Search arXiv⌕ Search

arXiv subjects

Xiaoxiao Li

Publications and source records attributed to Xiaoxiao Li.

At least 19 recordsLinked to original sources
arXiv2026

VisualCoding: Learning Executable Skills for Reliable Visual Reasoning

Learning reusable skills can improve visual reasoning agents, yet applying a skill still requires the agent to implement or invoke a procedure and produce the final answer. Errors and hallucinations at these steps can prevent acquired experience from translating into reliable single-run performance. We propose VisualCoding, which combines execution-guided acquisition of reusable visual programs with direct adoption of their answers on applicable inputs. Execution-Guided Program Acquisition composes visual primitives and additional code into family-level solvers, then iteratively tests and revises them on labeled training instances. Scope-Aware Direct Execution runs a retained solver and returns its valid in-scope answer. We evaluate VisualCoding on four models across four benchmarks spanning 69 task families. VisualCoding exceeds the strongest evaluated baseline in mean accuracy by 4.45 percentage points with Gemini-3.1-Pro and by 14.78 points when reusing the frozen Pro-acquired artifacts with Flash-Lite. Ablations support execution-guided acquisition and show that direct execution outperforms textual or callable reuse of the same programs. Code is available at https://github.com/jiaq-liu/VisualCoding.

cs.CV↗
arXiv2026

VLMProbe: An Open-source Library for VLM Probing and Interpretation

Understanding how vision-language models use visual information requires connecting internal representations to model behavior. We present vlmprobe, an open-source library that integrates lens readouts, attention attribution, causal interventions, and sparse autoencoder (SAE) analysis within a common model interface. Shared adapters, calibration and checkpoint utilities, evaluation pipelines, and visualizations support systematic analysis across architectures. Using this framework, we compare methods across vision-language, language, and vision models through matched case studies and controlled evaluations of pretrained sparse dictionaries. The comparisons show that representation readouts and spatial explanations depend on the analysis objective and observation site, and that local attribution estimates can disagree with measured intervention effects. SAE evaluations further expose distinct trade-offs in reconstruction fidelity, concept readout, and spatial localization. Together, these results support combining multiple forms of evidence to evaluate an interpretation. vlmprobe provides a reusable basis for selecting methods, comparing their conclusions, and testing hypotheses about model behavior. The code is available on github.com/ubc-tea/vlm-probe

cs.CV↗
arXiv2026

EvalRx: Agentic Diagnosis and Verified Repair of Foundation Models

Model evaluation reveals where models fail, but rarely why or whether a proposed fix genuinely works. We introduce EvalRx, an open-source agentic framework for statistically grounded model diagnosis and repair. EvalRx discovers systematic failure modes, explores candidate mechanisms, formulates falsifiable hypotheses, and verifies them by executing frozen, pre-specified tests before using supported diagnoses to guide tiered repair. Candidate repairs are validated against the unchanged model and promoted only when they improve the target behavior without unacceptable regression. Across language, vision-language, and audio-language settings spanning Gemini, Qwen, Gemma, and Nemotron, EvalRx improves performance in all 12 evaluated model family-modality combinations, with gains ranging from +1.4 to +10.9 percentage points (+5.8 points on average). These results show that model evaluation can serve not only to measure failures, but to support an auditable process for diagnosing and repairing them.

cs.AI↗
arXiv2026

Multi-Reward Hacking: From Aggregation-Induced Drift to Trust-Anchored Defense

To improve task performance while encouraging desirable response properties, LLM post-training often combines a primary reward (e.g., correctness) with auxiliary rewards (e.g., brevity or efficiency). However, how these rewards are aggregated can critically affect training outcomes. We show that symmetric aggregation, which treats reward channels equally, can induce a failure where aggregate reward improves while primary-task performance degrades, even when individual rewards are meaningful. We investigate this aggregation-induced failure through a one-step multi-reward decision formulation, characterizing when it constitutes multi-reward hacking. We prove that, under symmetric aggregation, deterministic conditional-mean GRPO converges to auxiliary-favoring behavior when auxiliary rewards are more reliably obtained, regardless of initialization. We propose Trust-Anchored Gating (TAG), which first elects an anchor channel based on the policy drift induced by optimizing each channel alone, then gates auxiliary credit on anchor satisfaction. Unlike calibration-sensitive static reweighting, TAG structurally eliminates credit for auxiliary-only outcomes while preserving additional credit for joint success. In the same analytical setting, TAG provably reverses this selection, converging to the primary-favoring action from any initialization when the primary channel is anchored. Experiments across Qwen and Llama models on mathematics and program synthesis reveal substantial degradation under symmetric, constrained, and normalized aggregation, whereas TAG mitigates this failure. On Qwen3-4B with a brevity reward, TAG achieves 0.581 pass@4 versus 0.220 for additive aggregation, while shortening responses by 44% relative to correctness-only training.

