Skip to main content
Search

University of Virginia

12 related articles

Subscribe to updates
  1. arXiv

    Where semantic priors should enter causal discovery: orientation gains hold, but query placement and admission certificates do not yet

    Observational data often leave causal directions unresolved or unstable, and large language models can propose directions from variable names and descriptions. This study splits the use of semantic information into query allocation, proposal admission and graph retention, and tests each on the Asia, Sachs, Child, Insurance and Alarm benchmark networks. Semantic orientation beats coin-flip orientation, but targeting unstable edges shows no established advantage over matched-random allocation; the tested admission certificate can only abstain given the available calibration tasks, and learned references contain the true DAG in only three of 80 planned tasks.

  2. arXiv

    SAM 3 at 4-bit weights stops grabbing the wrong object: SPQ lifts RefCOCO hit rate from 42.1 to 72.5

    Quantizing SAM 3's image-encoder weights to 4 bits while keeping activations at 8 bits produces prompt-alignment drift: masks look plausible but land on the wrong object or lose the target across video frames. SPQ gives more calibration weight to parameters that affect text-concept scores, constrains feature error along low-variance residual directions, and folds a mean-error correction estimated from paired video rollouts into existing biases. Under W4A8, RefCOCO Pr@0.5 rises from 42.1 for PTQ4SAM to 72.5 while retaining 54.8 COCO mask AP, with no added inference operators.

  3. Reward the search, not just the answer: process supervision lifts retrieval recall and holds up out of domain

    Search agents for multi-hop information seeking are trained on step-level judgments of each search action instead of only on final-answer correctness. RAG-Gym casts search as a Markov decision process and compares architecture, parameter optimization and action evaluation, combining a reasoning-reflection architecture, preference optimization and critic-guided inference into Re2Search++. Trained on HotpotQA, it beats outcome-supervised baselines on the unseen Bamboogle benchmark with higher retrieval recall at comparable or fewer search steps, and a critic trained on Llama-3.1-8B raises GPT-4o-mini MedQA accuracy.

  4. arXiv

    MEA uses reward-driven multi-agent optimization to raise explanation faithfulness by up to 34%

    The work presents MEA, a multi-agent framework in which a Proposer agent selects and configures explanation tools based on the question and modality while an Actor agent is optimized end-to-end against faithfulness to turn outputs into natural-language explanations; the authors also introduce question types spanning feature attribution, counterfactual reasoning, and spurious feature detection, each paired with a perturbation-based faithfulness metric, find that frontier LLMs systematically produce unfaithful explanations, and report that MEA outperforms post hoc explainers, agentic, and closed-source baselines across six datasets, with faithfulness gains of +28% (tabular), +21% (text), and +34% (vision) over the untrained backbone.

  5. arXiv

    Voice agents skip token-by-token decoding: a frozen LLM reads one hidden state, answering 80 decisions on eight utterances in about 0.1 s

    Full-duplex voice agents make many small closed decisions, and current systems answer them by slow autoregressive decoding. DuplexJev feeds ASR-encoder hidden states through a small connector into a frozen LLM and reads each question as a single-token distribution over its options, decoding nothing; an 8-GPU node answers 80 decisions about eight utterances in about 0.1 s. With a last-layer connector, spoken QA stays close to reading the transcript (90% vs. 91%); a cross-attention connector brings gender and emotion accuracy to 90% (from 55% and 28%) while spoken QA drops only one point (83% to 82%).

  6. arXiv

    DUET co-evolves solver and grader agents in alternating rounds, improving both task execution and grading agreement across four agent benchmarks

    DUET treats both a solver agent and a grader agent as optimization targets: each round it adaptively selects training tasks, the solver executes them, the grader scores the outcomes with evidence-grounded critiques, and a tool-using update module revises the optimizable component (system prompt or skill library) of either agent, alternating between solver and grader, with training updates using no reference answers or ground-truth labels; across SpreadsheetBench, GDPval, τ-Bench, and Finance Agent Benchmark, DUET improves both solver and grader performance and surpasses fixed-grader baselines SkillOpt and GEPA on solver metrics, even when those baselines use the benchmark reference evaluator.

  7. arXiv

    SRFT lets LLM agents learn from their own failures, cutting AgentDojo attack success from 7.59% to 1.26%

    The work proposes Self-Reflection Fine-Tuning (SRFT): it injects attack instructions into clean expert trajectories, samples the agent's own potentially failing actions under the attacked context, and has an expert model generate structured self-reflection reasoning that contrasts unsafe and optimal actions, which then supervises fine-tuning; instantiated as SR-Agent on Llama-3.1-8B-Instruct and Qwen3-8B, it reduces AgentDojo attack success rate from 7.59% to 1.26% and from 16.97% to 1.05% respectively, and keeps a low attack success rate under the RL-Hammer adaptive attack.

  8. arXiv

    MemLife builds entity-grounded first-person text memories read by a time-indexed agentic reader, beating the strongest training-free baseline by 4.6–12.0% on four long-horizon egocentric video benchmarks, with MemOpt adding 2.7–5.0% by training only the memory writer under the FIRM reward

    The work introduces MemLife, a multimodal memory system that compacts egocentric video into time- and entity-anchored first-person text episodes retrieved by a time-indexed agentic reader, improving over the strongest training-free baseline by 4.6–12.0% across four long-horizon benchmarks without training or query-time video access, and further proposes MemOpt, which applies reinforcement learning only to the memory writer under the faithful, informative, and retrievable FIRM reward, adding 2.7–5.0% with gains that transfer across writer and reader backbones, memory systems, and out-of-domain video and question distributions.

  9. A 45-member modified Delphi consensus: screen for eating disorders before GLP-1 receptor agonist prescribing, and generally avoid these agents in active anorexia and bulimia nervosa

    Using a three-round modified Delphi method, a 45-member panel of clinicians and researchers in diabetes, obesity and/or eating disorders plus people with lived experience reached an average 91.7% consensus on recommendations for GLP-1 receptor agonist use in eating disorder contexts, covering pre-treatment screening, risk-stratified monitoring, standardized educational materials and concurrent evidence-based psychotherapy, stating that these agents should generally be avoided in active eating disorders, particularly anorexia nervosa, atypical anorexia nervosa and bulimia nervosa with marked dietary restraint and/or significant weight suppression, and setting out research priorities in epidemiology, regulation and binge eating disorder trials.

  10. arXiv

    CalcSeg combines confidence-aware curriculum learning with slice-wise self-attention to raise myocardial scar segmentation Dice to 0.677 and low-confidence Dice to 0.644 on single-stack LGE-CMR

    The work presents CalcSeg, a confidence-aware latent 3D context curriculum learning framework that scores each sample using Dice, percentage scar-burden error, and epistemic uncertainty from Monte Carlo Dropout, expands training from easy to hard cases across stages, and uses slice-wise self-attention to infer subject-level 3D anatomical context from single-stack 2D LGE-CMR; on LGE-CMR data from four sites and two segmentation challenges (MICCAI 2012 LV Infarct, EMIDEC 2020) comprising 976 patients, it reaches myocardial scar Dice of 0.677±0.24, low-confidence Dice of 0.644±0.22, scar error of 38.88%, and low-confidence scar error of 35.03%, outperforming TransUNet, AttentionUNet, UNETR, ScarNet, and ScarNet with supervised curriculum learning using expert difficulty labels.