Searcharxiv⌕ Search

arXiv subjects

Hong Yu

Publications and source records attributed to Hong Yu.

At least 19 recordsLinked to original sources
arXiv2026

EpiLoRA: Fusing Many LoRA Adapters into Low-Precision GEMMs for Serving and RL

Low-rank adapters (LoRA) let one base model serve many fine-tuned variants: multi-tenant engines keep a single copy of the base and batch the requests of many tenants, each with its own adapter. As base models move to 8-bit (FP8) and 4-bit formats, their general matrix multiplications (GEMMs) become cheaper, but the adapters do not: engines run them as separate small GPU kernels beside the GEMM, whose latency takes a growing share of every decoding step. An adapter projects its input down to a low rank and back up, and our key observation is that the GEMM needs the down-projection only in its epilogue, the final stage in which it writes its output. EpiLoRA, a plugin for the vLLM engine, therefore launches the down-projection first and the low-precision GEMM as a dependent kernel that streams its weights meanwhile and adds every tenant's up-projection in its epilogue. On NVIDIA Hopper GPUs, EpiLoRA decodes for a single adapter with up to 24% lower time per token than vLLM's fastest configuration, and faster than SGLang and TensorRT-LLM; with many tenants, its adapter kernels deliver more throughput than vLLM's on most GPU-model pairs we tested. The same low-precision base also trains: a LoRA reinforcement-learning trainer that computes on the engine's FP8 weights needs no 16-bit copy, so a 27B model trains on one 48 GB GPU, and an engine that then keeps activations in 16 bits removes most of the remaining mismatch between the model generating training data and the one trained.

cs.DC↗
arXiv2026

RIVO: Lightweight Adaptive Rerouting for Deep Research Agents

Deep Research agents rely on long-horizon reasoning and iterative tool use to solve complex information-seeking tasks. However, adaptive decision-making over these trajectories remains challenging, as agents may stop prematurely when further evidence is needed, continue unnecessarily after sufficient evidence has been gathered, or fail to redirect unproductive trajectories. Existing solutions either require costly model retraining and dense turn-level supervision, or rely on heuristic control rules that are not directly optimized for downstream research utility. This motivates a lightweight, utility-aware alternative for Deep Research control. We formulate adaptive rerouting in long-horizon Deep Research as a state-conditioned trajectory correction problem. Based on this view, we introduce RIVO, a lightweight plug-in framework for adaptive trajectory correction. The agent estimates its research state from lightweight signals and determines whether trajectory redirection is needed. Once triggered, the plug-in modifies the Host token logits to steer generation toward a more promising research direction. Theoretical analysis shows that, under a value-alignment condition and bounded local curvature, a sufficiently small logit intervention can locally improve expected downstream research utility. We further construct BrowseCore, a curated dataset with a paired local corpus, for efficient and controlled Deep Research experimentation. Building on this resource, the plug-in is trained from downstream outcomes, enabling it to learn corrections that improve subsequent research and final task performance. Extensive experiments across multiple Deep Research benchmarks and control settings demonstrate the effectiveness of our method, showing that lightweight logit intervention can improve long-horizon research without retraining the underlying Host model.

cs.CL↗
arXiv2026

SAGE: Sense-Aligned Multimodal Grounding Evaluation with Gloss-Augmented Reasoning

This paper introduces Sense-Aligned Grounding Evaluation (SAGE), a unified 10-candidate image-selection benchmark that reformulates VWSD, VerSe, and WikiDiverse into a shared term-centric protocol covering noun-, verb-, and entity-centered grounding. SAGE includes confusable hard negatives and a context-swapped VerSe diagnostic to probe target-term grounding beyond direct caption--image compatibility. We further propose Gloss-Augmented Reasoning (GAR), a training-free framework combining definition grounding with constraint-guided semantic verification. Experiments show that strong VWSD performance does not transfer uniformly across grounding settings. With Gemma3-12B, definition augmentation improves all three subsets, while constraint-guided reasoning improves VWSD and yields particularly strong gains on VerSe, but is less effective on WikiDiverse. Combining the two signals achieves the best macro-average performance, improving Hits@1 from 65.90 to 71.22. These results show that SAGE supports both unified evaluation and fine-grained analysis of explicit semantic reasoning across grounding settings.

