimental impact was disproportionately amplified for individuals with high motion sickness susceptibility. Conversely, directional congruency, where motion direction matched the visual cue, improved performance for all participants. These findings suggest that motion's impact on cognition is not monolithic, and that system design for mobile HCI can be informed by strategies that actively shape motion, such as minimizing lateral interference while maximizing directional congruency.

cs.HC↗
arXiv2025

K2-Think: A Parameter-Efficient Reasoning System

K2-Think is a reasoning system that achieves state-of-the-art performance with a 32B parameter model, matching or surpassing much larger models like GPT-OSS 120B and DeepSeek v3.1. Built on the Qwen2.5 base model, our system shows that smaller models can compete at the highest levels by combining advanced post-training and test-time computation techniques. The approach is based on six key technical pillars: Long Chain-of-thought Supervised Finetuning, Reinforcement Learning with Verifiable Rewards (RLVR), Agentic planning prior to reasoning, Test-time Scaling, Speculative Decoding, and Inference-optimized Hardware, all using publicly available open-source datasets. K2-Think excels in mathematical reasoning, achieving state-of-the-art scores on public benchmarks for open-source models, while also performing strongly in other areas such as Code and Science. Our results confirm that a more parameter-efficient model like K2-Think 32B can compete with state-of-the-art systems through an integrated post-training recipe that includes long chain-of-thought training and strategic inference-time enhancements, making open-source reasoning systems more accessible and affordable. K2-Think is freely available at k2think.ai, offering best-in-class inference speeds of over 2,000 tokens per second per request via the Cerebras Wafer-Scale Engine.

cs.LG↗
arXiv2025

Revisiting Reinforcement Learning for LLM Reasoning from A Cross-Domain Perspective

Reinforcement learning (RL) has emerged as a promising approach to improve large language model (LLM) reasoning, yet most open efforts focus narrowly on math and code, limiting our understanding of its broader applicability to general reasoning. A key challenge lies in the lack of reliable, scalable RL reward signals across diverse reasoning domains. We introduce Guru, a curated RL reasoning corpus of 92K verifiable examples spanning six reasoning domains--Math, Code, Science, Logic, Simulation, and Tabular--each built through domain-specific reward design, deduplication, and filtering to ensure reliability and effectiveness for RL training. Based on Guru, we systematically revisit established findings in RL for LLM reasoning and observe significant variation across domains. For example, while prior work suggests that RL primarily elicits existing knowledge from pretrained models, our results reveal a more nuanced pattern: domains frequently seen during pretraining (Math, Code, Science) easily benefit from cross-domain RL training, while domains with limited pretraining exposure (Logic, Simulation, and Tabular) require in-domain training to achieve meaningful performance gains, suggesting that RL is likely to facilitate genuine skill acquisition. Finally, we present Guru-7B and Guru-32B, two models that achieve state-of-the-art performance among open models RL-trained with publicly available data, outperforming best baselines by 7.9% and 6.7% on our 17-task evaluation suite across six reasoning domains. We also show that our models effectively improve the Pass@k performance of their base models, particularly on complex tasks less likely to appear in pretraining data. We release data, models, training and evaluation code to facilitate general-purpose reasoning at: https://github.com/LLM360/Reasoning360

cs.LG↗
arXiv2024

Improving Code Search with Hard Negative Sampling Based on Fine-tuning

Pre-trained code models have emerged as the state-of-the-art paradigm for code search tasks. The paradigm involves pre-training the model on search-irrelevant tasks such as masked language modeling, followed by the fine-tuning stage, which focuses on the search-relevant task. The typical fine-tuning method is to employ a dual-encoder architecture to encode semantic embeddings of query and code separately, and then calculate their similarity based on the embeddings. However, the typical dual-encoder architecture falls short in modeling token-level interactions between query and code, which limits the capabilities of model. To address this limitation, we introduce a cross-encoder architecture for code search that jointly encodes the concatenation of query and code. We further introduce a Retriever-Ranker (RR) framework that cascades the dual-encoder and cross-encoder to promote the efficiency of evaluation and online serving. Moreover, we present a ranking-based hard negative sampling (PS) method to improve the ability of cross-encoder to distinguish hard negative codes, which further enhances the cascaded RR framework. Experiments on four datasets using three code models demonstrate the superiority of our proposed method. We have made the code available at https://github.com/DongHande/R2PS.

cs.SE↗
arXiv2024

Scaling Laws Behind Code Understanding Model

The scaling law is becoming a fundamental law in many machine learning areas. That is, test error falls off with the power law when increasing training data, model size, and computing resource. However, whether this law is suitable for the task of code understanding is not well studied, and most current language models for code understanding are about 100M parameters, which are relatively "small" compared to large language models. In this paper, we conduct extensive experiments to investigate the scaling law for the code understanding task by varying training data, model size, and computing resource. We validate that the test error of code understanding models falls off with the power law when using larger models, indicating that the scaling law is suitable for the code understanding task. Besides, we apply different scales of models to two downstream code understanding tasks, and find that the performance increases with larger scale of models. Finally, we train a large-scale code understanding model named CoLSBERT with 1.5B parameters on a large dataset using more computing resource, which outperforms previous work by a large margin. We will release our code and the CoLSBERT model when our paper is published.

