arXiv subjects

Yu Mi

Publications and source records attributed to Yu Mi.

10 recordsLinked to original sources

arXiv2026

SCOPE: Selective Critical-Step On-Policy Self-Distillation for Agentic Reinforcement Learning

Outcome-based reinforcement learning provides coarse supervision for long-horizon LLM agents: failed trajectories may contain correct steps, while successful trajectories can contain errors. Trajectory-level skills can also shift teacher preferences at unrelated steps, introducing off-target distillation signals. We introduce SCOPE, a framework for Selective Critical-Step On-Policy Self-Distillation. After a supervised warm-up, the policy analyzes its own trajectories, generates step-specific hindsight skills, and distills their guidance only at selected error steps. A bounded signed teacher-student coefficient stabilizes auxiliary updates while preserving positive reinforcement and negative correction. A success-relative pairwise length reward further encourages efficient task completion. Experiments on ALFWorld and WebShop demonstrate consistent improvements across multiple model backbones. With Qwen2.5-3B, SCOPE achieves 97.7% success on ALFWorld and 85.2% on WebShop, outperforming SEED by 5.9 and 6.3 percentage points, respectively. On unseen ALFWorld tasks, SCOPE reaches 89.6% success, improving over GRPO by 23.2 percentage points with 46.7% fewer interactions. Scaling to Qwen3-32B, SCOPE achieves 99.2% success on seen ALFWorld tasks and 94.0% on unseen tasks.

cs.AI↗

arXiv2026

EOPSA: Efficient On-Policy Self-Distilled Safety Alignment

On-Policy Self-Distillation (OPSD) has emerged as a promising paradigm for safety alignment, delivering dense, token-level supervision by distilling from a teacher conditioned on refusal-oriented privileged prompts. However, we reveal that this paradigm suffers from critical inefficiencies that degrade both training efficiency and general reasoning capabilities. Specifically, we diagnose two fundamental bottlenecks: (1) supervisory collapse over extended rollouts, where the teacher's corrective efficacy degrades precipitously as the student's generation prefix lengthens, injecting noisy gradients into late-stage tokens; and (2) gradient dilution from stylistic shifts, where the distillation objective is dominated by safety-irrelevant stylistic discrepancies induced by privileged prompting, washing out genuine safety signals and impairing base reasoning. To resolve these issues, we propose Efficient On-Policy Self-Distilled Safety Alignment (EOPSA), which concentrates computational and gradient budgets exclusively on reliably supervised, safety-critical tokens. EOPSA incorporates two coordinated mechanisms: (i) Adaptive Rollout Scheduling, which dynamically bounds the generation horizon guided by a novel Teacher Rescue Rate (TRR) metric to operate strictly within reliable supervision regimes; and (ii) Selective Distillation, which filters out safety-neutral tokens to restrict gradient updates exclusively to safety-pivotal transitions. Extensive evaluations across reasoning models up to 32B parameters demonstrate that EOPSA slashes rollout computation by $\sim$50% and backpropagates through merely $\sim$2% of tokens, consistently outperforming full-token distillation baselines in both safety compliance and reasoning retention.

cs.AI↗

arXiv2026

Mitigating Cross-Modal Hallucinations by Correcting Modality Misrouting in Omnimodal LLMs

Omnimodal large language models (OLLMs) are susceptible to cross-modal hallucinations under audio-visual interference. We find that, for a substantial fraction of incorrect joint predictions, the question-relevant unimodal branch can still recover the correct answer, suggesting that task-relevant evidence can remain recoverable from the relevant modality even when it is not reflected in the joint prediction. We refer to this recoverable failure pattern as \emph{modality misrouting}. We propose CaRe (Conflict-aware Residual Intervention), a training-free inference-time method for identifying and correcting modality misrouting. CaRe uses a text-only modality probe to identify a candidate modality-specific counterfactual branch, and flags potential misrouting when its prediction conflicts with the original joint prediction. For such cases, CaRe computes a residual direction from the discrepancy between the counterfactual and fused representations and uses it to steer the original multimodal state toward the relevant modality evidence, while retaining both modalities in the original audio-visual inference path. The intervention strength is adapted based on the confidence of the counterfactual branch. Experiments on CMM and AVHBench across two OLLMs show that CaRe consistently outperforms the evaluated inference-time hallucination-mitigation baselines.

