acceptodds
Under review as a conference paper at ICLR 2027

Not Proven Is Not Proven-Not: Language Models Read the “Not” in a Question as Negation as Failure

Abstract

Language models are routinely asked about texts that leave some facts open, and then the right answer is unknown. We find that a “not” turns such gaps into confident answers: when a text mentions a property P but never settles whether x has it, models judge “x is P” unknown yet “x is not P” true. The models read the “not” as negation as failure, taking “it is not established that x is P” for “it is established that x is not P”. Both questions look P up in the text, and only the negated one reads a failed lookup as an answer. Editing the text moves the error where this account predicts: removing the property word largely removes it, a rule that merely mentions the word brings it back, and writing “not P” into that rule spreads it to the plain question. Inside the model, the lookup runs through the attention of the question's property word to the text, and while answering true the model still registers, and even says, that the fact is not established. Across fourteen open models the asymmetry tracks this lookup, which is present before instruction tuning. Instructions mostly make models abstain, while thinking replaces the lookup with a proof. What is not proven is read as proven not.

Primary area:
foundation or frontier models, including LLMs
Keywords:
large language models, negation, negation as failure, reasoning under incomplete information, closed-world assumption, interpretability, attention interventions, abstention

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_977jOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper investigates LLM errors (we could call them hallucinations) in the context of negation. The main takeaway is that by default their outputs change depending on whether a property is mentioned, and whether the mention is negated. This change in the prompt / input leads to different LLM outputs: unknown or not (when answering a question). In other words, LLMs sometime (wrongly “assume” that a property is not true when nothing is stated about that property for a specific entity. The authors work with 14 LLMs

Critical strengths
  • LLMs generally underperform when the input has a negation, and understanding why is certainly important. Building on previous work, the motivation to work on negation is solid.

  • The authors present results with 14 LLMs. That is certainly impressive, but I have to wonder why those 14.

Critical weaknesses
  • Unless I am missing something, the main conclusion is that stating a property holds (for some entity) makes LLMs hallucinate that the same property does not hold for another entity. That seems thin to me for a full paper.

  • Following up with the above, would explicitly stating how to treat negation in the system prompt (or input prompt) solve the problem? For example, what happens if you prompt with “Answer the question below based on the facts below. Do not assume anything except what is explicitly stated.”

  • While the quantitative results are extensive (14 LLMs), what I learned is what these LLM output in the very narrow problem the paper presents. There are no actual causes, but rather some extensive LLM testing (observing outputs given inputs).

  • 264: “and pretrained checkpoints already show the lookup”. Where is the evidence for this, and what checkpoints are the authors referring to?

  • What are the chat templates in 297?

The paper is very hard to follow, to the point that it is a major weakness and not just a matter of a few quick edits

  • “both questions look P up in the text, and only the negated one reads a failed lookup as an answer” LLMs don’t look for or at anything, they process the input text (always, regardless of what the input is). LLMs don’t read per se, and if we want to say that processing the input is “reading”, they read whatever is given to them in the input. They don’t choose what to read, even if they appear to ignore a chunk of the input based on observations of their outputs.

  • “Editing the text moves the error where this account predicts: removing the property word largely removes it, a rule that merely mentions the word brings it back, and writing “not P ” into that rule spreads it to the plain question.” I think they mean that LLMs output change if the reverse the polarity of a statement (the statement states a property, based on some examples in the paper), but that wording is extremely convoluted.

These two examples are from the abstract alone. The rest of the paper is not much better. Many terms are undefined or explained in any reasonable detail. For example, the title in 5.1 (“The Lookup Is the Property Word’s Read”) is hard to parse.

The last two paragraphs in Section 2 seem to be a draft with no direct connection to the remaining of the section. “Rule texts” have not been defined. I can’t tell what “open” means in 128.

Here is another example of weird word choices: 144 lists “rule texts, everyday records, and human-written paragraphs alike”. Yet in 156, they state “We wrote 240 records […]”. So the records are also human-written (by the authors)? I find the paper very hard to comprehend.

From the conclusions: “When a model does not find X in a text that mentions P, it keeps the plain question open and judges the negated one true, reading what is not established as established not.” I could guess, but what is X? Keeping a question open means to not answer? That would contradict the paper (the LLM always answer?). “Reading” should probably be rephrased as “assuming”, “making up”, or “hallucinating”.

There are too many examples to list. The paper needs to be rewritten.

Questions for authors
  • Can you justify the need for 14 LLMs, and specifically those? I don’t have a problem, but I do wonder why those.

  • Are you missing a reference in 264 (definition of lookup effect)?

  • See weaknesses. If a question raises to the level of a weakness, I wrote it there.

