Searcharxiv⌕ Search

arXiv subjects

Xiaoxiao Li

Publications and source records attributed to Xiaoxiao Li.

At least 55 records · Page 3Linked to original sources
arXiv2026

Spend Less, Reason Better: Budget-Aware Value Tree Search for LLM Agents

Test-time scaling has become a dominant paradigm for improving LLM agent reliability, yet current approaches treat compute as an abundant resource, allowing agents to exhaust token and tool budgets on redundant steps or dead-end trajectories. Existing budget-aware methods either require expensive fine-tuning or rely on coarse, trajectory-level heuristics that cannot intervene mid-execution. We propose the Budget-Aware Value Tree (BAVT), a training-free inference-time framework that models multi-hop reasoning as a dynamic search tree guided by step-level value estimation within a single LLM backbone. Another key innovation is a budget-conditioned node selection mechanism that uses the remaining resource ratio as a natural scaling exponent over node values, providing a principled, parameter-free transition from broad exploration to greedy exploitation as the budget depletes. To combat the well-known overconfidence of LLM self-evaluation, BAVT employs a residual value predictor that scores relative progress rather than absolute state quality, enabling reliable pruning of uninformative or redundant tool calls. We further provide a theoretical convergence guarantee, proving that BAVT reaches a terminal answer with probability at least $1-ε$ under an explicit finite budget bound. Extensive evaluations on four multi-hop QA benchmarks across two model families demonstrate that BAVT consistently outperforms parallel sampling baselines. Most notably, BAVT under strict low-budget constraints surpasses baseline performance at $4\times$ the resource allocation, establishing that intelligent budget management fundamentally outperforms brute-force compute scaling.

cs.LG↗
arXiv2026

See-in-Pairs: Reference Image-Guided Comparative Vision-Language Models for Medical Diagnosis

Medical image diagnosis is challenging because many diseases resemble normal anatomy and exhibit substantial interpatient variability. Clinicians routinely rely on comparative diagnosis, such as referencing cross-patient healthy control images to identify subtle but clinically meaningful abnormalities. Although healthy reference images are abundant in practice, existing medical vision-language models (VLMs) primarily operate in a single-image or single-series setting and lack explicit mechanisms for comparative diagnosis. This work investigates whether incorporating clinically motivated comparison can enhance VLM performance. We show that providing VLMs with both a query image and a matched healthy reference image, accompanied by cross-patient comparative prompts, significantly improves diagnostic performance. This performance can be further augmented by lightweight supervised fine-tuning (SFT) on a small amount of data. At the same time, we evaluate multiple strategies for selecting reference images, including random sampling, demographic attribute matching, embedding-based retrieval, and cross-center selection, and find consistently strong performance across all settings. Finally, we investigate why comparative diagnosis is effective theoretically, and observe improved sample efficiency and tighter alignment between visual and textual representations. Our findings highlight the clinical relevance of comparison-based diagnosis, provide practical strategies for incorporating reference images into VLMs, and demonstrate improved performance across diverse medical imaging tasks.

cs.CV↗
arXiv2026

Communication Compression for Distributed Learning with Aggregate and Server-Guided Feedback

Distributed learning, particularly Federated Learning (FL), faces a significant bottleneck in the communication cost, particularly the uplink transmission of client-to-server updates, which is often constrained by asymmetric bandwidth limits at the edge. Biased compression techniques are effective in practice, but require error feedback mechanisms to provide theoretical guarantees and to ensure convergence when compression is aggressive. Standard error feedback, however, relies on client-specific control variates, which violates user privacy and is incompatible with stateless clients common in large-scale FL. This paper proposes two novel frameworks that enable biased compression without client-side state or control variates. The first, Compressed Aggregate Feedback (CAFe), uses the globally aggregated update from the previous round as a shared control variate for all clients. The second, Server-Guided Compressed Aggregate Feedback (CAFe-S), extends this idea to scenarios where the server possesses a small private dataset; it generates a server-guided candidate update to be used as a more accurate predictor. We consider Distributed Gradient Descent (DGD) as a representative algorithm and analytically prove CAFe's superiority to Distributed Compressed Gradient Descent (DCGD) with biased compression in the non-convex regime with bounded gradient dissimilarity. We further prove that CAFe-S converges to a stationary point, with a rate that improves as the server's data become more representative. Experimental results in FL scenarios validate the superiority of our approaches over existing compression schemes.