cs.AI↗

arXiv2026

Lowest Span Confidence: Zero-Shot Hallucination Detection from a Single LLM Response

Hallucinations in Large Language Models (LLMs), i.e., plausible but non-factual generations, pose a significant challenge to reliable deployment in high-stakes environments. However, many existing hallucination detectors require expensive repeated sampling for consistency checks or access to model-internal states unavailable in common API-based scenarios. To this end, we propose an efficient zero-shot metric called Lowest Span Confidence (LSC) for hallucination detection under minimal resource assumptions. Concretely, LSC evaluates the local confidence of adjacent complete-word spans. By selecting the lowest aggregated confidence across neighboring words whose token widths can vary, LSC captures localized uncertainty associated with factual inconsistency. This boundary-aligned smoothing reduces the global dilution of perplexity and the sensitivity of minimum token probability to isolated noise. Our main evaluation spans four model families {Llama-2, Qwen2.5, Gemma-2, Mistral} and seven benchmarks {NQ, TriviaQA, SQuAD, CoQA, HotpotQA, RAGTruth, FELM}. Additional analyses examine word reconstruction, span width, and the role of adjacency in preserving local confidence. Across these settings, LSC is competitive with methods that use multiple responses or model-internal information while requiring only one response and its output token probabilities, without training a separate detector or using an auxiliary model.

cs.CL↗

arXiv2026

Ignorance or Incompetence? Constructing Knowledge-Gated, Verifiable Tasks for LLM Agents

Professional agent tasks often depend on conventions that are absent from public corpora, yet benchmarks rarely control whether an agent has access to those conventions. We introduce a knowledge-gated task-construction protocol that separates a task instruction from a compact artefact containing private conventions, reference tables, and utility operators. Construction-time provenance, byte-identical task instructions across the provided- and withheld-artefact conditions, leak audits, and executable witnesses make dependence on the artefact explicit and testable. Across fifteen calibration tasks, one frontier agent configuration achieves a 68.0% pass rate with the artefact and 0% without it; on one task, a plausible but incorrect artefact also yields 0% across five trials. Deterministic solvers and rule corpora provide exact ground truth for structured tasks, while named criterion-level rubrics support outputs that cannot be checked by a single executable oracle. A configuration-relative calibration screen retains seven tasks satisfying our five-trial empirical knowledge-gating screen. These experiments validate the behavior of the construction protocol; they do not establish that the retained tasks improve post-training. We publicly release part of the task suite and supporting tooling at https://github.com/DatagridsAI/Knowledge-Gated-Task-Construction.

cs.AI↗

arXiv2026

LFRAG: Layout-oriented Fine-grained Retrieval-Augmented Generation on Multimodal Document Understanding

Multimodal Retrieval-Augmented Generation (RAG) has emerged as an effective paradigm for enhancing Large Language Models (LLMs) with external knowledge. However, existing multimodal RAG systems predominantly rely on coarse-grained page-level retrieval, which fails to capture fine-grained semantic and layout structures in visually rich documents, thereby compromising retrieval accuracy and leading to redundant context in downstream tasks. To address these issues, we propose Layout-oriented Fine-grained Retrieval-Augmented Generation (LFRAG), a novel framework that advances multimodal RAG from page-level to block-level retrieval. We perform layout segmentation to construct semantically coherent fine-grained retrieval units and design a semantic-layout fusion encoder that integrates local semantics with global context via cross-attention. With block-level late interaction retrieval, LFRAG enables precise query-content alignment and reduces irrelevant content for downstream generation. To enable rigorous evaluation, we construct LFDocQA, a large-scale benchmark with block-level annotations spanning diverse document types, designed to assess both multimodal document retrieval and question answering with greater granularity than existing datasets. Extensive experiments on LFDocQA demonstrate that LFRAG achieves state-of-the-art performance on retrieval tasks, outperforms the best baseline by 7.20% in answer accuracy, and reduces token consumption by 73.07% in generation tasks, confirming LFRAG as an accurate and efficient framework for multimodal RAG over visually rich documents. Our code and datasets will be released soon.

