acceptodds
Under review as a conference paper at ICLR 2027

Multiple Task-State Descriptions Interfere with In-Context Computation

Abstract

Adding a second description of a task state lowers language-model accuracy on a downstream question about it, even when the two agree and even when one is an exact copy of the other. When one description updates the other, the model fails on most items where it can separately resolve the current state and compute from it, once it must do both together. We call this a composition penalty. Neutralizing one token’s representation in the obsolete description restores most of the lost accuracy, showing that the obsolete state remains causally active. Duplication and stale-state costs recur in verbatim shell-tool output, and on an independently written benchmark the penalty persists on items whose state the model resolves correctly, including for Claude Opus 5.5, which solves every controlled chain. In nineteen novels, deleting the sentences that support an obsolete state improves accuracy more than deleting as much other text. Reasoning closes some controlled gaps but leaves a gap on long histories. Compiling the current state into a single description before computing nearly eliminates the penalty in the controlled setting. Its end-to-end benefit depends on compilation fidelity: on long histories, inaccurate compilations can outweigh the benefit and reduce overall accuracy.

Primary area:
foundation or frontier models, including LLMs
Keywords:
large language models, in-context computation, state tracking, contextual interference, compositional reasoning, activation patching

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_XoaXOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies if coexisting descriptions of a "task state" impair the downstream results, even if the descriptions are identical. The authors conducted controlled experiments on this problem and the results reveal that current models can roughly identify updated facts yet still fail to use them, and even repeating identical information would negatively affect their performance. Rewriting the context into a single source-of-truth description often helps but can also affect performance on longer horizon when the rewrite itself induces errors.

Critical strengths
  • the central claim of the paper is well supported by the results, and the controlled experiments were relatively well executed
  • the authors also report failed compilation/etc cases and are honest about the boundary of these experiments and attempted solutions, which should be appreciated
Critical weaknesses

The main complaint is the controlled experiment might imply a narrower scope than the headline suggest and primarily limited to the models that have been evaluated. If such a narrowly scoped observation study merits enough contribution per ICLR standard. To push it further is if such kind of observation study can offer deeper insights to the community on how we could amend this type of failure in real-world LLM uses.


Update: Thank you to the authors for the detailed response and the effort of updating the manuscript and adding new information about my question. The new budget comparison has addressed one of my concerns and I appreciate the effort to revise and sharpen the intro and abstract. Therefore I've increased my overall recommendation score.

Questions for authors

From the reasoning budget lens, does rewriting the context before answering still outperform, e.g., simply letting the model reason longer?

2 comments

Thank you for your positive feedback and for reconsidering your assessment. We appreciate your time and thoughtful review.

Area_Chair_gtrwOpenReview ↗

Please engage with the authors during the discussion period.

Dear reviewer, The authors have provided a response to your review. Please ensure that you have read it and taken it into consideration when making your final recommendation. Even if you do not change your score, please post a response with your rationale.

Reviewer_1GgxOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary
  1. The paper looks at how having multiple descriptions of the same task state in context can hurt an LLM’s accuracy on that task.
  2. They show that this is even the case if the descriptions are identical.
  3. They introduce “composition penalty” for the case where one description updates another: models can correctly work out the current state and compute from it separately, but fail when asked to do both together.
  4. They first demonstrate this with a synthetic sensor-counting task before looking at related effects in different domains (shell-tool output, STALE, and novels from OAKS).
  5. They perform several controlled comparisons, activation patching, and targeted evidence deletion to investigate what drives the dip in accuracy.
  6. They propose replacing the history with a compiled current-state description before answering, which largely closes the controlled gap.
  7. However they also show: a. Letting the model reason also closes some controlled gaps, though the authors report a gap remaining on long histories. b. With longer histories, inaccurate compilations can make performance worse.
Critical strengths
  1. Depth into the problem / controlled experiments a. The amount of controlled experiments to explore / rule out alternative explanations was great (e.g., testing prior commitment and context length, and showing through identical duplicates that conflict isn’t needed for the accuracy drop). All the experiments are well motivated and the methodology seems sound and well defended (AppF showing why the more obvious patching setup would have been misleading particularly) b. It made the results a lot more convincing and a bit baffling (in a good way). c. Extension to the other domains (shell output, STALE, and the OAKS novels) gave me confidence this isn’t just an artifact of the synthetic setup.
  2. Honesty and good practice a. Appreciated the reporting and discussion of the results that don’t fit neatly (compile-then-compute sometimes hurting on STALE). b. This made me trust the positive results more and also made the weaknesses below feel more like scoping issues than red flags.
Critical weaknesses
  1. Some results need clearer separation (minor) a. STALE’s headline number (resolved-state vs. history accuracy, T4) is not the papers main defined composition penalty (E1) as it mixes obsolete-state interference with retrieval and other long-context demands. b. The actual full-population gap looks negative (~ −0.48 for DeepSeek-V4-Pro, −0.11 for Opus 5.5). c. The conditional results support it but they come from a subset selected on an outcome. d. This bit should be a bit clearer as the long-history results are a key part of the case that the problem persists beyond smaller models and constrained decoding

  2. There is a lot going on and it’s not clear to me the cost is specifically about composing the state change the question needs a. AppJ shows decision-irrelevant updates cause as big a penalty as decision relevant ones, this fits the “multiple descriptions interfere” narrative but not quite “composing the state change the question needs”. b. Also you show that reasoning closes the gap in some settings (AppI and AppJ), but the results in AppQ are with thinking disabled c. This makes it a bit unclear whether compiling beats just letting the model reason on the same items d. Think I need a bit more convincing it’s really about composition or some broader cost of processing structured context.

