Searcharxiv⌕ Search

arXiv subjects

Yitong Qiao

Publications and source records attributed to Yitong Qiao.

12 recordsLinked to original sources

Large Language Models without QK-Normalization

Query-Key Normalization (QK-Norm) is widely adopted in modern large language models to control the growth of attention logits and improve training stability. However, query and key norms are not merely sources of instability. The query norm controls the sharpness of the attention distribution, while the key norm determines how strongly a token attracts different queries. These norms therefore carry token-specific information about attention selectivity and token importance. By normalizing both, QK-Norm suppresses this information. This raises a natural question: can we control query-key representations without discarding their sensitivity to magnitude? To address this question, we propose QK-MAP, a Magnitude-Aware Projection for constructing queries and keys. Rather than removing magnitude through normalization, QK-MAP transforms the input representation using multiple learnable frequencies, so that changes in magnitude are encoded as bounded phase patterns. This allows magnitude to be represented through phase and direction rather than vector length alone. Its periodic features have a fixed norm while remaining sensitive to magnitude changes, separating magnitude encoding from norm growth. Multiple frequencies further help distinguish magnitudes and reduce periodic ambiguity. Experiments across model scales from 190M to 7B parameters show that QK-MAP achieves lower training and validation loss than both Standard attention and QK-Norm. Further analyses show that QK-MAP preserves sensitivity to input scale, assigns less attention to the first token, and produces stronger attention contrast for relevant evidence in synthetic long-context tests. These results demonstrate that query-key scale can be effectively controlled without sacrificing token-specific magnitude information.

cs.CL↗

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↗

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↗

EHR-RobustGym: Benchmarking and Training Agents for Robust Clinical Reasoning

In hospital workflows, electronic health records (EHRs) are often noisy, and may not contain the evidence needed to confirm events or measurements referenced in a clinical query. Even when database retrieval succeeds, clinical agents can overlook such discrepancies and return plausible but unsupported answers. We introduce EHR-RobustGym, a scalable and interactive environment for evaluating and training robust clinical agents grounded in noisy EHRs. Built on MIMIC-IV hospital records (365K patients, 31 tables, and over 500M records), EHR-RobustGym comprises 5,486 Clean-Noise pairs spanning six clinical intents and both patient-level and population-level queries. The pairs test robustness to Record-level, Value-level, and Query-level noise, while interactive SQL/Python execution and outcome verification support trajectory collection and training. Evaluating multiple LLMs reveals substantial robustness gaps: average task success across proprietary and large-scale open-weight models drops from 62.2% on Clean questions to 37.9% on Noise questions. At k=4, pass^k consistency falls below 50% for most evaluated models, exposing instability in clinical task completion. Supervised fine-tuning and reinforcement learning in EHR-RobustGym improve performance, with gains generalizing to five external EHR benchmarks. Together, these results position EHR-RobustGym as a testbed for evaluating and improving the evidence-grounded robustness of clinical agents.

cs.AI↗

From Imitation to Reward Discovery: On-Policy Warmup for Agentic RL

Reinforcement learning with a verifiable reward (RLVR) offers a scalable approach to training language-model agents, yet sparse outcome rewards can leave early training with little signal for policy improvement. We identify an On-Policy Acceleration Phenomenon: in our main comparisons, RLVR initialized with on-policy distillation reaches high performance earlier in training and achieves both higher average performance during subsequent RLVR and higher final performance than the alternative baselines. Motivated by this observation, we study On-Policy Warmup (OPW), a teacher-guided stage in which the student trains with teacher supervision on its own interaction trajectories before transitioning to RLVR. Unlike imitation on fixed teacher-generated trajectories, OPW targets states induced by the student's own decisions, including imperfect actions and recovery situations. We provide a theoretical explanation by connecting on-policy reverse-KL distillation to trajectory-level distribution matching. Under a competent teacher and sufficiently small population distillation loss, this connection yields a lower bound on initial verifier success and a corresponding bound on reward-discovery complexity. For group-relative RLVR, we further characterize when increased success probability produces more reward-informative groups. Together, our findings support on-policy distillation as an effective warmup for agentic RLVR and identify initial reward discovery as a mechanism that can contribute to the observed acceleration.

cs.LG↗

Can LLMs Value the Right Evidence? Evidence-Value Misalignment in Dynamic Medical Diagnosis

