acceptodds
Under review as a conference paper at ICLR 2027

Remembered but Not Used: Understanding In-Context Update Failures in LLMs

Abstract

As preferences, goals, and facts change, LLM agents must use the current state while earlier versions remain in context. Yet they can answer with an old value of the same variable, a failure that we call **stale binding**. To study when models use outdated information and why, we introduce Controlled In-Context Memory (CICM), a benchmark for tracking and using updated information in conversations and agent logs. We observe that even frontier reasoning models can fail to recover the current state. We find that in open-source models probes can still recover the updated value when the model answers with an old one, pointing to a failure to select information that remains available. Component tests in Qwen and Pythia identify a mechanism for this selection failure: attention drift, where attention favors old values over the current one when producing an answer. We study a one-layer transformer to mathematically understand how this phenomenon happens: when attention scores are similar, several old values can together receive more attention than the current value. Guided by this explanation, we redirect attention toward the current value without further training. When the current value is requested directly, adjusting this intervention for each input corrects most old-value errors across various model families while preserving nearly all initially correct answers. Reliable context management therefore requires more than remembering updated information: models must use it to guide their answers.

Primary area:
generative models
Keywords:
Large Language Models, Contextual Management, mechanism analysis, attention mechanism

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_pa54OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies "stale binding": when a variable is updated several times in context, an LLM answers with an old value of the same variable. It introduces CICM, a set of preference dialogues, constraint decisions and operational logs with programmatic labels. Linear probes are used to argue that the current value stays readable when the model answers with an old one, and the failure is attributed to "attention drift" toward old values (Qwen, Pythia). A one-layer softmax/RoPE analysis explains how several old values can outweigh the current one, and test-time attention routing shifts attention from old-value positions to the current value's positions.

Critical strengths
  1. The problem is real and relevant to agents whose context keeps changing.
  2. The experiments are extensive, and the limitations are reported candidly in the appendix.
Critical weaknesses
  1. Limited novelty over PI-LLM. That repeated updates to the same key lead the model to return old values is already established by PI-LLM (Wang & Sun, 2025), and the synthetic overwrite task in Section 2 is close to its setting. The new parts are the name "stale binding" and the internal analysis, which has the problems below.
  2. The theory is trivial. When several old values compete in the softmax, their combined attention naturally takes mass from the current value; this follows from QK similarity and needs no derivation. Propositions 3 and 7 amount to adding a term to the softmax denominator. The authors also note that their measured QK margins do not estimate B(δ) (Appendix A.1), so Section 5 does not connect to the experiments.
  3. The test-time repair needs the answer in advance. Attention routing (Section 6) relies on a rule-based parser that already locates the current value, and Appendix E.1 admits the parser alone could answer the question. If routing could truly move a model from old to new information, the interference problem would be solved; what it shows is only that attention can be steered once the answer is known. It is a diagnostic, not the "training-free repair" claimed in the abstract and Contribution 4. Inserting the parsed value as a plain reminder reaches 99.27% on Qwen2.5-3B and 98.44% on Mistral-7B, above routing, and the largest model tested, Qwen2.5-14B, gains only 1.17 points (Table 2).
  4. The labels behind the mechanism analysis are ambiguous. In the near-old half of the preference core, a turn says "I am still thinking about [old value] when I compare options for my [slot]", which a reasonable reader could take as a renewed preference. Rewording it as explicitly historical raises Qwen's accuracy on these dialogues from 12.3% to 62.5% (Table 14). Yet the probes, attention analysis, Figure 4 factorial and routing all use the original wording (Appendix F states they were not re-estimated), and the routing parser hard-codes this phrase as a non-update (Appendix E.1).
  5. The frontier-model results measure a different failure. The 2,048-operation logs are permutation composition, a known long-chain state-tracking problem. Figure 10 shows Opus 4.8 losing track of the slot at operation 514, a reasoning error rather than selection of a stale value. That every error names an old occupant of the queried slot is likely uninformative: after thousands of cycles over 16 slots, most identifiers have probably occupied every slot, so a random wrong answer would usually qualify. The deferred-log comparison (12/17 vs. 16/17, p = 0.22) is not significant and covers only 17 of 40 scenarios. These results still carry the "even frontier models fail" framing in the abstract, introduction and conclusion.
Questions for authors
  1. Beyond PI-LLM, what is the new behavioral finding?
  2. In the operational logs, how many of the 16 identifiers have ever occupied the queried slot? What is the chance that a random wrong answer counts as an "old occupant"?
  3. Can you rerun the probes, attention comparison and routing on the clarified dialogues (historical wording or matched placement)?
  4. Is there a routing variant that does not need the parser's current value, for example one that uses the model's own estimate?
Reviewer_Jc7sOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates a failure called stale binding in which LLMs return old values for a variable despite newer updates being available in the context. The authors introduce CICM, a benchmark covering preference updates, changing decision constraints, and operational logs, with programmatic labels to distinguish stale binding errors from other types of mistakes. Their experiments show that stale binding errors become more likely as variables are updated repeatedly. Through attention analysis and component-replacement experiments, the authors propose attention drift as a potential mechanism underlying these failures. Finally, they introduce an inference-time attention intervention that substantially improves accuracy on direct value-retrieval tasks.

