When a language model reads the same fact, the way it learned that fact changes how strongly it sticks with its opinion when new information arrives. For example, reading a summary of coin flips and actually seeing the sequence of heads and tails both produce the same current guess, but the summary-given guess is far less persistent against new evidence than the sequence-given guess. This matters because it means current agreement across models or prompts does not guarantee future agreement, and systems that rely on matched outputs as a sign of robust belief need to test updating under new evidence.
Same answer from different prompts does not mean same belief updating in language models
Li et al. show that beliefs induced via observed sequences, summaries, or activation steering carry different persistence when new evidence arrives
Top University
Weihan Li · Junhao Wu · Xinlei Chen · Tianshi Zheng
University of Tokyo · RWTH Aachen University · Harbin Institute of Technology (Shenzhen) · The Hong Kong University of Science and Technology
Research Digest··3 min read
In one line
Matched predictions from different belief-setting routes produce different belief-update dynamics in language models.
Before you read this
Three ideas this paper builds onactivation steeringActivation steering modifies a model's internal vector representations during inference to shift behavior, such as editing outputs, removing capabilities, or controlling trust.64 papers →bayesian updateA method for updating the probability of a hypothesis based on new evidence.16 papers →confidence calibrationConfidence calibration aligns a model's predicted confidence with its actual accuracy across predictions.82 papers →
Read it as
The authors tested three open-weight language models on a coin-flip bias inference task, preparing beliefs through four textual routes (observed sequences, count summaries, source assertions, instructions) and two activation interventions (steering, state transplantation).
The authors designed a Bernoulli inference task where a language model must estimate the bias of a coin from observed flips, with an analytic Bayesian posterior as reference.
How it works
before and afterThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.