cs.SE↗
arXiv2023

Survey of Code Search Based on Deep Learning

Code writing is repetitive and predictable, inspiring us to develop various code intelligence techniques. This survey focuses on code search, that is, to retrieve code that matches a given query by effectively capturing the semantic similarity between the query and code. Deep learning, being able to extract complex semantics information, has achieved great success in this field. Recently, various deep learning methods, such as graph neural networks and pretraining models, have been applied to code search with significant progress. Deep learning is now the leading paradigm for code search. In this survey, we provide a comprehensive overview of deep learning-based code search. We review the existing deep learning-based code search framework which maps query/code to vectors and measures their similarity. Furthermore, we propose a new taxonomy to illustrate the state-of-the-art deep learning-based code search in a three-steps process: query semantics modeling, code semantics modeling, and matching modeling which involves the deep learning model training. Finally, we suggest potential avenues for future research in this promising field.

cs.SE↗
arXiv2023

AcademicGPT: Empowering Academic Research

Large Language Models (LLMs) have demonstrated exceptional capabilities across various natural language processing tasks. Yet, many of these advanced LLMs are tailored for broad, general-purpose applications. In this technical report, we introduce AcademicGPT, designed specifically to empower academic research. AcademicGPT is a continual training model derived from LLaMA2-70B. Our training corpus mainly consists of academic papers, thesis, content from some academic domain, high-quality Chinese data and others. While it may not be extensive in data scale, AcademicGPT marks our initial venture into a domain-specific GPT tailored for research area. We evaluate AcademicGPT on several established public benchmarks such as MMLU and CEval, as well as on some specialized academic benchmarks like PubMedQA, SCIEval, and our newly-created ComputerScienceQA, to demonstrate its ability from general knowledge ability, to Chinese ability, and to academic ability. Building upon AcademicGPT's foundation model, we also developed several applications catered to the academic area, including General Academic Question Answering, AI-assisted Paper Reading, Paper Review, and AI-assisted Title and Abstract Generation.

cs.CL↗
arXiv2023

Bridging Code Semantic and LLMs: Semantic Chain-of-Thought Prompting for Code Generation

Large language models (LLMs) have showcased remarkable prowess in code generation. However, automated code generation is still challenging since it requires a high-level semantic mapping between natural language requirements and codes. Most existing LLMs-based approaches for code generation rely on decoder-only causal language models often treate codes merely as plain text tokens, i.e., feeding the requirements as a prompt input, and outputing code as flat sequence of tokens, potentially missing the rich semantic features inherent in source code. To bridge this gap, this paper proposes the "Semantic Chain-of-Thought" approach to intruduce semantic information of code, named SeCoT. Our motivation is that the semantic information of the source code (\eg data flow and control flow) describes more precise program execution behavior, intention and function. By guiding LLM consider and integrate semantic information, we can achieve a more granular understanding and representation of code, enhancing code generation accuracy. Meanwhile, while traditional techniques leveraging such semantic information require complex static or dynamic code analysis to obtain features such as data flow and control flow, SeCoT demonstrates that this process can be fully automated via the intrinsic capabilities of LLMs (i.e., in-context learning), while being generalizable and applicable to challenging domains. While SeCoT can be applied with different LLMs, this paper focuses on the powerful GPT-style models: ChatGPT(close-source model) and WizardCoder(open-source model). The experimental study on three popular DL benchmarks (i.e., HumanEval, HumanEval-ET and MBPP) shows that SeCoT can achieves state-of-the-art performance, greatly improving the potential for large models and code generation.

cs.CL↗
arXiv2023

Towards Effective and Compact Contextual Representation for Conformer Transducer Speech Recognition Systems

Current ASR systems are mainly trained and evaluated at the utterance level. Long range cross utterance context can be incorporated. A key task is to derive a suitable compact representation of the most relevant history contexts. In contrast to previous researches based on either LSTM-RNN encoded histories that attenuate the information from longer range contexts, or frame level concatenation of transformer context embeddings, in this paper compact low-dimensional cross utterance contextual features are learned in the Conformer-Transducer Encoder using specially designed attention pooling layers that are applied over efficiently cached preceding utterances history vectors. Experiments on the 1000-hr Gigaspeech corpus demonstrate that the proposed contextualized streaming Conformer-Transducers outperform the baseline using utterance internal context only with statistically significant WER reductions of 0.7% to 0.5% absolute (4.3% to 3.1% relative) on the dev and test data.

eess.AS↗