cs.LG↗
arXiv2026

Token Hidden Reward: Steering Exploration-Exploitation in Group Relative Deep Reinforcement Learning

Reinforcement learning with verifiable rewards has significantly advanced the reasoning capabilities of large language models, yet how to explicitly steer training toward exploration or exploitation remains an open problem. We introduce Token Hidden Reward (THR), a token-level metric that quantifies each token's influence on the likelihood of correct responses under Group Relative Policy Optimization (GRPO). We find that training dynamics are dominated by a small subset of tokens with high absolute THR values. Most interestingly, tokens with positive THR strengthen confidence in correct outputs, thus favoring exploitation, while tokens with negative THR preserve probability mass for alternative outputs, enabling exploration. This insight suggests a natural intervention: a THR-guided reweighting algorithm that modulates GRPO's learning signals to explicitly bias training toward exploitation or exploration. We validate the efficacy of this algorithm on diverse math reasoning benchmarks. By amplifying tokens with positive THR value and weakening negative ones, our algorithm improves greedy-decoding accuracy, favoring exploitation. The reverse strategy yields consistent gains in Pass@K accuracy, favoring exploration. We further demonstrate that our algorithm integrates seamlessly with other RL objectives such as GSPO and generalizes across architectures including Llama. These findings establish THR as a principled and fine-grained mechanism for dynamically controlling exploration and exploitation in RL-tuned LLMs, providing new tools for targeted fine-tuning in reasoning-intensive applications.

cs.LG↗
arXiv2026

Rethinking LoRA for Data Heterogeneous Federated Learning: Subspace and State Alignment

Low-Rank Adaptation (LoRA) is widely used for federated fine-tuning. Yet under non-IID settings, it can substantially underperform full-parameter fine-tuning. Through with-high-probability robustness analysis, we uncover that this gap can be attributed to two coupled mismatches: (i) update-space mismatch, where clients optimize in a low-rank subspace but aggregation occurs in the full space; and (ii) optimizer-state mismatch, where unsynchronized adaptive states amplify drift across rounds. We propose FedGaLore, which combines client-side GaLore-style gradient-subspace optimization with server-side drift-robust synchronization of projected second-moment states via spectral shared-signal extraction, to address this challenge. Across NLU, vision, and NLG benchmarks, FedGaLore improves robustness and accuracy over state-of-the-art federated LoRA baselines in non-IID settings.

cs.LG↗
arXiv2026

On Group Relative Policy Optimization Collapse in Agent Search: The Lazy Likelihood-Displacement

Tool-integrated (TI) reinforcement learning (RL) enables large language models (LLMs) to perform multi-step reasoning by interacting with external tools such as search engines and retrievers. Group Relative Policy Optimization (GRPO), exemplified by the recent Search-R1, offers fast convergence and a value-free formulation that makes it appealing for this setting, yet consistently suffers from training collapse. We identify Lazy Likelihood Displacement (LLD), a systematic reduction or stagnation in the likelihood of both correct and incorrect responses, as the core mechanism driving this failure. LLD emerges early and triggers a self-reinforcing LLD Death Spiral, where declining likelihood leads to low-confidence responses, inflating gradients, and ultimately causing collapse. We empirically characterize this process across models on a Search-R1-style, search-integrated question answering task, revealing a consistent three-phase trajectory: early stagnation, steady decay, and accelerated collapse. To address this, we propose a likelihood-preserving regularization LLDS that activates only when a response action's likelihood decreases, and regularizes only the tokens responsible. This fine-grained structure mitigates LLD with minimal interference. Our method stabilizes training, prevents gradient explosion, and yields substantial performance improvements across seven benchmarks, including relative improvements of +45.2% on Qwen2.5-3B and +37.1% on Qwen2.5-7B over vanilla GRPO training. Our results establish LLD as a previously overlooked bottleneck in GRPO-based TIRL and provide a practical path toward stable, scalable training of tool-integrated RL.

