#01Nov 7, 2026
cs.CV
Can General-Purpose Generative AI Supply Training Data for Medical Vision Models? A Multimodal Retrospective Study
Jiachuan Peng, Kyle Lam, Jierui Qu and 10 more
Background: Well-annotated, diverse medical imaging datasets remain a major constraint for vision-based medical AI. We evaluated Nano Banana, a general-purpose image generation and editing model, as a source of synthetic medical images for model training, subgroup augmentation, and model auditing. Methods: In this retrospective study across four imaging modalities, chest radiography, fundus photography, and CT models were trained only on synthetic images and externally validated on real-world data against real-data-trained models. Carotid ultrasound models were trained on a private Human Phenotype Project cohort (aged 40-75 years) with and without synthetic augmentation. Clinicians rated image plausibility. Evaluation used AUROC, AUPR, calibration, and decision curve analysis. Findings: Synthetic-only models achieved AUROCs of 0.877 (95% CI 0.843-0.910) for pneumonia (n=14863), 0.963 (0.948-0.975) for retinal detachment (n=1149), and 0.711 (0.600-0.823) for renal calculus (n=4122), comparable to real-data-trained models, but their calibration and net benefit were not consistently more favorable. With mixed synthetic-real training, a carotid ultrasound model produced an ordinal vascular age index that stratified 633 external subjects into subgroups differing in BMI and glucose (both p<0.01), linking vascular aging to cardiometabolic health. To stress-test shortcut learning, editing a moderate tessellated pattern into fundus photographs reduced the true-class probability for central serous chorioretinopathy from 0.996 to 0.822. Clinician-guided editing raised low-quality CT plausibility from 3.10/5 to 4.30/5. Interpretation: Synthetic data can complement real data in model training, subgroup representation, and model testing via controlled editing, but its clinical value should be assessed with calibration, decision curve analysis, and clinician review, not discrimination alone.
#02Nov 7, 2026
cs.AI
TurnFlow: Benchmarking User-State Reasoning in Multi-Turn LLM Agents via Scene-Guided User Simulation
Zhixun Chen, Ming Li, Anjie Liu and 3 more
Large Language Models (LLMs) are increasingly deployed through conversational interfaces, yet strong single-turn performance does not imply reliability in multi-turn settings. Multi-turn assistance requires a model to maintain an evolving representation of user requirements and to reason over it before complying with later requests. Failures can therefore be subtle: models may fail to recognize a changed constraint or hidden conflict, or recognize it without enforcing it in the final response. Existing benchmarks primarily test fully specified prompts, long-context recall, or controlled domain environments, leaving open-domain requirement reconciliation undermeasured. We introduce TurnFlow, an interactive benchmark for evaluating whether models preserve, revise, and act on user requirements across turns. TurnFlow combines adaptive user simulation with verifiable judging, allowing conversations to unfold naturally while keeping hidden requirements and pass/fail criteria controlled. In our evaluation, nine LLMs achieve overall pass rates from 42.86% to 59.59% on 490 scenarios. Performance varies substantially across challenge types: models perform well on Stateful Multi-Constraint Following but remain considerably less reliable on Scope Creep and Goal Shift. These results reveal persistent limitations in reasoning over evolving user requirements across turns.
#03Nov 7, 2026
cs.AI
Represented but Not Used: Value-Policy Dissociation in LLM Agents
Yuzhang Luo, Mohamed Aghzal, Liangming Pan and 1 more
Large language model (LLM) agents need to choose actions whose consequences unfold over multiple future steps, yet poor decisions do not reveal whether the model failed to represent those consequences or failed to use them. Mechanistic interpretability has studied internal representations of environment state and plans, but sequential decision making also requires evaluating how candidate actions change future outcomes. Reinforcement learning formalizes this as action value. We study this computation in Qwen2.5-7B-Instruct, Qwen2.5-32B-Instruct, and GPT-OSS-20B across a partially observed Grid environment and WebShop, a simulated web shopping environment. We find that the relative values of candidate actions are linearly decodable from intermediate hidden states, including at states where the policy prefers a worse action. We call this gap between value representation and policy choice value-policy dissociation. To investigate how represented value relates to action selection, we compare direct interventions along the value readout direction with Natural Value Transport (NVT), a transformation learned from hidden state changes that naturally result from differences in action value. While readout-direction interventions have limited and inconsistent effects, NVT more consistently influences action preferences, particularly at later layers. We also test a closed-loop controller that uses the value readout as a sensor and NVT as an actuator, observing improvements in task outcomes. These results suggest that LLM agents can represent useful prospective value without reliably translating it into policy choice.
#04Nov 7, 2026
cs.CL
Test-Time Scaling for Diffusion Language Models with Frequency-Aware Remasking
Bowen Zuo, Yue Yu, Dongruo Zhou and 1 more
Diffusion language models (DLMs) generate text through iterative denoising, where decisions about which tokens to commit or remask strongly affect output quality. Existing certainty-aware strategies rely on instantaneous confidence or entropy, but these local signals can be misleading on difficult reasoning problems. We formulate DLM test-time scaling as adaptive allocation of both inference compute and decoding strategies. Our two-stage approach uses a fast base sampler to solve easy questions, then applies frequency-aware remasking to the remaining hard questions. This strategy tracks how often each token's prediction changes across denoising steps and preferentially commits frequently changing tokens, encouraging samples to explore alternative reasoning paths rather than collapse onto the model's preferred trajectory. Experiments with LLaDA-8B-Instruct and Dream-7B-Instruct show consistent improvements over a strong adaptive allocation baseline, with absolute accuracy gains of up to 3.73 percentage points (LLaDA) and 7.33 percentage points (Dream) on MATH-500, and absolute coverage gains of up to 11.39 percentage points (LLaDA) and 10.98 percentage points (Dream) on HumanEval. On questions unsolved by the base certainty-aware strategy after 100 samples, frequency-aware remasking recovers solutions for 3.75% to 18.75% of examples.
#05Nov 7, 2026
cs.HC
Safer for Whom? A Systematic Exploration of AI Safety in Eating Disorder Interactions
Jinyuan Wang, Minh Duc Chu, Frances Ross and 2 more
AI assistants advise users about diet, exercise, and weight loss---areas in which unsafe guidance can worsen body dissatisfaction, dietary restriction, compulsive exercise, and other behaviors associated with eating disorders (EDs). Existing evaluations emphasize diagnostic accuracy, overtly harmful prompts, or synthetic scenarios, providing limited insight into whether models recognize implicit vulnerability in natural interactions or protect users consistently. We evaluate responses from three open-weight large language models to 100 questions sourced from an online pro-anorexia community. We grade responses using clinically informed criteria that extend a seven-dimensional taxonomy of ED-related harms. We also construct matched conditions in which questions are accompanied by information about the user varying race, gender, body mass index (BMI), and explicit ED vulnerability. Safety varies substantially across models and harm dimensions. High-risk responses are approximately twice as frequent without bios, suggesting that models do not consistently recognize vulnerability from language alone. Explicit signals of ED vulnerability improve safety by 10--14 points across models. Responses also differ significantly by race, gender, and BMI, with body size producing the largest disparities. These patterns are consistent with reliance on the stereotype that EDs primarily affect thin, White women, potentially leaving users outside this profile with less protective responses.