A correct diagnosis reached from insufficient or misleading evidence can pose a clinical hazard, yet outcome-based accuracy may reward such lucky guesses. We call this mismatch between diagnostic decisions and the value of available evidence Evidence-Value Misalignment (EVM). To disentangle evidential grounding independently from diagnostic accuracy, we introduce MedEVM, a dynamic benchmarking environment comprising 1,050 cases across 24 disease systems. Observations arrive turn by turn, requiring models to continuously calibrate its decision by deciding whether to wait for more evidence or submit a diagnosis. Across 9 LLMs, four interesting patterns are observed. (1) Miscalibrated evidence tracking. Making a diagnosis often fails to calibrate evidence sufficiency, even in more capable models, and even worsens in reasoning mode. (2) Misaligned diagnosis submission. Confidence in the correct diagnosis often fails to ensure timely submission despite sufficient evidence. (3) Evidence order matters. Reordering the same evidence changes diagnoses even when model confidence remains similar. (4) Misleading evidence remains influential. Added misleading evidence redirects diagnoses even after prior evidence becomes sufficient. We further verify that EVM predicts errors and that preventing premature submission improves accuracy. These findings motivate Evidence-Verified Diagnosis Harness (EVD-Harness). It decouples diagnosis generation from submission through an offline Contrastive Diagnostic Wiki and three online control stages, namely observation management, proposal and witness verification, and diagnosis submission control. Across five LLMs, EVD-Harness improves accuracy by 12.0--51.1 percentage points while mitigating EVM-related failures. Our results demonstrate that verifying evidential support before submission can make diagnostic decisions more reliable.

cs.CL↗

EHR-Complex: Benchmarking Medical Agents for Complex Clinical Reasoning

Clinical agents promise to democratize access to electronic health records (EHRs), yet existing benchmarks fail to reflect the complexity of practical EHR analysis, e.g., often operating on idealized, clean EHRs via static SQL generation rather than interactive execution. In this work, we introduce EHR-Complex, a large-scale benchmark designed for interactive clinical database reasoning. Built on the large MIMIC-IV substrate (365K patients, 31 tables, 500M+ records), EHR-Complex comprises about 52K tasks spanning six clinical intents, supporting both patient-level and population-level queries, where each task requires an agent to interact with a sandboxed environment by executing SQL queries or Python code. Notably, EHR-Complex considers the real-world SQL task complexity for longitudinal multi-table aggregation and compositional reasoning, resulting in 31.93 SQL structural components per query on average. Evaluation results on EHR-Complex reveal the clinical difficulty of these EHR reasoning scenarios, with the top-performing model achieving only 62.3% exact-match accuracy. Pass^k consistency drops below 50% for nearly all evaluated models at k=4, exposing broad stochastic fragility. A fine-grained analysis of more than 3,800 failed trajectories for representative LLMs reveals three dominant failure modes: SQL logic errors, medical-code lookup failures, and semantic misunderstandings. EHR-Complex provides a rigorous testbed for clinical agents and highlights remaining gaps in robust reasoning for large-scale EHR analysis.

cs.AI↗

SpaMEM: Benchmarking Dynamic Spatial Reasoning via Perception-Memory Integration in Embodied Environments

Multimodal large language models (MLLMs) have advanced static visual--spatial reasoning, yet they often fail to preserve long-horizon spatial coherence in embodied settings where beliefs must be continuously revised from egocentric observations under environmental change. We introduce SpaMEM (Spatial Memory from Action Sequences), a large-scale diagnostic benchmark that isolates the mechanics of spatial belief evolution via action-conditioned scene transformations (spawn, place, remove) over long interaction horizons. SpaMEM is built on a physically grounded dataset with 10,601,392 high-fidelity images across four modalities (RGB, depth, instance, semantic segmentation), collected from 25,000+ interaction sequences in 1,000 procedurally generated houses. We formalize embodied spatial reasoning as a three-level hierarchy with 15 diagnostic tasks: Level 1 measures atomic spatial perception from single observations; Level 2 probes temporal reasoning with oracle textual state histories to factor out perceptual noise; and Level 3 requires end-to-end belief maintenance from raw visual streams under the same task dimensions. We further evaluate both short-term (step-wise) updates and long-term (episodic) reconstruction. Benchmarking representative open-source VLM families reveals a consistent stacked bottleneck: coordinate-consistent grounding remains a hard ceiling, and the sharp collapse from Level 2 to Level 3 exposes a pronounced symbolic scaffolding dependency, where models succeed with text-based bookkeeping but struggle to sustain robust visual memory. SpaMEM provides a granular diagnostic standard and motivates explicit mechanisms for state representation, belief revision, and long-horizon episodic integration. A subset of SpaMEM is publicly available at https://huggingface.co/datasets/mill-ct-liao/SpaMEM.

cs.CV↗

Do LLMs Build World Models From Text? A Multilingual Diagnostic of Spatial Reasoning