cs.IR↗

arXiv2025

Tree embedding based mapping system for low-latency mobile applications in multi-access networks

Low-latency applications like AR/VR and online gaming need fast, stable connections. New technologies such as V2X, LEO satellites, and 6G bring unique challenges in mobility management. Traditional solutions based on centralized or distributed anchors often fall short in supporting rapid mobility due to inefficient routing, low versatility, and insufficient multi-access support. In this paper, we design a new end-to-end system for tracking multi-connected mobile devices at scale and optimizing performance for latency-sensitive, highly dynamic applications. Our system, based on the locator/ID separation principle, extends to multi-access networks without requiring specialized routers or caching. Using a novel tree embedding-based overlay, we enable fast session setup while allowing endpoints to directly handle mobility between them. Evaluation with real network data shows our solution cuts connection latency to 7.42% inflation over the shortest path, compared to LISP's 359\% due to cache misses. It also significantly reduces location update overhead and disruption time during mobility.

cs.NI↗

arXiv2025

scAgent: Universal Single-Cell Annotation via a LLM Agent

Cell type annotation is critical for understanding cellular heterogeneity. Based on single-cell RNA-seq data and deep learning models, good progress has been made in annotating a fixed number of cell types within a specific tissue. However, universal cell annotation, which can generalize across tissues, discover novel cell types, and extend to novel cell types, remains less explored. To fill this gap, this paper proposes scAgent, a universal cell annotation framework based on Large Language Models (LLMs). scAgent can identify cell types and discover novel cell types in diverse tissues; furthermore, it is data efficient to learn novel cell types. Experimental studies in 160 cell types and 35 tissues demonstrate the superior performance of scAgent in general cell-type annotation, novel cell discovery, and extensibility to novel cell type.

cs.CL↗

arXiv2024

A Survey on LoRA of Large Language Models

Low-Rank Adaptation~(LoRA), which updates the dense neural network layers with pluggable low-rank matrices, is one of the best performed parameter efficient fine-tuning paradigms. Furthermore, it has significant advantages in cross-task generalization and privacy-preserving. Hence, LoRA has gained much attention recently, and the number of related literature demonstrates exponential growth. It is necessary to conduct a comprehensive overview of the current progress on LoRA. This survey categorizes and reviews the progress from the perspectives of (1) downstream adaptation improving variants that improve LoRA's performance on downstream tasks; (2) cross-task generalization methods that mix multiple LoRA plugins to achieve cross-task generalization; (3) efficiency-improving methods that boost the computation-efficiency of LoRA; (4) data privacy-preserving methods that use LoRA in federated learning; (5) application. Besides, this survey also discusses the future directions in this field. At last, we provide a Github page~\footnote{\href{https://github.com/ZJU-LLMs/Awesome-LoRAs.git}{https://github.com/ZJU-LLMs/Awesome-LoRAs.git}} for readers to check the updates and initiate discussions on this survey paper.

cs.LG↗

arXiv2024

FinSQL: Model-Agnostic LLMs-based Text-to-SQL Framework for Financial Analysis

Text-to-SQL, which provides zero-code interface for operating relational databases, has gained much attention in financial analysis; because, financial professionals may not well-skilled in SQL programming. However, until now, there is no practical Text-to-SQL benchmark dataset for financial analysis, and existing Text-to-SQL methods have not considered the unique characteristics of databases in financial applications, such as commonly existing wide tables. To address these issues, we collect a practical Text-to-SQL benchmark dataset and propose a model-agnostic Large Language Model (LLMs)-based Text-to-SQL framework for financial analysis. The benchmark dataset, BULL, is collected from the practical financial analysis business of Hundsun Technologies Inc., including databases for fund, stock, and macro economy. Besides, the proposed LLMs-based Text-to-SQL framework, FinSQL, provides a systematic treatment for financial Text-to-SQL from the perspectives of prompt construction, parameter-efficient fine-tuning and output calibration. Extensive experimental results on BULL demonstrate that FinSQL achieves the state-of-the-art Text-to-SQL performance at a small cost; furthermore, FinSQL can bring up to 36.64% performance improvement in scenarios requiring few-shot cross-database model transfer.

cs.CL↗