cs.CL↗
arXiv2026

Second Thoughts: Exploring and Planning through Comparative Imagination in World Models

World models let agents compare imagined outcomes before acting, but an agent must still decide which alternatives to explore and when a proposed plan should replace its learned policy. We introduce SDI (Sibling Disagreement in Imagination), which connects these decisions through a shared actor-based reference. SDI compares action sequences imagined from the same model state, using repeated rollouts of the actor's own sequence to measure variation under fixed behavior. This reference scales both the spread across candidate returns, which provides an exploration signal, and a proposed plan's predicted advantage, which governs action acceptance. The agent executes a planned action when its advantage exceeds the reference variability by a fixed factor; otherwise, it follows the actor. SDI reuses the base agent's world model, actor, and critic without requiring an additional disagreement ensemble. Exploration and planning guide training data collection, while evaluation uses the learned actor alone. Across Meta-World, Adroit, and DM-Control, SDI achieves a higher observed task-average score than every compared baseline in each suite and matches or exceeds the strongest baseline's performance on 15 of 18 tasks. Component ablations and robustness analyses further examine the roles of exploration, planning exposure, and action acceptance. Paired evaluations from identical simulator states show that the gate improves both recovery from actor failures and preservation of successful behavior relative to ungated planning, while substantially reducing return degradation. Together, these results support comparing imagined alternatives against a shared actor reference as a unified approach to exploring and planning with world models.

cs.LG↗
arXiv2026

Measure First, Then Allocate: Calibrated Rollout Budgets for MCP Agents in Reinforcement Learning

Training tool-using language models requires costly interaction, yet identical rewards within a rollout group provide no group-relative learning signal. We use reward disagreement to guide harness evaluation and the allocation of rollouts for policy training in Model Context Protocol (MCP) environments. We first introduce Discord, which tests harness edits using exact paired inference and decision-preserving early stopping: only disagreeing paired outcomes favor one harness over another. This insight motivates prospective allocation in BRACE (Bank-calibrated Rollout Allocation under Correlated Episodes). BRACE combines historical task evidence, uncertainty in task success rates, and a correlated reward model to forecast mixed-reward groups before allocating generation budget. Controlled harness experiments favor Discord, while live MCP experiments expose a retention-cost trade-off. At three Qwen3-VL model scales, BRACE exceeds the strongest plotted baseline peaks with fewer new training rollouts; historical-bank collection is counted separately. It also leads the compared methods on BFCL and NESTFUL without further training. Ablations probe the bank, estimator, and allocation score; prospective evaluation checks forecast calibration. These results support forecasting reward contrast before generation.

cs.LG↗
arXiv2026

Is Anyone Enforcing the Cartel? What Reinforcement Learning Teaches LLM Agents in Markets

Evaluations of large language model (LLM) agents report that pricing agents collude at others' expense. Yet collusion is defined by enforcement: a cartel, in which competitors share monopoly profits, holds only if a member who cheats is punished. We introduce an audit that tests enforcement by forcing a member to cheat in counterfactual copies of a market, and use it to study what reinforcement learning (RL) teaches agents in pricing and trading. Small LLMs trained by RL learn what their training rivals reward, but only what they readily try. In self-play, they learn fixed rules ignoring their rivals; against a punishing rival, to be deterred; against a cheating rival, to match its cuts, never the costly price war that would deter it, which they rarely try. A war forced into training is learned, but mostly only where it was forced. Against their own copies, deterred and matching agents can earn the same cartel profits, but only matching ones can enforce them, and then thinly: as a by-product and only against large cuts. Q-learners enforce collusion only in pricing. Prompted LLMs reach cartel-like outcomes mostly through rules that ignore the market. Outcome measures cannot tell these cases apart; the audit can.