cs.LG↗
arXiv2026

Mitigating Memorization Where It Happens

Large language models (LLMs) can memorize and reproduce training sequences verbatim, raising privacy and copyright concerns. Existing mitigation methods apply interventions uniformly, often degrading performance on the majority of tokens that generalize normally. We study a token-level log-likelihood ratio between a target model and a domain-competent reference with lower exposure to the evaluated sequences. Exposure-controlled experiments support its use for localizing positions associated with sequence-level verbatim recall. Across evaluated settings, a minority of positions carries a disproportionate share of the positive signal, motivating selective activation correction. From the ideal memorization reduction objective, we derive a probe--steer framework which decomposes intervention into a probe that detects memorization-relevant activations and a steer that applies targeted correction only when the probe exceeds a threshold. We propose Gated Subspace Steering (GSS) as its practical instantiation: the optimal probe--steer pair admits a closed-form solution as the leading singular vectors of a gradient-weighted activation matrix. Across four benchmarks, GSS matches or exceeds state-of-the-art memorization reduction while requiring 10--100$\times$ less compute than optimization-based alternatives.

cs.LG↗
arXiv2026

PathLang: A Language-Centered Benchmark for Vision-Language Models in Computational Pathology

Pathology vision-language models (VLMs) have shown strong visual perception ability, but their robustness in the language domain remains poorly characterized. Existing pathology VLM benchmarks largely rely on canonical closed-set prompts or perturb only generic templates, treating language as a fixed evaluation component rather than a variable axis of model behavior. In clinical practice, however, diagnostic language varies across reports, institutions, and candidate diagnoses. We introduce PathLang, a language-centered and clinically grounded zero-shot benchmark. PathLang holds the underlying slides, ground-truth labels, and image-text evaluation direction fixed while systematically varying only the diagnostic language, so that performance differences reflect how a diagnosis is phrased rather than what is imaged. The language variation follows how pathologists actually rephrase diagnoses (terminology, specificity, and reporting style), and all prompts and candidate pools are validated by six board-certified pathologists. PathLang covers four task families: (1) zero-shot classification with image-text alignment analysis, (2) cross-modal retrieval, (3) paraphrase robustness, including semantic-equivalence paraphrases, length and reporting-style variation, and prompt ensembling, and (4) open-vocabulary diagnosis retrieval over four candidate pools with distinct forms of semantic competition. Across nine VLMs and five public datasets spanning four organs, we find that performance is highly sensitive to clinically equivalent paraphrases, varies substantially across forms of semantic competition, and that image-text alignment quality does not necessarily translate into inter-class separability. We release the prompt corpus, candidate pools, pre-computed text embeddings, and evaluation code at https://anonymous.4open.science/r/PathLang.

cs.CV↗
arXiv2026

MemCo: Memory-Centric Collaboration for Generalizing LLM Agents to Unseen Environments

Large language model (LLM) agents increasingly operate in interactive environments, where they need to make sequential decisions through observation, action, and feedback. Although memory can help agents reuse experience, existing work designs memory in isolation, where collecting enough trajectories to populate it is expensive. Existing shared-memory approaches mitigate isolated experience by pooling episodic memories across tasks and environments. However, retrieving shared memory is challenged by the granularity, where retrieved memories can be either too specific to preserve current grounding or too coarse to support the next action. In this work, we propose MemCo, a memory-centric collaboration framework for generalizing LLM agents to unseen interactive environments. It maintains complementary local and global memory spaces, preserving environment-specific details locally while promoting transferable workflows induced from local trajectories to global memory. During online interaction, MemCo routes relevant local and global memories in terms of the agent's current state and decision phase, enabling agents to reuse the experience of other agents without blindly transferring environment-specific details. Experiments on interactive decision-making benchmarks show that MemCo improves task success and reduces redundant exploration compared with isolate-memory and shared-memory baselines. Our code is available at https://github.com/SYannL/nvdamas.

cs.AI↗
arXiv2026

Local Superior Soups: A Catalyst for Model Merging in Cross-Silo Federated Learning