Whether large language models (LLMs) construct internal spatial world models from pure-text descriptions remains contested, and whether such capabilities transfer across languages has not been systematically studied. We introduce MentalMap, a multilingual diagnostic benchmark with a six-level capability hierarchy (L0-L5) spanning atomic spatial facts to generative world-graph construction, together with four diagnostic axes probing frame of reference, reading-direction bias, reasoning-effort allocation, and hallucination. MentalMap is built from 100 ProcTHOR household scenes, covers eight typologically diverse languages plus a structured-text control, and contains 39 task families across 1,950 evaluation cells. Evaluating thirteen LLMs across scales and model families, we identify a universal L3 reasoning cliff: no model retains even half of its L0 performance on viewpoint reasoning once baseline atomic accuracy exceeds 40%. The cliff persists across languages, scales, and prompting strategies, while structured-output failures and reasoning patterns vary substantially across models. Human evaluation under the identical pure-text protocol reproduces the same failure pattern, suggesting that the bottleneck arises from text-only working memory constraints rather than being specific to current LLM architectures. Our findings reframe pure-text spatial reasoning as a multi-axis world-modeling problem and motivate multimodal and scratchpad-augmented reasoning as future directions.

cs.AI↗

Increasing Computation Resolves Conflicts in Vision Language Models

Cognitive control, the ability to coordinate competing information sources in pursuit of goals, is fundamental to intelligent behavior. We systematically investigate whether Vision Language Models (VLMs) exhibit cognitive control and how computational resources modulate conflict resolution. We construct a benchmark of 4,410 tasks across seven conflict paradigms (Stroop, Flanker, and five realistic variants) spanning multiple difficulty levels and visual complexities, testing 47 VLMs with rigorous experimental control. We find that VLMs exhibit robust congruency effects across all tasks, with larger models systematically resolving conflicts more effectively than smaller models. Critically, VLMs reproduce the fine-grained demand-resource relationship observed in human temporal dynamics: larger models drop below chance on incongruent high-conflict trials while smaller models fail to meaningfully engage and perform at chance, mirroring human behavior at short processing times and establishing parameter count as a proxy for conflict resolution capacity. These findings demonstrate that human-like cognitive control emerges from optimization dynamics in large-scale neural networks, suggesting that adaptive flexibility under conflict may naturally arise through scaling.

cs.NE↗

Blind Spot Navigation: Evolutionary Discovery of Sensitive Semantic Concepts for LVLMs

Adversarial attacks aim to generate malicious inputs that mislead deep models, but beyond causing model failure, they cannot provide certain interpretable information such as ``\textit{What content in inputs make models more likely to fail?}'' However, this information is crucial for researchers to specifically improve model robustness. Recent research suggests that models may be particularly sensitive to certain semantics in visual inputs (such as ``wet,'' ``foggy''), making them prone to errors. Inspired by this, in this paper we conducted the first exploration on large vision-language models (LVLMs) and found that LVLMs indeed are susceptible to hallucinations and various errors when facing specific semantic concepts in images. To efficiently search for these sensitive concepts, we integrated large language models (LLMs) and text-to-image (T2I) models to propose a novel semantic evolution framework. Randomly initialized semantic concepts undergo LLM-based crossover and mutation operations to form image descriptions, which are then converted by T2I models into visual inputs for LVLMs. The task-specific performance of LVLMs on each input is quantified as fitness scores for the involved semantics and serves as reward signals to further guide LLMs in exploring concepts that induce LVLMs. Extensive experiments on seven mainstream LVLMs and two multimodal tasks demonstrate the effectiveness of our method. Additionally, we provide interesting findings about the sensitive semantics of LVLMs, aiming to inspire further in-depth research.

cs.CV↗

Evaluating Robustness of Large Audio Language Models to Audio Injection: An Empirical Study

Large Audio-Language Models (LALMs) are increasingly deployed in real-world applications, yet their robustness against malicious audio injection attacks remains underexplored. This study systematically evaluates five leading LALMs across four attack scenarios: Audio Interference Attack, Instruction Following Attack, Context Injection Attack, and Judgment Hijacking Attack. Using metrics like Defense Success Rate, Context Robustness Score, and Judgment Robustness Index, their vulnerabilities and resilience were quantitatively assessed. Experimental results reveal significant performance disparities among models; no single model consistently outperforms others across all attack types. The position of malicious content critically influences attack effectiveness, particularly when placed at the beginning of sequences. A negative correlation between instruction-following capability and robustness suggests models adhering strictly to instructions may be more susceptible, contrasting with greater resistance by safety-aligned models. Additionally, system prompts show mixed effectiveness, indicating the need for tailored strategies. This work introduces a benchmark framework and highlights the importance of integrating robustness into training pipelines. Findings emphasize developing multi-modal defenses and architectural designs that decouple capability from susceptibility for secure LALMs deployment.

cs.CL↗