Reviewer_mtZ9OpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates a distinctive failure in language models' reasoning under incomplete information: models may judge an affirmative proposition as unknown while incorrectly accepting its negation as true. The authors attribute this asymmetry to negation as failure (NAF). Experiments across 14 open-weight models, controlled text interventions, and internal activation analyses characterize the phenomenon and investigate its underlying mechanisms. The paper also evaluates prompting and thinking-enabled inference as mitigation strategies.

Critical strengths
  1. Novel and meaningful problem discovery. The paper identifies a distinctive asymmetric negation failure under incomplete information. On 8,654 open ProofWriter pairs, Qwen3-32B incorrectly accepts the negated proposition while leaving the affirmative unknown in 2,387 cases, versus only 60 cases showing the reverse pattern. The additional record and ANLI experiments support the relevance of this phenomenon beyond synthetic rule reasoning.
  2. Well-designed interventions for investigating the mechanism. Section 4 systematically manipulates predicate occurrence, rule conditions, and entity associations. For Qwen3-32B, masking the target predicate reduces committed negated answers from 86% to 17%, while restoring a rule mentioning that predicate increases negated commitments by 52 percentage points but plain commitments by only 3 points. These controlled experiments provide compelling evidence of context-dependent retrieval effects.
  3. Multi-level mechanistic analysis and mitigation evaluation. Section 5 combines attention blocking, residual-state transplantation, and activation probing to investigate the internal computation associated with the error. Section 6 further distinguishes increased abstention from improved reasoning, showing that thinking raises Qwen3-32B's known-item accuracy from approximately 61% to 88% while substantially reducing the asymmetry.
Critical weaknesses
  1. Insufficient identification of the proposed NAF mechanism. Although the experiments establish a strong connection between contextual retrieval and negation errors, they do not uniquely identify negation as failure as the underlying computation. In particular, Table 21 reports more additional plain-true than plain-false responses for Qwen3-32B after negating a rule condition, which is not fully explained by the proposed mechanism.
  2. Limited specificity of the internal causal evidence. The attention-blocking experiments identify an important evidence-access pathway but also substantially disrupt correct factual retrieval. For Qwen3-32B, correct answers to directly stated facts decrease from 203/204 to 3/204. Despite the additional transplantation and ablation analyses, the specific computation that converts retrieval failure into an affirmative negation judgment remains insufficiently isolated.
  3. Limited methodological contribution beyond mechanistic diagnosis. The paper completes a coherent research cycle of problem discovery, mechanistic explanation, and evaluation of existing mitigation strategies, but does not develop and validate a new method informed by its findings. Section 6 primarily examines prompting and existing thinking-enabled inference. While a novel algorithm is not necessary for a valuable empirical contribution, the absence of a targeted, mechanism-driven solution limits the demonstrated practical impact of the findings.
Questions for authors
  1. How does the proposed NAF mechanism explain the unexpected increase in plain-true responses observed in Table 21?
  2. Which experimental evidence most directly distinguishes NAF from alternative context-dependent retrieval or response heuristics?
  3. Can the authors clarify how the identified mechanism could guide targeted mitigation beyond existing prompting and thinking strategies, based on the current findings?
Reviewer_6oKNOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper investigates how LLMs incorrectly interpret missing evidence as evidence for negation. The authors attribute this behavior to negation as failure (NAF) and study it across multiple models using controlled text edits and mechanistic interventions.

Critical strengths

The paper identifies an interesting behavioral phenomenon and provides extensive experiments across different models and settings. The controlled interventions also offer potentially useful insights into how LLMs process negation.

Critical weaknesses

1. Major presentation and clarity issues.

I found this paper very difficult to follow, even after reading the main text carefully. The writing is extremely dense, starting from the abstract, and many important concepts are introduced without sufficient explanation or motivation.

More importantly, the paper does not clearly distinguish between empirical observations, hypotheses, and mechanistic conclusions. The discussion moves between NAF, lookup mechanisms, internal representations, and proof-based reasoning without presenting a sufficiently coherent overall picture.

Although some individual findings are interesting, I struggled to understand the complete idea the authors are trying to convey and how the different experiments fit together. This substantially limits my ability to properly assess the paper's contributions. I believe the paper needs significant improvements in writing, structure, and conceptual clarity.

2. Insufficient justification for the proposed mechanism.

The mechanistic interpretation appears stronger than what the experiments establish. For example, Section 5.1 concludes that affirmative and negated questions share a lookup process, with negation affecting its interpretation afterward. However, the presented interventions do not clearly distinguish this explanation from polarity-dependent processing that occurs earlier. While the results are interesting, I am not convinced that they establish the specific computational mechanism claimed by the authors.

Questions for authors
  1. Could the authors provide a clearer, step-by-step explanation of the central idea, explicitly distinguishing observations from hypotheses and mechanistic conclusions?
  2. What evidence specifically distinguishes the proposed NAF mechanism from simpler context-dependent heuristics?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.