acceptodds
Withdrawn from ICLR 2027

Access, Wording, and Knowledge Attribution in Language Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Evaluations of theory of mind ask what an agent knows, and typically separate that question from reading by pairing it with a non-mental control over the same story. We introduce a controlled diagnostic in which the epistemic rule is stated in the prompt, and which manipulates which agents have perceptual access to an event, so that a correct answer is derivable from the text rather than from commonsense. Using it, we report a replicated positive result and a boundary on it. On nine open-weight models from four families, recalling the two locations the question depends on raises balanced accuracy by 9.0 to 22.8 points, under an identical explicit instruction naming the comparison to perform. With both additions present five models reach 97.9 to 99.7, while the four that begin at or near chance reach only 59.6 to 77.6 — even though those same four score 95.4 to 99.4 when the witnesses are named and answering reduces to checking list membership. Stating the procedure is not what they were missing, and can cost up to 22.6 points when the facts are already supplied. We reach this by auditing our own control, which named the actor of the event while the knowledge question required recovering it. With designation matched we no longer detect the growth of the gap with group size that motivated the study. Seemingly minor changes in item wording matter throughout: a sentence that is true and wholly redundant costs up to 12 points, on seven models of nine. These results argue for measurement robustness before stronger conclusions are drawn about the mechanisms underlying model behaviour, and we release all prespecifications, reproduction gates and internal controls, including the internal control that failed.

Primary area:
datasets and benchmarks
Keywords:
Theory of Mind, Epistemic Reasoning, Information Access, Knowledge Attribution, Language Models, Model Evaluation, Benchmark Design, Measurement Robustness

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_1g58OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies knowledge attribution in a synthetic task whose perceptual rule is stated in the prompt. Each story places 2 to 20 agents in rooms and describes one event, and the model is asked whether a queried person knows the fact the event reveals. This question is paired with a spatial control asking whether that person is in the same room as the actor. Once the two questions designate the actor in the same way, the growth of the gap with the number of agents is no longer detected, but the sentence added for this (MAP) itself lowers spatial accuracy on seven of nine models.

The main result comes from a 2x2 design. With an instruction naming the comparison held constant, restating the two relevant rooms (RECALL) raises balanced accuracy by 9.0-22.8 points on all nine models. Five models then reach 97.9-99.7, while the four that start near chance reach only 59.6-77.6, even though they score 95.4-99.4 when the witnesses are listed; the authors describe this as a threshold. An explicit perceptual block yields almost no "knows" answers, and an audit of ToMi and BigToM finds no similar designation mismatch.

Critical strengths
  1. The recall effect is measured with the instruction held constant, which separates supplying the premises from stating the procedure. Q2 is +9.0 to +22.8 on all nine models from four families with every interval excluding zero, and recomputing the contrasts from the cell means in Table 5 reproduces Tables 3-4 to within 0.1.

  2. The authors audit their own control and report what they found, including a prespecified internal control that failed. Stories are asserted byte-identical across cells and all contrasts are paired on the same 2,200 scenarios. Replacing the actor's name with a description in the spatial question alone costs 19.4 points on gemma-2-27b, a concrete case of a control that does not share the retrieval demand of the target question.

Critical weaknesses
  1. The novelty is limited and the scope is narrow. The main result, that restating the relevant facts next to the question helps, is close to known effects of where information sits in the context [1] and to ToM methods that filter the story down to the relevant events before answering [2], [3]. The cautionary claim, that the control supplies the actor the knowledge question leaves to be recovered, is shown only for the authors' own benchmark, and the audit of ToMi and BigToM found the same problem in neither. The evidence comes from one synthetic task with a stated rule, one wording per condition and nine models up to 32B.

  2. Wording effects are as large as the effects interpreted, so the secondary claims are not robust. One true, redundant sentence (MAP) moves accuracy by 1.3-12.2 points on seven of nine models, comparable to the naming effect (-4.6 to +5.0) and Q1 (-3.6 to +13.6). The designation-matched analysis relies on this sentence, and the two conditions compared for the threshold also differ by two sentences. Each condition uses a single wording, so this variation is not reflected in the intervals.

[1] N. F. Liu et al. Lost in the middle: How language models use long contexts. TACL, 2024.

[2] A. Wilf et al. Think twice: Perspective-taking improves large language models' theory-of-mind capabilities. ACL, 2024.

[3] M. Sclar et al. Minding language models' (lack of) theory of mind: A plug-and-play multi-character belief tracker. ACL, 2023.

Questions for authors
  1. Please report the main contrasts (Q2, the instruction cost, the naming effect and the Fig. 2 difference) over several paraphrases of MAP, RECALL, INSTR and the question. If they hold across these variants, weakness 2 is largely resolved.

  2. Please run the length- and position-matched irrelevant-recall control proposed in Sec. 10. A benefit that appears only for relevant facts would separate the main result from the known effect of information position.

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

This paper reports an empirical analysis on theory of mind / social reasoning datasets.

Critical strengths

None

Critical weaknesses

The paper is not well-written and does not seem worth reviewing further.

Questions for authors

None

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

This paper examines why language models misjudge what people know. Retrieving story facts may be part of the problem. Using 2200 synthetic scenarios, the authors audit control questions and vary factual reminders and procedural instructions across nine models from four families. Reminders help every model. Balanced accuracy improves by 9.0 to 22.8 points, but four models remain below 78 percent.

Critical strengths

The controlled comparisons are useful.

  • Across nine models from four families, recalling relevant facts improves accuracy even when both conditions already state the comparison to perform.
  • Pairing 2200 scenarios and reproducing the baseline results provide useful checks on the reported comparisons.
  • The failed MAP control shows that adding a true, redundant sentence can cost 12.2 points, even when the question already names the actor and requires no extra retrieval.
Critical weaknesses
  • Figure 2 changes instructions alongside representation, and Qwen2.5-7B loses 22.6 points from instruction alone, leaving the interpretation in terms of inference confounded.
  • Nine heterogeneous models and one intermediate observation cannot establish a discontinuity without comparison against a smooth relationship.
  • I don’t think the broader impact of this audit is clear, since neither of the two benchmarks audited shows the designation confound the authors found in their synthetic task.
Questions for authors

Which conditions does Figure 2 compare? Its values appear to subtract Table 2's named scores from Table 5's scores, although Section 7 calls this a recall comparison.

For Qwen2.5-7B, REC reaches 82.2, while both reach 59.6. Does the separation persist with instruction, length, and position matched?

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

Related papers

Discussion (0)

Sign in to comment.