acceptodds
Withdrawn from ICLR 2027

ALERT: Attention-based Long-context Early-warning via Routing Trajectories

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Long conversations can retain relevant evidence while a language model answers incorrectly. We introduce ALERT, which predicts retrieval failure before the first answer token. A learned ranker selects one history message using attention at the final prompt position; a second probe predicts failure from that source's attention profile. The LLM remains frozen, and warnings trigger re-injection of selected evidence while preserving the full history. Paired evaluations through 32K separate localization, failure prediction, and recovery across NoLiMa-style tasks, BABILong, turn-based RULER, and Scatter-MCQ. In an initial confirmation on 576 new BABILong stories at 16K, frozen monitors achieve failure AUROC 0.710 for Gemma-4-E4B and 0.759 for Qwen3-4B, versus 0.518 and 0.517 for raw next-token margin. For these two models, respectively, frozen-monitor AUROCs are 0.941 and 0.802 on RULER, and 0.672 and 0.662 on Scatter-MCQ. Selective re-injection improves accuracy by 14.9 and 17.0 percentage points on BABILong, and 9.5 and 5.7 points on Scatter-MCQ, respectively. These results support pre-generation failure prediction and net recovery, although neither BABILong policy exceeds always-re-injection accuracy. Source-selection errors and frequent warnings remain constraints on effective selective intervention.

Primary area:
interpretability and explainable AI
Keywords:
Long-context language models, in-context retrieval, retrieval failure detection, pre-generation monitoring, attention mechanisms, mechanistic interpretability.

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_6vVWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

ALERT is a white-box monitor that predicts long-context retrieval failures before the first answer token is generated. A supervised linear ranker uses attention at the final prompt position to select one earlier message. A second probe then predicts answer failure from that message's attention strength, its distribution across layers, and its competition with the question and the rest of the history. Both probes are fitted separately for each LLM and benchmark, with the LLM kept frozen. A threshold calibrated for at least 95% failure recall determines whether to repeat an excerpt of at most 64 tokens immediately before the unchanged question.

The evaluation tests ten open-weight models on NoLiMa, RULER, BABILong, and Scatter-MCQ adapted into multi-turn histories of 8K, 16K, and 32K tokens. Detection works best on targeted retrieval, with pooled AUROC of 0.970 on NoLiMa and 0.955 on RULER for Gemma-4-E4B in Table 1. Results are weaker on BABILong and Scatter-MCQ. For example, Qwen3-4B reaches BABILong AUROC of 0.786, 0.805, and 0.831 at 8K, 16K, and 32K in Table 2. Selective re-injection brings sizeable accuracy gains in Table 3, raising aggregate BABILong accuracy by 14.8–18.5 points and Scatter-MCQ accuracy by 5.7–9.5 points for the three intervention models, although it is triggered on 71.3–98.2% of prompts on those tasks.

Critical strengths
  1. Distinguishing source localization from failure prediction and intervention makes the results much easier to interpret. Finding the supporting message does not necessarily mean the model will answer correctly. In the motivating example in Figure 4, Gemma-4-E4B gives that message 15.7% of its layer-35 attention at 16K but still produces the wrong location. Reporting source recall@1 alongside AUROC and false-positive rate in Tables 1 and 2, then accuracy changes and intervention rates in Table 3, helps distinguish errors in finding the source, assessing answer risk, and acting on the warning.

  2. The fitting protocol takes several plausible sources of leakage seriously. Training the failure detector on out-of-fold ranked sources, as described in Section 4, exposes it to realistic localization errors rather than only ideal source selections. The alert threshold is chosen on a disjoint calibration set and then frozen. Section 5 further describes benchmark-specific partitions that hold out facts, relation families, background contexts, or question identities. These are appropriate safeguards for a supervised monitor, and the paper explains them clearly.

