Searcharxiv⌕ Search

arXiv subjects

Qiancheng Zhou

Publications and source records attributed to Qiancheng Zhou.

3 recordsLinked to original sources

Acting by Intent, Reacting by Sight: Visual Conditioned Reflexes for Real-Time VLM Agents

Vision-Language Models (VLMs) have shown strong capabilities as generalist agents that perceive, reason, and act in interactive environments. Yet real-time environments such as video games demand both deliberate control for exploration and split-second reactions to fast-moving hazards. Existing agents struggle to handle both: relying solely on direct VLM actions is too slow for sudden events, whereas delegating execution to fixed scripts leaves the agent blind at the moment of action, leading to frequent misfires whenever the scene shifts. Humans overcome this by pairing normal conscious control with conditioned reflexes-while the mind handles deliberate actions and shapes reflex rules, the reflex itself still relies on live vision to make the final instant decision. Emulating this mechanism, we propose ViCoR (Visual Conditioned Reflex). Alongside normal gameplay actions, the main VLM defines and updates lightweight situational guidance for recurring events, while a fast visual reflex branch (invoking the VLM in no-thinking mode) inspects live local views with motion cues to make the final sub-second action decisions. By offloading 67.8% of high-frequency reactions to the reflex branch (680 ms per check), ViCoR reduces average per-action decision latency by 2.6x (from 14.0 s to 5.4 s) and cuts token consumption by 36.9%. Across 17 diverse 2D and 3D games in the GameWorld benchmark, ViCoR achieves state-of-the-art performance, improving task success rate by +11.8% (from 52.9% to 64.7%) and average normalized progress by +13.9% (from 66.5% to 80.4%) over the strongest deliberative VLM and script baselines. In fast-paced real-time games, ViCoR refines its reflexes across episodes to boost progress by 1.67x, and even with a low-effort main planner, surpasses high-effort deliberative baselines by +8.1% at one-third of the token cost.

cs.CV↗

From Points to Targets: Online Foreground Affinity and Background Learning for Infrared Small Target Detection

Infrared small-target detection must identify a few faint target pixels in extensive background clutter while keeping false alarms low. Under single-point supervision, each target has only one labeled foreground pixel with unknown spatial extent during training, whereas the detector must recover complete targets from the infrared image without any point guidance at inference. Existing methods typically reconstruct regional pseudo-labels from local cues or intermediate predictions, which can introduce boundary noise around blurry target edges. Instead of turning uncertain boundaries into dense pseudo-masks, we propose Online Foreground Affinity and Background Learning (OABL), a single-stage framework that learns target structure directly from sparse points. First, for online foreground affinity learning, each annotated target point serves as a feature anchor, and local spatial proximity and intensity similarity weight the feature alignment between the point and surrounding pixels, guiding target-related representations without hard boundary labels. Second, for online background learning, we exclude the ambiguous neighborhood around each target point and select high-response false alarms outside this region as active negatives. Averaging the background loss over this active set keeps sparse target-like clutter influential as easy background pixels are suppressed. Jointly optimizing these objectives in a single stage updates the detector directly without generating or refreshing persistent pseudo-masks. Across three public datasets and the combined SIRST3 benchmark, OABL improves detection probability by 1.06--7.82 percentage points over matched PAL training at a false-alarm area rate of $2 \times 10^{-5}$ and avoids PAL's multi-stage full-pool label updates.

cs.CV↗

Locked at the Entrance, Open Inside: Where RLVR Narrows the Solution Space

Reinforcement learning with verifiable rewards (RLVR) substantially improves single-sample accuracy (pass@1) but causes the policy's solution space to contract, diminishing the returns of test-time scaling. In this work, we investigate where inside a reasoning trajectory this breadth is lost: does the policy fail to access a valid solution family, or does it fail to execute computation once initiated? To disentangle access from execution, we analyze the Countdown task, whose solution space can be exhaustively enumerated into discrete entrance families defined by the first operand and operator, across PPO on Qwen2.5-3B and GRPO on Qwen2.5-3B-Instruct. Across both training setups, solution coverage falls by up to 67%, halving even on problems solved across all checkpoints. We show that this contraction is heavily concentrated at the entrance: per-token likelihood shifts are 11x--16x larger prior to the first arithmetic operation than during downstream reasoning. Supplying only an unselected entrance prefix restores completion rates in low-access families by over an order of magnitude (0.018 -> 0.212 under PPO), demonstrating that alternative solutions remain executable but are no longer initiated. Guided by this localization, we find that while surface prompting fails to recover diversity, entrance-targeted interventions succeed: late-layer parameter interpolation with early checkpoints increases solution coverage by 37% at no loss in pass@1. Finally, we show that early-step entropy collapse recurs across six math benchmarks with 7B and 14B models, but is not an inevitable byproduct of reasoning optimization: an SFT baseline preserves more than double the coverage, and staged SFT--DPO--RLVR pipelines retain early-step entropy. In summary, reasoning breadth is lost at the door, not inside the room. Code: https://github.com/ershiyidian/early-branch-locking.

cs.LG↗