Federated learning (FL) is a learning paradigm that enables collaborative training of models using decentralized data. Recently, the utilization of pre-trained weight initialization in FL has been demonstrated to effectively improve model performance. However, the evolving complexity of current pre-trained models, characterized by a substantial increase in parameters, markedly intensifies the challenges associated with communication rounds required for their adaptation to FL. To address these communication cost issues and increase the performance of pre-trained model adaptation in FL, we propose an innovative model interpolation-based local training technique called ``Local Superior Soups.'' Our method enhances local training across different clients, encouraging the exploration of a connected low-loss basin within a few communication rounds through regularized model interpolation. This approach acts as a catalyst for the seamless adaptation of pre-trained models in in FL. We demonstrated its effectiveness and efficiency across diverse widely-used FL datasets. Our code is available at https://github.com/ubc-tea/Local-Superior-Soups.

cs.LG↗
arXiv2026

RVCBench: Benchmarking the Robustness of Voice Cloning Across Modern Audio Generation Models

Modern Voice Cloning (VC) can synthesize speech that closely matches a target speaker from only seconds of reference audio, enabling applications such as personalized speech interfaces and dubbing. In practical deployments, modern audio generation models inevitably encounter noisy reference audios, imperfect text prompts, multilingual and long-form generation settings, downstream post-processing, and adversarial perturbations, all of which can significantly hurt robustness. Despite rapid progress in VC driven by autoregressive codec-token language models and diffusion-based models, robustness under realistic deployment shifts remains underexplored. This paper introduces RVCBench, a comprehensive dataset and benchmark that evaluates Robustness in Voice Clone. RVCBench contributes a large-scale, task-aligned robustness dataset that instantiates realistic deployment shifts through controlled text-audio pairing, multilingual and long-form scenarios, expressive prompts, post-processing conditions, and passive or proactive audio perturbations. Covering 18 robustness evaluations, 204 unique speakers, and 14,370 utterance-level evaluation items, RVCBench enables unified evaluation of input sensitivity, generation stability, output resilience, and perturbation robustness. We evaluate 18 representative modern open-source VC models and reveal systematic vulnerabilities in content consistency, speaker similarity, long-form stability, post-processing resilience, adversarial robustness, and detector-facing separability. We open-source the toolkit and dataset to support reproducible evaluation and future research.

cs.SD↗
arXiv2026

Learning to Ask: Information Acquisition for SLM-LLM Collaboration, under a budget

Collaboration between a small language model (SLM) and a large language model (LLM) offers an opportunity to combine the efficiency of smaller models with the strong reasoning capabilities of larger ones. Existing approaches primarily frame such collaboration as a computation allocation problem, determining which model should handle each portion of the reasoning process. In black-box API-based settings, however, this paradigm can be inefficient due to coarse-grained delegation or repeated transmission of context across model switches. In this work, we instead formulate SLM-LLM collaboration as an information acquisition problem, under an API budget constraint. The SLM remains the primary reasoner and selectively queries a black-box LLM advisor only when needed, issuing targeted queries rather than delegating the reasoning process itself. To realize this strategy, we develop a three-stage RLVR framework that learns whether to call the advisor, how to formulate useful queries, and how to integrate the collaboration into the reasoning process by jointly refining advisor invocation and information use. Across mathematical reasoning and coding tasks, our approach improves the performance--cost tradeoff over existing collaboration baselines and, in some settings, matches or exceeds oracle problem-level routing. Finally, we show that our strategy can transfer to other advisor model families, without further training.

cs.AI↗
arXiv2026

Self-Evolving Agents via Likelihood-Guided Tool-Space Optimization

Self-evolving agents can continually improve their behavior, while tools define the executable action space through which they interact with the environment. However, exposing the full tool library to model introduces substantial irrelevant context and can impair tool-use decisions. We study tool-space self-evolution, where each recurring task type maintains a persistent tool space which is constructed from accumulated output experience. We identify three limitations of existing methods: (1) output-unaware selection: they rely primarily on tool descriptions or model priors rather than observed tool outputs; (2) statelessness across request: they select tools independently for each request without consolidating prior output experience into persistent task-specific state; (3) inference cost: they repeatedly search, rank, or reason over candidate tools for subsequent requests of the same task. We address these limitations through output-aware tool scoring, persistent task-specific tool spaces, amortized tool selection, and reusable configurations across models. We introduce LOTS (Likelihood-Only Tool Scoring), which evolves an agent's tool space from accumulated output experience while keeping model parameters fixed. After each request, LOTS holds the model's generated answer and estimates each tool's contribution by measuring how much the answer likelihood changes when its observed output is removed. These contributions are aggregated within each recurring task to rank tools and update its persistent space. Across three benchmarks, LOTS improves task performance while substantially reducing tool context. More importantly, sequential experiments demonstrate that task-specific spaces persist and continue to improve over time, while cross-model experiments show that learned configurations transfer across different models.