cs.CL↗
arXiv2026

ArcFlow: Training Flow Policies for Multiple Solver Budgets

Flow-matching policies adjust inference cost through the number of solver steps, but reinforcement learning at a single fixed solver budget can hide degradation at cheaper deployment budgets and even erase useful one-step behavior present in a pretrained policy. We introduce Action-Ratio-Clip Flow (ArcFlow), a deployment-aware algorithm for training one flow policy across a specified range of solver budgets. ArcFlow combines action-aligned supervision, budget exposure, and worst-budget checkpoint selection. For reconstruction objectives, its bridge aggregates temporal probes into one surrogate ratio and one PPO-style clip per action, with a boundary residual exactly equal to deployed one-step reconstruction error. Cycling exposes the shared field to the deployment range, and selection optimizes its performance floor. For native velocity objectives, multi-budget endpoint supervision applies the same principle without replacing the native update. Both realizations preserve the architecture and deployment-time cost. Across DM Control, manipulation, and locomotion, ArcFlow attains the highest mean one-step return among five flow-policy baselines on 25 of 30 tasks under each method's reported training settings. Component studies identify complementary improvements from boundary supervision, the fixed-grid bridge, budget exposure, and deployment selection. These results establish deployment-budget-aware training as an effective route to strong low-step control with one policy serving multiple inference costs.

cs.LG↗
arXiv2026

Does Multi-Agent Collaboration Reduce Hallucination in Clinical Tasks?

Multi-agent LLM systems improve performance on clinical tasks, and that improvement is often read as evidence of a more trustworthy system. However, the two are not the same claim. A clinician acts on the stated reasoning and not only on the conclusion, so whether collaboration reduces hallucination, and what decides when it does, has to be measured in the reasoning too. Yet most evaluations score only the final answer. We evaluate four systems (a single-LLM baseline and three multi-agent topologies) across four models and three clinical tasks, scoring hallucination, under a factuality and faithfulness taxonomy, inside the rationale each system generates and reporting it alongside the task metric. We report that collaboration buys task performance but not lower hallucination as a rule, and no topology is reliably better than another. The effect differs by model: of the $36$ model--system--task comparisons, $12$ are large enough to resolve --- four decreases and eight increases, with no model on both sides. What sets the level, and the direction of any change, is the model rather than the topology or the number of agents. A self-correction experiment is consistent with this: shown its own flagged errors, every model except Llama-3.2-3B-Instruct reduces hallucination; the one model for which collaboration lowers hallucination gains the most, and Llama-3.2-3B-Instruct, the model whose hallucination it amplifies most, does not reduce its hallucination at all. We find that collaboration redistributes what its models can already do; it does not supply what they cannot.

cs.AI↗
arXiv2026

DischargeBench: Pre-Deployment Evaluation of LLM Discharge Educators through Persona-Grounded Patient Simulation

Effective discharge education is personalized: a clinician adapts the discharge plan to each patient's literacy, recall, and personality. Before large language models (LLMs) take on this role with real patients, they need to be evaluated in conversations with diverse patients, yet existing evaluations target static, artifact-generation, or question-answering tasks. We introduce **DischargeBench**, a pre-deployment benchmark in which a candidate LLM educator teaches a persona-grounded simulated patient in an open-ended, multi-turn conversation. We curate **MIMIC-IV-Ext-DischargeBench**, 320 cases across 16 ICD chapters, each paired with a patient persona that varies personality, education level, health literacy, and recall. Each conversation is scored on Conversation Quality, Topic Checklist Score (TCS), Simulated Comprehension Score (SCS), and a Claim-Level Safety Audit (CLSA) by an LLM judge validated against physician annotations. Across eight closed- and open-source LLMs, the best model scored only 0.70 on TCS and 0.72 on SCS, and distrustful and anxious patients lowered comprehension for every model. DischargeBench is intended as a screening step before human evaluation, not as evidence of readiness to educate patients.

