first-evidence-dominance-is-representational-not-decoding CLM-0053mechanismsingle paperReviewed by AI and Russ Miller
Given several conflicting but equally legitimate passages, a model's answer is driven disproportionately by whichever passage comes first, and the skew sits in the representation of the combined prompt rather than in decoding — which is why reordering the passages does not remove it and why steering the activations does.
Evidence for: First impressions last (holds)
Capability: Checking claims against evidence · Knowledge, Retrieval-augmented QA
Observed on
2026, seven models from 8B to 120B, open and closed.
Sources
- Shapley attribution over evidence pieces gives the first passage a large majority of the contribution, a layer-wise projection shows combined activations shifted toward the first evidence at every layer, a reordering control leaves the dominance in place, and steering toward an evidence-neutral centroid raises accuracy — a causal handle on the proposed mechanism.
- The prior account of which conflicting source wins: models follow the majority and lean toward their own prior. Position is an additional determinant alongside those, not a replacement for them.
Either way, bring evidence: a link, or what you saw.
Disagreeing is the most useful thing you can do here. Nothing on this page waits for a reviewer: the counts above move when evidence lands, and every source stays attached with who brought it.
Related claims
- Where a passage sits in a long input changes how much the model uses it — accuracy is highest when the needed information is at the very start or the very end and lowest when it is in the middle — so ordering retrieved passages to put the most relevant ones first is a real lever on accuracy.Losing information in long inputs · contested
- When reward models are scored by sequence log-probability on contextual preference pairs, abstentions are systematically shorter than the unsupported answers they are preferred over, so measured grounding/abstention judgement partly tracks response length; rewriting rejected and refusal responses to match the preferred ones in length while keeping the grounding failure raises pairwise accuracy far more than adding more in-domain legal preference data (up to +25.6 points over baseline for Ministral-8B).Judging another's work as well as a careful reviewer would · reviewed by AI
- When deception in narrative text is analysed by instantiating fixed argumentation schemes with an LLM (premise extraction, then critical-question evaluation), the scheme assessments are stable across ten different LLM backbones and repeated runs for clear-cut deception cases but grow inconsistent on ambiguous intelligence-style narratives.Strategic deception and detecting it · reviewed by AI
- When an LLM gives an answer plus an explanation, deleting from the input those concepts the explanation does not credit and re-querying on the reduced input raises measured faithfulness over both standard prompting and prompts that ask for faithful explanations, because the re-query removes the unmentioned factors that could have influenced the original answer.Saying what actually drove the answer · reviewed by AI
- When an LLM answers a question and gives an explanation, removing from the input the concepts its explanation did not credit and re-querying on the reduced input raises measured explanation faithfulness over both standard prompting and prompts that ask for faithfulness, because the removal deletes the unmentioned influences rather than asking the model to report them.Saying what actually drove the answer · reviewed by AI
Notes
Held apart from the claim about reordering because they are different assertions. This one says where the bias lives; that one says whether a remedy works. The dataset is partly semi-synthetic and real retrieval is untested, which is why this is single-paper and why the era is recorded.