Searcharxiv⌕ Search

arXiv subjects

QingHao Zhang

Publications and source records attributed to QingHao Zhang.

3 recordsLinked to original sources

Do the Details Help the Whole? One-Way Transfer from Person Actions to Group Intent in VLMs on JRDB-Social

Social scene datasets label people at two levels: what each person is doing and what the group is there for. Training on both levels at once is often assumed to help each of them. We test this assumption for vision-language models (VLMs) on JRDB-Social, treating person actions as the details and group intent as the whole. We fine-tune Qwen3.5 4B under eight training arms with three seeds, repeat the core arms on Qwen3.5 9B and Qwen3-VL 4B, and judge every arm with a verdict rule fixed before testing. Writing each member's actions before the intent raises intent macro-F1 from 30.1 to 31.0, and writing the intent first reaches 31.2, while action F1 stays within 0.2 points of an action-only model. Under the selected recipe, no arm trained with intent improves actions. The gain requires correct, image-grounded actions: empty action lists or actions borrowed from another group with the same intent lower intent F1. Forced-prefix tests, occlusion, member masking, linear probes and component grafting show that the intent follows the written actions, relies more on people and partly resides in the vision tower. Supplying actions as input instead of as targets hurts both levels.

cs.CV↗

Can We Make an LLM Dumber, or Is It Just Playing Dumb? An Empirical Study of Believable Performance Throttling

A benchmark score is commonly read as a proxy for model capability. We study the inverse problem: can a model be driven to a target score, and do its failures resemble those of a naturally weaker model? We operationalize believable performance reduction along three axes: target calibration, item-difficulty structure, and error-distribution similarity to a real weak reference. We introduce CapLadder, an IRT-calibrated suite of 1,985 items spanning mathematics, code, multiple-choice reasoning, and tool use. A simple weak-teacher LoRA dial reaches unseen target levels on reasoning with a 1.9-point held-out calibration error, whereas prompting, temperature, and weight noise cannot combine comparable range and precision. Yet on multiple choice, the dial converges to an answer-token attractor: up to 97% of its newly introduced errors land on one letter, with the attractor changing from D in Qwen to A in Llama. Logit interventions identify a dominant near-constant shift: subtracting the estimated shift on held-out items recovers 52 to 59% of lost accuracy and removes the dominant attractor, while option permutation confirms that the bias tracks token position rather than option content. With high inter-annotator agreement ($κ=0.94$), human-coded error-category distributions nonetheless match the weak reference on reasoning (JS = 0.00), showing that local plausibility and aggregate detectability differ. Structured tool calls exhibit more natural degradation in the Qwen family (JS = 0.09), while a globally calibrated dial transfers poorly across benchmarks (14.2-point error on two usable benchmarks). Precise score control, therefore, does not imply a portable or believable reduction in skill.

cs.CL↗

Guides That Cause Actions: An Offline Study of Guide-Action Mutual Reinforcement in Multimodal Web Agents

Web agents are usually evaluated in live environments, where environment state and judge models drift between runs, so the same checkpoint rarely reproduces the same score, making controlled studies of training phenomena impractical. We present WebMRE, an offline benchmark of 541 tasks and 5,293 steps derived from successful WebArena trajectories, with fully audited test labels and a deterministic protocol that scores a checkpoint identically on every run without any environment. Each step pairs a human oriented guide sentence with a grounded action, enabling the first study of the mutual reinforcement effect between them in web agents. Averaged over three seeds the effect holds for both models in both decoding orders and grows with scale: jointly decoding a guide lifts element selection over an action only reference by 0.9 and 0.2 points for Qwen3.5-4B and by 1.7 and 2.2 points for Qwen3.5-9B. A mediation analysis shows that the guide is a causal channel rather than commentary: forcing the gold guide as a decoding prefix lifts action accuracy from .422 to .684, another step's guide collapses it to .055, and a paraphrase that renames the target still recovers half of the gain, so the channel carries instruction meaning and not only the label string. The same channel yields an offline reward that only a replayable protocol makes computable, though optimizing it from a strong checkpoint brings no gain yet. Our fine tuned models outperform GPT-5.5, Claude Opus 4.8, and Gemini 3.5 Flash, run zero shot, on every offline metric.

cs.CL↗