cs.CL↗
arXiv2026

OverdoseMoE: A Multi-Expert Framework for Opioid Overdose Risk Prediction

Opioid overdose remains a major clinical and public health burden, highlighting the need for scalable approaches to identify patients at high risk. Here, we investigate diagnosis-specific adaptation for 180-day opioid overdose risk prediction from patients' preceding one-year longitudinal ICD histories. We develop OODMAMBA and OODQWEN through continued pretraining on longitudinal diagnostic sequences followed by task-specific fine-tuning. Building on the stronger Qwen-based predictors, we further propose OVERDOSEMOE, a multi-expert framework that integrates models of different scales using complementary expert-weighting strategies. Diagnosis-specific adaptation consistently improved predictive performance over general-purpose language-model baselines, with OODQWEN achieving an AUPRC of 24.47 and an AUROC of 68.56. OVERDOSEMOE further improved discrimination and precision, achieving an AUPRC of 25.17 and an AUROC of 69.49 while outperforming the strongest single-model baselines. Among patients ranked in the top 5% of predicted risk, OVERDOSEMOE identified substantially enriched overdose risk, achieving a PPV of 25.38% while retaining meaningful recall. Evaluation on an independent MIMIC-IV cohort further demonstrated cross-cohort robustness, with complementary weighting strategies showing advantages across different performance measures. These findings demonstrate that diagnosis-specific language-model adaptation combined with multi-expert integration can improve opioid overdose risk stratification and support more robust prediction across heterogeneous electronic health record populations.

cs.CL↗
arXiv2026

The Visual Target Matters: Learning across the Visual Hierarchy for Brain-to-Image Retrieval

Brain-to-image retrieval seeks to identify the visual stimulus that elicited a non-invasive neural response. Candidate images are typically represented by pretrained vision models, whose internal representations vary in abstraction across depth. Existing methods usually train the neural encoder to recover a fixed final-layer visual target. Under this formulation, the visual hierarchy is reduced to a single prescribed endpoint, preventing representations at other depths from directly shaping the visual target. This limitation motivates learning how information across visual depths should contribute to the retrieval target. To this end, we introduce NeuroGlyph, which learns a trial-independent visual target from multiple depths of a frozen visual backbone. NeuroGlyph decomposes the target into factor-specific subspaces. Each subspace learns an image-conditioned allocation over visual depth. The resulting subspaces are fused into a single embedding for retrieval. Across THINGS-EEG and THINGS-MEG, NeuroGlyph outperforms final-layer supervision in all controlled comparisons. It also surpasses the post hoc best fixed-layer oracle in three of four comparisons. Parameter-matched ablations support both factorized target construction and image-conditioned depth allocation. Under comparable 200-way retrieval protocols, NeuroGlyph achieves the strongest system-level performance in six of eight reported metrics. These results support learning retrieval targets across the visual hierarchy rather than prescribing one visual depth.

cs.CV↗
arXiv2026

MultiViewDx: Evidence-Linked Multi-View Clinical Diagnosis

Medical multimodal large language models (MLLMs) can perform well on existing medical visual question answering (MedVQA) benchmarks, but their training data often does not match clinical diagnosis. Most supervision is organized around isolated images or short QA pairs, leaving two structures weakly specified: how evidence leads to a decision, and how views, series, modalities, and patient context from the same case are linked. We introduce MultiViewDx, a partly physician-validated multimodal instruction dataset for evidence-linked multi-view medical imaging diagnosis. MultiViewDx uses the clinical case as the supervision unit. It links imaging studies with patient context, normalizes heterogeneous reports into an evidence-linked workflow (evidence -> findings -> differential discussion -> diagnosis), and uses a unified image-text retriever to constrain instruction synthesis to source-supported evidence. It covers X-ray, CT, MRI, ultrasound, histopathology, and other clinical visual sources. We fine-tune MultiViewDx-8B-AN and evaluate it on both existing MedVQA benchmarks and real-world case-based diagnostic reasoning. Across four MedVQA benchmarks, it achieves the best average accuracy among compared systems (79.0%), outperforming HuatuoGPT-Vision-34B (66.7%) and Claude3-Opus (55.7%). Beyond MedVQA, on JAMA Clinical Challenge cases, it receives the strongest overall rating under a physician-designed rubric for key clinical points, diagnostic inference, and evidence grounding. Controlled ablations and clinician evaluation show that both case-level multi-view organization and evidence-linked reasoning targets contribute to the gain.