Questions for authors
  1. How do you reconcile App J with the irrelevant maintenance log being fairly harmless in Sec 3.1? Does it change how you'd define the composition penalty?
1 comment
Area_Chair_gtrwOpenReview ↗

Please engage with the authors during the discussion period.

Dear reviewer, The authors have provided a response to your review. Please ensure that you have read it and taken it into consideration when making your final recommendation. Even if you do not change your score, please post a response with your rationale.

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

This paper studies why language models struggle to compute from a context containing multiple descriptions of the same task state. Its controlled experiments separate resolving an updated state from answering a question about that state. Models often succeed at both components separately but fail when asked to combine them; even an exact duplicate of a correct report can reduce accuracy. Activation patching provides evidence that the obsolete-state representation affects the answer. The paper also examines related effects in shell output, STALE histories, and novels. Compiling the context into a single current-state description largely removes the controlled composition penalty, but its usefulness on longer histories depends on whether the compilation is accurate. The main contribution is an empirical account of this failure mode and its boundaries.

Critical strengths
  • The separation of state resolution and downstream computation is convincing. In Section 5, the joint-task deficit remains on items where the model both writes the correct state and answers correctly from that state. This is stronger evidence than an aggregate comparison between an easier prompt and a harder one.
  • The paper tests several plausible alternative explanations. Exact duplication removes disagreement between sources; the irrelevant-context control addresses added length; and the restatement controls in Table 2 show that an additional model call alone does not reproduce the benefit of compilation. Together, these experiments make the controlled finding quite persuasive.
  • The evidence extends beyond the authors' synthetic reports. The novel experiment is particularly useful because it compares deletion of annotated obsolete evidence with deletion of similar amounts of other text. I also appreciate that the paper reports where its proposed solution fails: on STALE, compilation reduces end-to-end accuracy in four of the five configurations in Table 4.
Critical weaknesses
  1. The connection to long-history failures remains incomplete. Reasoning can eliminate the short-task gap, while in the shell experiment an unrelated listing costs 0.306 accuracy versus 0.421 for a stale listing. On STALE, replacing a history with a resolved state also removes substantial processing demands. The novel deletion experiment provides more direct evidence, but the advantage over matched deletion is smaller, at 0.044. The evidence supports stale-state interference as a contributor, but does not establish how much of the broader deficit shares the controlled mechanism.

  2. Patching establishes causal relevance more clearly than removal of interference. The donor replaces the old value with the current one, both removing conflict and reinforcing the correct state (Section 4). As Appendix E acknowledges, the unchanged-prefix control does not match the activation change. This leaves some ambiguity in interpreting the intervention as “neutralizing” obsolete information. The result convincingly shows that the token matters, but does not uniquely distinguish suppression of the old state from reinforcement of the new one.

  3. The remedy has limited demonstrated benefit beyond short states. Compilation lowers STALE accuracy in four of five configurations (Table 4), and the verifier-gated route gains only 0.010 (Section 8). Replacing the history with an inaccurate compilation discards information that could otherwise support recovery. These limitations are acknowledged, but they leave the paper with a stronger diagnosis than a practical solution for accumulated agent histories.

Questions for authors
  1. On shell-output tasks, how do stale, duplicated, and matched-irrelevant contexts compare when the model can reason before answering? Which differences persist under that response protocol?
  2. Does suppressing the obsolete value without inserting the current-value activation recover similar accuracy?
  3. Could retaining access to the history when compilation is uncertain improve on the current verifier-gated route, while preserving the benefit on faithfully compiled items?

Comments on the paper

1 comment

Dear Area Chair and Reviewers,

Thank you for the constructive discussion. We show that correctly identifying the current facts does not guarantee that a model can use them to answer a question. Models can succeed at both tasks separately yet fail on the same items when answering from the update history. We observe this gap when models answer without first generating intermediate reasoning. Even identical state descriptions can reduce accuracy.

  • Stronger causal evidence. In new experiments, blocking attention to obsolete values improves accuracy across three model families without inserting corrected activations. Targeted exclusion outperforms the matched-control mean in each model. Deleting obsolete evidence in novels also outperforms matched deletion (Sections 4 and 9; Appendix R).

  • Clearer scope. Additional budget and large-model evaluations show that reasoning can close short-task gaps. We distinguish the controlled composition penalty from the broader history/state gap that remains in STALE (Sections 5–7 and Appendix V).

  • Practical implications. Deterministic tracking reconstructs the tested shell states exactly from supported commands. In an exploratory comparison on all 400 STALE scenarios, giving the compiler the question improves Kimi's end-to-end state consistency by 9.25 percentage points over direct answering. The pipeline uses two calls and more output tokens (Section 6 and Appendix V.2).

We incorporated the new results into the manuscript, revised it for clarity, and updated the supplement with inputs, responses, and reproduction code. Thank you again for your thoughtful feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.