cs.AI↗
arXiv2026

SlideBank: A Persistent Hierarchical Evidence Bank for Consistent Whole-Slide Reasoning

Whole-slide images (WSIs) are challenging for vision-language reasoning because diagnostically relevant morphology is sparse, heterogeneous, and distributed across gigapixel-scale images and multiple spatial resolutions. Existing WSI models and pathology agents can aggregate slide features or actively acquire evidence, but the information retained after exploration is often difficult to access semantically while preserving its connection to the original visual evidence. We introduce SlideBank, a training-free framework that represents each WSI as a persistent, concept-indexed, and spatially grounded evidence bank. SlideBank performs question-independent coarse-to-fine exploration to identify informative regions and multi-scale views, converts them into explicit morphological observations, and grounds pathology signals to their supporting patches and WSI coordinates. At inference time, questions are routed to relevant signals and evidence scales, and the linked global, regional, and patch evidence is integrated through confidence-based cross-level consensus. Experiments on WSI-VQA and SlideBench-BCNB show that with Patho-R1, SlideBank reaches 52.77% on WSI-VQA and with Quilt-LLaVA, it reaches 50.92% average accuracy on SlideBench-BCNB, while structured signal-guided retrieval consistently outperforms random evidence sampling. Reusing the same bank across repeated queries further achieves over 99% rephrasing consistency and substantially reduces amortized inference cost through persistent evidence reuse.

cs.AI↗
arXiv2026

APIFlow-Bench: Measuring Whether Agents Survive Long, Dependent API Workflows

Tool-using agents are commonly evaluated by a single bit: whether an end-to-end workflow completed. This metric fails to distinguish failures that matter in production, such as expired credentials, malformed payloads, or correct execution followed by incorrect final delivery. We introduce APIFlow-Bench, a fully auditable benchmark for long-horizon, dependent REST-API workflows that decomposes performance into seven engineering capabilities and requires agents to produce answers supported by the actual call path. We generate synthetic API worlds forward, subtask by subtask; each subtask is admitted only after a zero-LLM self-test triad verifies its grader and an oracle establishes solvability, and an adversarial audit identified and fixed six grader exploits. Grading is deterministic and provenance-sensitive: a state check traces a mock-minted canary through the API data flow to the response the answer must originate from, and a typed answer card is verified field by field. We release all answer keys and 44,362 unredacted execution transcripts. Across 19 frontier and open-weight models under one neutral scaffold, we find: (1) longer dependency chains degrade success, from 93% on individual subtasks to 74% on clean 20-subtask chains and 61% when including the 8% of chain trials that a model-consensus screen flags as passed by no model; (2) reliability separates models more than best-case capability, with best-of-five spanning seven points but all-five-of-five reliability spanning 44 points; (3) the independent-error account of compounding failure does not fit the data: pass rates on 20-subtask chains are 33 percentage points above the product of subtask-level rates, and on the clean slice 77% of failing runs reached the correct final state and failed only at delivery.

cs.AI↗
arXiv2026

Toward Federated Large Language Models in Medicine: A Parameter-Efficient Framework for Privacy-Preserving, Multi-Institutional Adaptation

Large language models (LLMs) are increasingly adapted for medical applications, but most are trained using data from a single institution because privacy and governance constraints prevent multi-institutional data sharing. As a result, these models often generalize poorly across heterogeneous healthcare systems. We address this gap by introducing Fed-MedLoRA and Fed-MedLoRA+, a parameter-efficient federated framework for collaborative LLM adaptation across healthcare institutions. Fed-MedLoRA transmits only low-rank adapters rather than full model weights, reducing communication overhead. We also evaluate a privacy-preserving variant that applies Gaussian perturbation to transmitted adapter updates. Fed-MedLoRA+ further incorporates adaptive aggregation to better address cross-site heterogeneity in patient populations, annotation practices, and disease distributions. We evaluate the framework on clinical information extraction across five independent patient cohorts totaling 42,198 entities and 41,570 relations, and compare it with zero-shot and fine-tuned LLMs, domain-specific BERT models, and federated baselines. Across all settings, the proposed methods consistently improve extraction performance and generalize better to heterogeneous cohorts. In a real-world case study using clinical notes from the Yale New Haven Health System, the framework demonstrates strong performance under low-resource new-site deployment. These results suggest that federated, parameter-efficient LLM adaptation is feasible, scalable, and effective for multi-institutional clinical deployment.

cs.CL↗
arXiv2026

Can Generalist Vision Language Models (VLMs) Rival Specialist Medical VLMs? Benchmarking and Strategic Insights

