acceptodds
Under review as a conference paper at ICLR 2027

Evidence States Survive Where Answers Fail: Decodability and Causal Access in Language Models

Abstract

A record can support a claim, its explicit negation, both, or neither. Preserving these distinctions is essential for evidence reporting. We separate evidence-state recoverability, native reporting, and causal access using entity-role swaps that preserve the exact word multiset across all four support states. Across three instruction-tuned model families, the output contrast between absent and dual support is attenuated relative to the contrast between opposing one-sided support. On held-out signed records, same-query linear decoders distinguish absent from dual support with 87.5–100% accuracy, compared with 50.0–75.0% for native yes/no readout. This advantage persists over scalar decoders trained on identical examples and state labels. Residual interventions reveal task-conditioned causal access: a within-query state intervention in Mistral exceeds an equal-norm random control by 1.79 log-odds units. Averaged over both complementary questions, transferred directions underperform random controls on signed records. On unseen source- and time-scoped records, Mistral retains strong hidden-state decoding, while Qwen and Olmo show weaker transfer. Together, these findings establish a recoverability–reporting gap and identify question- and model-dependent boundaries on its causal portability.

Primary area:
interpretability and explainable AI
Keywords:
mechanistic interpretability, evidence-state representation, linear probing, activation interventions, knowledge conflicts

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_9eDZOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

Paper studies whether LLMs distinguish between a record that supports neither of two claims and a record that supports both. The authors build synthetic records where swapping entity roles changes the evidence state while keeping the same words. They show that the native yes/no answers distinguish absent from conflicting evidence poorly, while linear probes on hidden states do it almost perfectly. They then test steering along the probe directions and transfer to a harder task with source and time scope.

Critical strengths
  • The question of whether models separate missing from conflicting evidence is relevant for RAG and evidence reporting.

  • The stimulus construction is clean, and controlling for the exact word set rules out simple lexical shortcuts.

  • The authors include reasonable controls, such as random labels, text baselines and matched scalar decoders.

Critical weaknesses

The main finding is expected.

That linear probes recover more than the model answer is already known from prior work on latent knowledge and truth directions. The paper applies this to a new label set, but I do not see a new insight about how models process evidence.

The setup is very limited.

All experiments use short synthetic records with invented entities, a single yes/no token, and three 7B models. The test sets are small and probes are near ceiling, so it is hard to judge differences. I am not sure the results tell us anything about real documents, longer contexts or free generation.

The baselines are weak.

The scalar comparator is a classifier on a single output margin, so it is expected to lose to a classifier on the full hidden state. Simple prompting changes, few-shot examples, or asking the model to state the evidence state directly are not tested, although they may close the gap at the output level.

The causal results are mixed.

Steering works within one question for Mistral but is weak for Qwen. Directions do not transfer across questions, and for Qwen a plain answer direction works better than the state direction. Transfer to the source-scoped task works only for Mistral. So the causal claims are narrow, and the paper does not explain why the models differ.

The writing is hard to follow.

The text introduces many custom terms and reports results mostly as raw counts. A clearer presentation of the main result would help.

Questions for authors
Reviewer_cA8POpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper asks whether a language model retains distinctions between missing and conflicting evidence when its answers fail to report them. Entity-role swaps create four support states while preserving the exact word multiset. Across three instruction-tuned model families, hidden-state decoders are compared with native yes/no scores and scalar decoders trained on the same examples and labels. Hidden features distinguish absent from dual support more accurately on the controlled records. Donor-based residual interventions and complementary questions probe how that information influences answers. Transfer to source- and time-scoped records is strongest in Mistral. The findings establish a decoding gap, with more limited and model-dependent causal transfer.

Critical strengths
  1. Entity-role swaps alter evidence binding while preserving word counts, including negations. This construction addresses a simple lexical explanation for differences between the four support states.

  2. Hidden and scalar decoders share training examples and labels, and their features come from the same query computation as the native answer. This makes the decoding advantage more informative than a comparison with native output alone.

  3. Complementary questions and direct answer-steering controls sharpen the interpretation of patch effects. Grouping related records by family also respects the dependence built into the data construction.

Critical weaknesses
  1. The positive Mistral patch result shows movement toward a donor with a known state. On signed records, the comparison with direct answer steering is unresolved, and cross-question directions underperform random controls. The causal evidence therefore has a narrow scope.

  2. The strongest decoding result concerns absent versus dual support. Full four-state decoding is weaker and question-dependent, leaving less complete evidence for recovering the full support state than for separating these endpoints.

  3. All evaluations use synthetic records and a restricted yes/no interface. Source transfer also changes the answer cue for Qwen and Mistral, leaving evidence-selection difficulty and interface sensitivity mixed in that comparison.

Questions for authors

1.Which result separates evidence-state access from steering toward the donor's answer in the signed task?

2.How much of the source-transfer difference for Qwen and Mistral comes from the answer-cue change?

Reviewer_1bzXOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper advocates the separation of evidence-state recoverability, native reporting, and causal access using entity-role swaps and demonstrates the effectiveness in LLMs. The idea is well justified, simple, and seems working.

Critical strengths

The idea is simple, principled, and clearly described.

The experimental results demonstrate the effectiveness of the idea.

I particularly like the discussion on limitations, very informative.

Critical weaknesses

From the discussion on related work, it is unclear whether any similar or partial ideas have been investigated in literature. It would be helpful to make this discussion more explicit and direct.

What would be the implications and applications of the proposed techniques here to the downstream tasks?

Questions for authors

There may be some tricky cases other than that shown in lines 162-163, such as X is NOT glossy. Y is triangular. Would those cases be relevant? A similar case is mentioned in A.1, but the discussion seems not clear.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.