Critical strengths
  1. The paper presents a well-structured and comprehensive study of stale binding failures in LLMs, connecting benchmark construction, mechanistic analysis, theoretical modelling, and inference-time intervention. These components form a coherent research pipeline, moving beyond merely identifying a behavioural failure to investigating its potential mechanisms and exploring practical mitigation.
  2. The benchmark’s programmatic annotations and differentiated error categories also enable more fine-grained diagnosis than accuracy-based evaluation alone.
Critical weaknesses
  1. The paper examines prompt wording and placement and shows that these factors can substantially affect stale binding errors. However, it remains unclear to what extent the proposed attention-drift mechanism generalizes across different prompt formulations and update representations. Further analysis disentangling prompt-induced effects from more general value-selection behavior would strengthen the mechanistic claims.
  2. The proposed attention-routing intervention relies on a structured parser to identify the current value and outdated mentions, which may limit its applicability to real-world scenarios where updates are implicit, ambiguous, or expressed in diverse ways. It is also unclear how much of the observed improvement comes from attention routing itself versus the parser’s prior identification of the correct value. Evaluating the method under less structured settings and comparing it with simpler state-tracking or prompting baselines would help clarify its practical utility and generalizability.
Questions for authors
  1. To what extent does the proposed attention-drift mechanism generalize across different prompt formulations and update representations?
  2. How much of the improvement from attention routing can be attributed to the attention intervention itself, as opposed to the parser’s prior identification of the correct current value?
  3. The intervention relies on a structured parser to distinguish current values from outdated mentions. How robust would it be in real-world settings?
Reviewer_fv77OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a benchmark named CICM for evaluating a specific ability (key-value updating) in LLM agents, which allows for automatically construction from exsiting dataset (like PrefEval) and tracking the update failure. The paper also provides a mechanism analysis and identifies the potential reason for update failure (attention drift), with a proof-of-concept intervention method to mitigate update failures.

Critical strengths

This paper proposes a benchmark construction framework that can automatically generate value-update samples, allowing identifying the update failure, which is a potentially useful tool to get a large-scale benchmark targeted at key-value updates in agent-memory research.

Critical weaknesses
  1. Limited evaluation against existing agent-memory methods. As a paper introducing a benchmark for agent-memory management, this work does not include comparisons with advanced agent-memory methods in its evaluation. The paper aims to explore a specific problem (key-value updates) within a specific research domain (agent memory). However, its evaluation primarily relies on directly prompting LLMs with predefined templates, without incorporating dedicated memory-management mechanisms. Moreover, experiments with more advanced proprietary LLMs, such as GPT and Claude, suggest that some relatively simple key-value update tasks are already easy for these models. This raises the question of whether state-of-the-art agent-memory methods or LLM agents equipped with external tools could already address the problem effectively. Without such comparisons, it is difficult to assess the practical relevance and significance of the proposed benchmark.

  2. Limited novelty and depth of the mechanism analysis and intervention. The proposed mechanism analysis and intervention method appear relatively superficial and lack substantial new insights. The mechanism analysis largely revisits an existing phenomenon (attention drift) in Transformer architectures, while the theoretical section mainly formalizes an intuitive observation: introducing more outdated values increases the amount of competing information, making it more difficult to select the current value. As for the intervention method, it is more of a proof of concept than a practically deployable solution for real-world LLM systems. Specifically, it relies on a rule-based parser to identify the positions of old and updated values before performing attention intervention, even though the parser itself can already directly provide the correct answer. This substantially limits the practical significance of the proposed intervention.

Questions for authors

Given that the rule-based parser already identifies the correct updated value, what additional practical benefit does the proposed attention intervention provide over directly using the parser's output?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.