cs.CL↗
arXiv2026

When Evidence Conflicts: Reliability-aware Meta-review Generation

Generating coherent meta-reviews from multiple peer reviews is challenging when reviewer evidence conflicts and varies in reliability. Existing approaches typically formulate meta-review generation as a multi-document summarization task and aggregate reviewer feedback uniformly, making it difficult to determine which opinions should be prioritized under disagreement. In this paper, we study meta-review generation through reliability-aware evidence aggregation. Our framework first extracts aspect-level opinions from peer reviews and identifies conflicting evidence within each aspect. It then estimates opinion-level support and review-level quality to measure evidence reliability. Based on these signals, the framework assigns reliability-aware weights to reviewer feedback, enabling the generator to prioritize better-supported arguments while preserving diverse perspectives. Experiments demonstrate that our method consistently improves meta-review generation over strong baselines on both automatic and human evaluations, with clear gains in conflict recognition and resolution under high-conflict review scenarios. The code and implementation details are publicly available at https://github.com/Wangxz729/reliability-aware-meta-review.

cs.CL↗
arXiv2026

Adaptive Cortically Constrained EEG-Vision Alignment for Zero-Shot Brain-to-Image Retrieval

Zero-shot brain-to-image retrieval requires robust alignment between noisy EEG responses and visual representations. Existing EEG-vision alignment methods often operate in sensor space and apply fixed visual supervision to all responses, ignoring both spatial mixing in scalp EEG and response-wise variability in alignment reliability. We propose an adaptive cortically constrained EEG-vision alignment method for zero-shot brain-to-image retrieval. The method reconstructs EEG responses into predefined ROI-level source-pattern representations and encodes them with a Neuro-ROI Attention Encoder. To handle response-wise variability, we introduce an evidence-based adaptive visual supervision strategy that weights detail-controlled visual targets using model-based alignment evidence. On THINGS-EEG, the proposed method achieves strong 200-way zero-shot retrieval performance, with ROI-level attribution providing post hoc interpretability of the learned source-pattern representations. These results show that cortically constrained representation learning and adaptive supervision can jointly support EEG-vision alignment for zero-shot brain-to-image retrieval.

cs.CV↗
arXiv2026

MedQA-MM: Shortcuts Behind Medical Visual Reasoning

A benchmark score credits final answers, but not the route by which an item can be answered. In medical multimodal multiple-choice questions (MCQs), this distinction matters because a correct answer can be supported by the intended image finding or by benchmark-preserved cues in the wording of answers, non-visual clinical text, visible image text, artificial annotations, or device/context artifacts. We call the resulting score-level overinterpretation reasoning inflation. Here, a route is an observable input path that can support answer selection, not a claim about the model's hidden cognition. Across six medical multimodal MCQ datasets, we separate candidate cues from behavioral evidence through prompt- and image-side audits, modality ablations, and matched repairs that preserve the medical target and answer key. In a 13-configuration open-model panel, full-input accuracy is 62.63%, while text-only and options-only settings achieve 53.96% and 29.71%, respectively. Removing length-gap, absolute/conspicuous, and spatial/prepositional cues lowers accuracy by 6.58, 3.50, and 4.77 percentage points. We also construct MedQA-MM, a 1,000-item shortcut-mitigated subset, where text-only and options-only accuracy fall to 5.21% and 12.33%. This does not imply that models never use images; it shows that medical image-reasoning claims require route-level evidence.