cs.CL↗
arXiv2026

When Single-Agent with Skills Replace Multi-Agent Systems and When They Fail

Multi-agent AI systems have proven effective for complex reasoning. These systems are compounded by specialized agents, which collaborate through explicit communication, but incur substantial computational overhead. A natural question arises: can we achieve similar modularity benefits with a single agent that selects from a library of skills? We explore this question by viewing skills as internalized agent behaviors. From this perspective, a multi-agent system can be compiled into an equivalent single-agent system, trading inter-agent communication for skill selection. Our preliminary experiments suggest this approach can substantially reduce token usage and latency while maintaining competitive accuracy on reasoning benchmarks. However, this efficiency raises a deeper question that has received little attention: how does skill selection scale as libraries grow? Drawing on principles from cognitive science, we propose that LLM skill selection exhibits bounded capacity analogous to human decision-making. We investigate the scaling behavior of skill selection and observe a striking pattern. Rather than degrading gradually, selection accuracy remains stable up to a critical library size, then drops sharply, indicating a phase transition reminiscent of capacity limits in human cognition. Furthermore, we find evidence that semantic confusability among similar skills, rather than library size alone, plays a central role in this degradation. This perspective suggests that hierarchical organization, which has long helped humans manage complex choices, may similarly benefit AI systems. Our initial results with hierarchical routing support this hypothesis. This work opens new questions about the fundamental limits of semantic-based skill selection in LLMs and offers a cognitive-grounded framework and practical guidelines for designing scalable skill-based agents.

cs.AI↗
arXiv2025

A Closer Look at Personalized Fine-Tuning in Heterogeneous Federated Learning

Federated Learning (FL) enables decentralized, privacy-preserving model training but struggles to balance global generalization and local personalization due to non-identical data distributions across clients. Personalized Fine-Tuning (PFT), a popular post-hoc solution, fine-tunes the final global model locally but often overfits to skewed client distributions or fails under domain shifts. We propose adapting Linear Probing followed by full Fine-Tuning (LP-FT), a principled centralized strategy for alleviating feature distortion (Kumar et al., 2022), to the FL setting. Through systematic evaluation across seven datasets and six PFT variants, we demonstrate LP-FT's superiority in balancing personalization and generalization. Our analysis uncovers federated feature distortion, a phenomenon where local fine-tuning destabilizes globally learned features, and theoretically characterizes how LP-FT mitigates this via phased parameter updates. We further establish conditions (e.g., partial feature overlap, covariate-concept shift) under which LP-FT outperforms standard fine-tuning, offering actionable guidelines for deploying robust personalization in FL.

cs.LG↗
arXiv2025

HookMIL: Revisiting Context Modeling in Multiple Instance Learning for Computational Pathology

Multiple Instance Learning (MIL) has enabled weakly supervised analysis of whole-slide images (WSIs) in computational pathology. However, traditional MIL approaches often lose crucial contextual information, while transformer-based variants, though more expressive, suffer from quadratic complexity and redundant computations. To address these limitations, we propose HookMIL, a context-aware and computationally efficient MIL framework that leverages compact, learnable hook tokens for structured contextual aggregation. These tokens can be initialized from (i) key-patch visual features, (ii) text embeddings from vision-language pathology models, and (iii) spatially grounded features from spatial transcriptomics-vision models. This multimodal initialization enables Hook Tokens to incorporate rich textual and spatial priors, accelerating convergence and enhancing representation quality. During training, Hook tokens interact with instances through bidirectional attention with linear complexity. To further promote specialization, we introduce a Hook Diversity Loss that encourages each token to focus on distinct histopathological patterns. Additionally, a hook-to-hook communication mechanism refines contextual interactions while minimizing redundancy. Extensive experiments on four public pathology datasets demonstrate that HookMIL achieves state-of-the-art performance, with improved computational efficiency and interpretability. Codes are available at https://github.com/lingxitong/HookMIL.

cs.CV↗
arXiv2025

Class-wise Balancing Data Replay for Federated Class-Incremental Learning

Federated Class Incremental Learning (FCIL) aims to collaboratively process continuously increasing incoming tasks across multiple clients. Among various approaches, data replay has become a promising solution, which can alleviate forgetting by reintroducing representative samples from previous tasks. However, their performance is typically limited by class imbalance, both within the replay buffer due to limited global awareness and between replayed and newly arrived classes. To address this issue, we propose a class wise balancing data replay method for FCIL (FedCBDR), which employs a global coordination mechanism for class-level memory construction and reweights the learning objective to alleviate the aforementioned imbalances. Specifically, FedCBDR has two key components: 1) the global-perspective data replay module reconstructs global representations of prior task in a privacy-preserving manner, which then guides a class-aware and importance-sensitive sampling strategy to achieve balanced replay; 2) Subsequently, to handle class imbalance across tasks, the task aware temperature scaling module adaptively adjusts the temperature of logits at both class and instance levels based on task dynamics, which reduces the model's overconfidence in majority classes while enhancing its sensitivity to minority classes. Experimental results verified that FedCBDR achieves balanced class-wise sampling under heterogeneous data distributions and improves generalization under task imbalance between earlier and recent tasks, yielding a 2%-15% Top-1 accuracy improvement over six state-of-the-art methods.