Vision Language Models (VLMs) have shown promise in automating image diagnosis and interpretation in clinical settings. However, developing specialist medical VLMs requires substantial computational resources and carefully curated datasets, and it remains unclear under which conditions generalist and specialist medical VLMs each perform best. This study highlights the complementary strengths of specialist medical and generalist VLMs. Specialists remain valuable in modality-aligned use cases, but we find that efficiently fine-tuned generalist VLMs can achieve comparable or even superior performance in most tasks, particularly when transferring to unseen or rare OOD medical modalities. These results suggest that generalist VLMs, rather than being constrained by their lack of specialist medical pretraining, may offer a scalable and cost-effective pathway for advancing clinical AI development.

eess.IV↗
arXiv2026

DeCo-MIL: Debiased Counterfactual Reasoning for Long-Tailed Whole Slide Image Analysis

Multiple instance learning (MIL) is widely used for weakly supervised whole slide image (WSI) analysis. However, under long-tailed distributions, MIL-based WSI analysis faces a nested dual long-tail: an inter-slide class long tail and an intra-slide long tail of instance-level discriminative evidence. The two long tails are coupled: tail classes have few training slides, while their limited diagnostic evidence is concentrated in a few patches and obscured by abundant within-bag redundancy. This coupling biases models toward head classes and degrades rare-class recognition. To address this, we propose DeCo-MIL for long-tailed WSI analysis, which jointly alleviates the nested dual long-tail through frequency-debiased counterfactual reasoning. For the inner long tail, DeCo-MIL clusters patches into tissue-morphology anchors, replaces each anchor with its matched normal prototype to perform a counterfactual intervention, and estimates its counterfactual contribution to the ground-truth class using class-frequency-corrected predictions. These contributions guide redundancy masking to preserve scarce discriminative instances. For the outer long tail, DeCo-MIL constructs anchor-stratified pseudo-bags from redundancy-reduced bags and combines tail-aware oversampling with consistency regularization, increasing effective supervision for tail classes while preserving tissue-morphology composition. Extensive experiments on three long-tailed WSI benchmarks demonstrate that DeCo-MIL achieves state-of-the-art performance in both tail-class recognition and overall classification.

cs.CV↗
arXiv2026

CoEvoSkills: Self-Evolving Agent Skills via Co-Evolutionary Verification

Anthropic proposes the concept of skills for LLM agents to tackle multi-step professional tasks that simple tool invocations cannot address. A tool is a single, self-contained function, whereas a skill is a structured bundle of interdependent multi-file artifacts. Currently, skill generation is not only label-intensive due to manual authoring, but also may suffer from human--machine cognitive misalignment, which can lead to degraded agent performance, as evidenced by evaluations on SkillsBench. Therefore, we aim to enable agents to autonomously generate skills. However, existing self-evolving methods designed for tools cannot be directly applied to skills due to their increased complexity. To address these issues, we propose CoEvoSkills, a self-evolving skills framework that enables agents to autonomously construct complex, multi-file skill packages. Specifically, CoEvoSkills couples a Skill Generator that iteratively refines skills with a Surrogate Verifier that co-evolves to provide informative and actionable feedback without access to ground-truth test content. On SkillsBench, CoEvoSkills outperforms five baselines on both Claude Code and Codex, and generalizes strongly to six additional LLMs. The code is publicly available at https://github.com/Zhang-Henry/CoEvoSkills.

cs.AI↗
arXiv2026

MESA:Task-Adaptive Multi-Structure Evidence Selection for Long-Horizon Agent Memory

Long-horizon agents accumulate trajectories spanning hundreds of interleaved reasoning, action, and observation steps, where answering a query may depend on evidence buried far back in the history. External memory stores such trajectories as structured representations, yet each structure provides a distinct and incomplete view. Existing multi-memory systems either read a fixed set of structures for every query, inflating context and introducing noise, or route each query to a single structure, preventing the composition of complementary evidence. A controlled analysis on AMA-Bench shows that the optimal memory configuration is typically neither a single structure nor the full union, but a tailored composition of multiple structural memories that varies with query and task demands. Motivated by these findings, we formulate structure-level dynamic selection: selecting and fusing a query-adaptive subset from a library of specialized memory structures. We propose MESA (a Multi-structure Evidence Selection framework for long-horizon Agent), which builds five complementary structure views of each trajectory and learns from end-to-end answer-level feedback to select and fuse a query-specific subset for a frozen answer model. To learn under this weak supervision, MESA employs harness optimization with prior-guided search and UCB-guided scheduling to balance exploration and exploitation. On AMA-Bench, MESA outperforms the strongest baseline by 8.5% while using 41% fewer evidence tokens than the all-structure alternative.

cs.AI↗