cs.CV↗
arXiv2026

MUSES: A Benchmark for Prospective Intellectual-Roots Retrieval

Scientific discovery depends on finding prior literature that shapes what comes next. Existing retrieval systems optimize for relevance and popularity, often favoring central papers over less familiar works that later prove generative. We introduce \textbf{MUSES}, a million-instance benchmark for prospective intellectual-roots retrieval over a fixed 2.33M-paper corpus, with roughly 140K test instances per familiarity tier. To our knowledge, it is the first prospective benchmark at this scale with a shared retrieval task and author-confirmed paper-level root labels. Alongside it, \textbf{CiteRoots} pairs a scalable rhetorical layer over local citation text (LLM judge $κ= 0.896$ versus human gold) with a paper-level author-endorsed layer ($n = 1{,}518$ generative-inspiration pairs from 753 focal papers). MUSES organizes difficulty along two axes: a \emph{familiarity} axis spanning CiteNext, CiteNew, and CiteNew-Isolated, and a \emph{functional} axis spanning broad citations, rhetorical roots, and author-endorsed roots. Across 9 method classes, a lean multi-centroid retriever built on SPECTER2 is strongest. Hit@100 falls from 0.534 on CiteNext to 0.424 on CiteNew, 0.205 on rhetorical CiteNew, and 0.171 on author-endorsed CiteNew, a $3.1\times$ decline. In a registered eight-lens full-test audit, roughly half of broad-tier test instances remain unsolved at K=1{,}000. Rhetorical role and author endorsement are distinct: the same judge agrees with endorsement at $κ= 0.037$. We release MUSES, both CiteRoots layers, and a distilled open companion judge for future work on prospective retrieval and intellectual roots.

cs.IR↗
arXiv2026

Foundation Agents Meet Agentic Deep Research: Evidence-Grounded Clinical Code Forecasting

Next-encounter ICD forecasting predicts which standardized diagnosis codes will be documented at a future visit from the longitudinal record available beforehand. The task is prospective and multi-label: the target note does not yet exist, and several codes may be correct. Structured EHR foundation models capture recurrence and temporal progression, whereas language foundation models generate flexible diagnostic hypotheses. We introduce ICD-Deepresearch, a DeepResearch workflow that composes these predictive foundation models with medical search and ICD dictionaries. Because no source reveals the future code set, research evaluates candidate transitions by linking patient evidence, external clinical relations, and exact code semantics under a fixed top-K budget. Candidate Generation uses SparseEHR to produce an EHR Prior that initializes two bounded Research Expansion rounds; an independent GPT-5 Direct Forecast supplies complementary candidates. Final Selection validates, deduplicates, and jointly ranks both paths, after which a separate module writes rationales without changing predictions. Finally ICD-Deepresearch achieves patient-averaged precision/recall of 24.60/35.09% on MIMIC-III and 25.14/48.32% on MIMIC-IV. Physicians rate 51% and 68% of its retrieved documents useful, compared with 22% and 39% for standalone GPT-5 web search and 32% and 41% for Medical Deep Research. ICD-Deepresearch therefore improves over the registered local comparators while retrieving evidence with higher physician-rated usefulness than the standalone research systems

cs.CL↗
arXiv2026

RICE-PO: Turning Retrieval Interactions into Credit Signals for Reasoning Agents

Retrieval is increasingly moving from one-shot matching toward interactive reasoning, where language agents iteratively inspect evidence, reformulate queries, and search again. Training such agents raises a credit-assignment challenge: executable actions such as queries or summaries can be directly evaluated by the retriever, while latent reasoning steps are not directly observable and only affect future executable actions. This asymmetry makes outcome-level reward assignment unreliable, as the same final reward may credit reasoning steps that did not actually shape retrieval success. We propose RICE-PO, a critic-free policy optimization framework that converts retrieval interactions into localized learning signals. RICE-PO selects high-uncertainty executable actions as anchors, evaluates local counterfactual branches using retrieval metrics, and propagates credit to latent reasoning steps only when reasoning-to-action influence is strong and future residual effects are stable. On BRIGHT and BEIR, RICE-PO consistently outperforms prompt-based agents and group-based RL baselines under the same retriever setting. These results show that the structure of agent-environment interaction itself can provide useful supervision for training reasoning-based retrieval agents.

cs.CL↗