Confirmation Bias
A concept from cognitive psychology referring to the tendency of individuals to seek, interpret and remember information that supports existing hypotheses while ignoring contradictory evidence. In AI models it manifests as a tendency to confirm expected outcomes while neglecting failure signals.
Timeline (last 90 days)
Existing VLM-based task progress estimation frameworks generally suffer from confirmation bias, readily hallucinating progress during failed executions.
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.
Traditional pseudo-labeling suffers from confirmation bias: repeatedly using pseudo-labels from the same model causes biases to accumulate and amplify with each iteration.
All Facts (3)
Existing VLM-based task progress estimation frameworks generally suffer from confirmation bias, readily hallucinating progress during failed executions.
50%UnverifiedConfirmation 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.
50%UnverifiedTraditional pseudo-labeling suffers from confirmation bias: repeatedly using pseudo-labels from the same model causes biases to accumulate and amplify with each iteration.
50%