cs.LG↗
arXiv2025

Global Prompt Refinement with Non-Interfering Attention Masking for One-Shot Federated Learning

Federated Prompt Learning (FPL) enables communication-efficient adaptation by tuning lightweight prompts on top of frozen pre-trained models. Existing FPL methods typically rely on global information, which is only available after the second training round, to facilitate collaboration among client models. Therefore, they are inherently dependent on multi-round communication to fully exhibit their strengths. Moreover, existing one-shot federated learning methods typically focus on fitting seen tasks, but lack cross-task generalization. To bridge this gap, we propose the Global Prompt Refinement with Non-Interfering Attention Masking (GPR-NIAM) method for one-shot FPL. The core idea is to design a masking mechanism that restricts excessive interaction between the original text embeddings and the learnable prompt embeddings. GPR-NIAM achieves this through the collaboration of two key modules. Firstly, the attention isolation module suppresses attention from the learnable prompt tokens to the original text tokens, and reweights the reverse attention which preserves generalization across tasks. Secondly, the cross-silo collaborative refinement module integrates decentralized visual knowledge into a unified base and calibrates the global prompt through multi-source cross-modal knowledge alignment, further mitigating the inconsistency caused by data heterogeneity. Extensive experiments conducted on ten benchmark datasets under two tasks show that GPR-NIAM outperforms eight state-of-the-art methods in both class-level and domain-level generalization.

cs.CV↗
arXiv2025

Communication Compression for Distributed Learning without Control Variates

Distributed learning algorithms, such as the ones employed in Federated Learning (FL), require communication compression to reduce the cost of client uploads. The compression methods used in practice are often biased, making error feedback necessary both to achieve convergence under aggressive compression and to provide theoretical convergence guarantees. However, error feedback requires client-specific control variates, creating two key challenges: it violates privacy-preserving principles and demands stateful clients. In this paper, we propose Compressed Aggregate Feedback (CAFe), a novel distributed learning framework that allows highly compressible client updates by exploiting past aggregated updates, and does not require control variates. We consider Distributed Gradient Descent (DGD) as a representative algorithm and analytically prove CAFe's superiority to Distributed Compressed Gradient Descent (DCGD) with biased compression in the non-convex regime with bounded gradient dissimilarity. Experimental results confirm that CAFe outperforms existing distributed learning compression schemes.

cs.LG↗
arXiv2025

Transsion Multilingual Speech Recognition System for MLC-SLM 2025 Challenge

This paper presents the architecture and performance of a novel Multilingual Automatic Speech Recognition (ASR) system developed by the Transsion Speech Team for Track 1 of the MLC-SLM 2025 Challenge. The proposed system comprises three key components: 1) a frozen Whisper-large-v3 based speech encoder, leveraging large-scale pretraining to ensure robust acoustic feature extraction; 2) a trainable adaptor module using Linear-ReLU-Linear transformation mechanisms to effectively align speech and text representations; and 3) a frozen Qwen2.5-7B-Instruct large language model (LLM) integrated with trainable LoRA for optimized contextual linguistic decoding. By systematically combining pretrained models with task specific fine-tuning, the system achieved a word/character error rate (WER/CER) of 9.83% across 11 languages in the evaluation set and ranked third place among global participants.

eess.AS↗
arXiv2025

PTCMIL: Multiple Instance Learning via Prompt Token Clustering for Whole Slide Image Analysis

Multiple Instance Learning (MIL) has advanced WSI analysis but struggles with the complexity and heterogeneity of WSIs. Existing MIL methods face challenges in aggregating diverse patch information into robust WSI representations. While ViTs and clustering-based approaches show promise, they are computationally intensive and fail to capture task-specific and slide-specific variability. To address these limitations, we propose PTCMIL, a novel Prompt Token Clustering-based ViT for MIL aggregation. By introducing learnable prompt tokens into the ViT backbone, PTCMIL unifies clustering and prediction tasks in an end-to-end manner. It dynamically aligns clustering with downstream tasks, using projection-based clustering tailored to each WSI, reducing complexity while preserving patch heterogeneity. Through token merging and prototype-based pooling, PTCMIL efficiently captures task-relevant patterns. Extensive experiments on eight datasets demonstrate its superior performance in classification and survival analysis tasks, outperforming state-of-the-art methods. Systematic ablation studies confirm its robustness and strong interpretability. The code is released at https://github.com/ubc-tea/PTCMIL.

