Language model reasons for rejecting candidates influence their choices
Does a model's stated reason for rejecting a candidate do any work?
Computation and LanguageArtificial IntelligenceMachine Learning
Summary
Language models often explain why they reject one option by pointing out missing facts. This paper checks if adding the missing fact to the rejected option changes the model’s choice. The authors find that supplying the stated reason does affect which option the model chooses, but the effect is subtle and mixed with other factors like sentence placement and fluency. They also show that not all effects are clear-cut, and some controls behave unexpectedly. Overall, the paper shows that a model’s given reason for rejecting an option does have some influence, but it’s complex and still unclear how much content alone drives decisions.
What this means in practice
- •For natural language processing engineers: Improve model interpretability tools by verifying whether stated rejection reasons truly affect model choices during candidate ranking.
- •For ai system auditors: Develop auditing methods that test if language model explanations correspond to real internal decision processes by systematically inserting facts.
Authors
Archit Rastogi
Abstract
Asked to choose between candidates and explain the choice, a language model often rejects a rival by naming a fact its profile lacks: no director, no date of death. That sentence is a claim about the text in front of the model, and it can be tested without any judge. We insert a real corpus sentence stating the named fact into the rival's profile and ask again under greedy decoding. Two controls separate content from placement: a length-matched irrelevant sentence at the same profile, and the same two sentences at a third option the model never mentioned. In the largest of three runs, six open models on 2WikiMultihopQA, supplying the named fact at the profile the model named moves its choice more than the irrelevant control does, odds ratio 3.57 [1.54, 8.26], Holm p=0.0210, and this survives dropping any single model. The contrast the design was built to detect, the same fact at the option nobody named, does not clear correction, Holm p=0.2428. The strongest result in the family carries no content claim at all: the identical irrelevant sentence moves the choice more at the named rival than at the third option, Holm p=0.0008. Repair and control also differ in co-candidate mentions, relation template and fluency; post-hoc matching on the first two preserves the content effects' direction, matching fluency weakens one, so the content contrasts bound an effect rather than establish one. A forced single-token probability read disagrees in direction with the free-text choice on that same contrast, and three candidate explanations for the disagreement find no support. Every measurement is a string rule, so each was validated against the records it reads; validation caught eight defects. The largest, a choice-parsing rule that returned the option a model had just rejected in 17.1% of adjudicable responses, would have reported six surviving contrasts instead of four.