Unverified50% confidenceFactExact time
Confirmation bias in VLM task progress estimation manifests as follows: if the model implicitly assumes the task is heading toward success, it interprets ambiguous or failed actions as approaching the goal.
1
Sources
50%
Confidence
Long-term
Relevance
11/3/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe core assumption of the research is that the residual stream representations inside the model have already implicitly encoded information about whether a task will succeed.77% similarUnverifiedFirst verifying the model's understanding of the problem domain through explanatory tasks, then gradually transitioning to generative tasks—this phased prompting approach reduces the probability that the final output will deviate from expectations.76% similarUnverifiedAfter the model in Goal mode receives the prompt, it only estimates what to do in the first round; after completing and verifying that round, it then checks whether the task is finished, and the total number of rounds is unknown in advance.74% similarUnverifiedVision-language models (VLMs) suffer from prompt bias: when the prompt explicitly specifies a target category, the model tends to 'find' objects matching the description in the image even when they do not actually exist.73% similarUnverifiedIn scientific chart question answering, VLMs must complete multi-level subtasks such as OCR, legend recognition, axis reading, numerical comparison, and arithmetic reasoning, and a deviation in any step can cause the final answer to mismatch the ground-truth answer.72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1130296API
curl https://kongchang.com/api/v1/knowledge/claims/1130296MCP
get_claim(id=1130296)