cs.CV↗
arXiv2025

Triangle decompositions of PG(n-1,2)

We define a triangle design as a partition of the set of lines of a projective space into triangles, where a triangle consists of three pairwise intersecting lines with no common point. A triangle design is balanced if all points are involved in the same number of triangles. We construct balanced triangle designs in PG$(n-1,2)$ for all admissible $n$ (congruent to $1$ modulo $6$) and an infinite class of balanced block-divisible triangle designs. We also prove that the existence of a triangle design in PG$(n-1,2)$ invariant under the action of the Singer cycle group is equivalent to the existence of a partition of $Z_{2^n-1}\backslash\{0\}$ into special $18$-subsets and find such partitions for $n=7$, $13$, $19$. Keywords: Subspace design, graph decomposition, triangle design, Heffter's difference problem.

math.CO↗
arXiv2025

Enhancing Clinical Multiple-Choice Questions Benchmarks with Knowledge Graph Guided Distractor Generation

Clinical tasks such as diagnosis and treatment require strong decision-making abilities, highlighting the importance of rigorous evaluation benchmarks to assess the reliability of large language models (LLMs). In this work, we introduce a knowledge-guided data augmentation framework that enhances the difficulty of clinical multiple-choice question (MCQ) datasets by generating distractors (i.e., incorrect choices that are similar to the correct one and may confuse existing LLMs). Using our KG-based pipeline, the generated choices are both clinically plausible and deliberately misleading. Our approach involves multi-step, semantically informed walks on a medical knowledge graph to identify distractor paths-associations that are medically relevant but factually incorrect-which then guide the LLM in crafting more deceptive distractors. We apply the designed knowledge graph guided distractor generation (KGGDG) pipline, to six widely used medical QA benchmarks and show that it consistently reduces the accuracy of state-of-the-art LLMs. These findings establish KGGDG as a powerful tool for enabling more robust and diagnostic evaluations of medical LLMs.

cs.CL↗
arXiv2025

Cross-attention and Self-attention for Audio-visual Speaker Diarization in MISP-Meeting Challenge

This paper presents the system developed for Task 1 of the Multi-modal Information-based Speech Processing (MISP) 2025 Challenge. We introduce CASA-Net, an embedding fusion method designed for end-to-end audio-visual speaker diarization (AVSD) systems. CASA-Net incorporates a cross-attention (CA) module to effectively capture cross-modal interactions in audio-visual signals and employs a self-attention (SA) module to learn contextual relationships among audio-visual frames. To further enhance performance, we adopt a training strategy that integrates pseudo-label refinement and retraining, improving the accuracy of timestamp predictions. Additionally, median filtering and overlap averaging are applied as post-processing techniques to eliminate outliers and smooth prediction labels. Our system achieved a diarization error rate (DER) of 8.18% on the evaluation set, representing a relative improvement of 47.3% over the baseline DER of 15.52%.

cs.SD↗
arXiv2025

Why is plausibility surprisingly problematic as an XAI criterion?

Explainable artificial intelligence (XAI) is motivated by the problem of making AI predictions understandable, transparent, and responsible, as AI becomes increasingly impactful in society and high-stakes domains. The evaluation and optimization criteria of XAI are gatekeepers for XAI algorithms to achieve their expected goals and should withstand rigorous inspection. To improve the scientific rigor of XAI, we conduct a critical examination of a common XAI criterion: plausibility. Plausibility assesses how convincing the AI explanation is to humans, and is usually quantified by metrics of feature localization or feature correlation. Our examination shows that plausibility is invalid to measure explainability, and human explanations are not the ground truth for XAI, because doing so ignores the necessary assumptions underpinning an explanation. Our examination further reveals the consequences of using plausibility as an XAI criterion, including increasing misleading explanations that manipulate users, deteriorating users' trust in the AI system, undermining human autonomy, being unable to achieve complementary human-AI task performance, and abandoning other possible approaches of enhancing understandability. Due to the invalidity of measurements and the unethical issues, this position paper argues that the community should stop using plausibility as a criterion for the evaluation and optimization of XAI algorithms. We also delineate new research approaches to improve XAI in trustworthiness, understandability, and utility to users, including complementary human-AI task performance.

cs.AI↗