  3. Several detection results are strong across the range of ten models and four tasks, and the intervention offers a simple way to recover some failures. Among the Gemma models, pooled AUROC in Table 1 ranges from 0.939–0.970 on NoLiMa and 0.941–0.955 on RULER. Qwen3-4B also remains relatively stable on BABILong as context length increases in Table 2. The sizeable paired BABILong accuracy gains for Gemma-4-E4B, Qwen3-4B, and Llama-3.2-11B in Table 3 are encouraging, particularly because re-injection preserves the original history and repeats only one short excerpt.

Critical weaknesses
  1. The results do not yet isolate how much predictive value comes from routing to the selected source rather than from simpler confidence or attention features. The direct test of the central claim about information beyond next-token confidence uses only one 576-story BABILong cohort at 16K. ALERT reaches AUROC 0.710 for Gemma-4-E4B and 0.759 for Qwen3-4B there, compared with 0.518 and 0.517 for raw next-token margin. That is encouraging, but it covers only one benchmark, one context length, and one confidence feature. Tables 1 and 2 do not compare confidence-only, confidence-plus-routing, or source-agnostic attention baselines on the same splits, so the strong detection scores alone cannot establish the specific contribution of source routing.

  2. The warning policy is close to always re-injecting in the settings where the intervention helps most, leaving its selective advantage unproven. Intervention rates in Table 3 are 71.3–91.5% on BABILong and 97.5–98.2% on Scatter-MCQ. The Introduction gives 43.6% accuracy for selective re-injection with Gemma-4-E4B on the 576-story cohort, compared with 44.4% for always re-injecting, and the abstract states that neither BABILong policy exceeds always re-injection accuracy. Although Section 4 describes always-on, confidence-triggered, lexical, random, and gold-source controls, their results are not shown in the main paper.

  3. The broad detection study should not be taken as evidence that one fitted monitor is broadly reusable, or that intervention works across the same range of models. Detection is evaluated for ten models, whereas Table 3 tests re-injection for only three. Moreover, each ranker and detector is fitted separately for every model and benchmark, and the paper does not test transfer across models or tasks. Some model and task combinations in Tables 1 and 2 also have undefined policies or very high false-positive rates at the required recall. These limitations matter when judging how widely the reported gains would apply.

Questions for authors
  1. Could you compare next-token confidence features, source-agnostic attention statistics, ALERT routing features, and confidence combined with routing across all four benchmarks? Please use the same training, calibration, and test partitions for every variant.

  2. Could you report the always-on, confidence-triggered, lexical, length-matched random, and gold-source re-injection controls described in Section 4, together with rescue and harm counts? I would like to see how much benefit comes from the warning policy rather than from repeating any plausible context.

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

The paper proposes ALERT, a method for predicting long-context retrieval failure before the first answer token is generated. They design a three-(four- if prefill counted) stage pipeline to mitigate retrieval failure. A linear ranker scores each history turn, using the attention, by 14 hand-designed features and return one, then predict whether the answer is a failure.

Critical strengths
  • The evaluation separates source localization (R@1), failure prediction (AUROC/FPR), and recovery (paired rescues and intervention rate). It also documents cases where attention reaches the correct evidence but the answer is still wrong. This framing is useful: it shows that “the model attends to the evidence” and “the model uses it correctly” are different targets.
Critical weaknesses
  • I don’t think the pipeline contributes the benefits as a whole, even based on the evidence in the paper. Essentially, the results show two regimes: (1) On BABILong and Scatter-MCQ, the detector is not helpful (FPR 0.70–1.00, with 71–98% of prompts triggering intervention), but re-injection improves accuracy. (2) On RULER, re-injection brings almost no gain (+0.2 pp) while the detector is accurate (AUROC 0.955 / 0.824). Therefore, the improvement in final accuracy can be attributed mostly to the operation of putting the evidence back, i.e., re-injection, rather than to the detector.

  • Every stage of the pipeline closely follows prior work: QRHead, Lookback Lens and AggTruth, and re-injection is close to the task repetition strategy in Hankache et al. (2025). I do agree that novelty alone does not equal contribution, but this work reads more like an engineering combination of existing components than a new method.

Questions for authors
  • Please report always-on re-injection and gold-source re-injection accuracy alongside selective re-injection for all benchmarks in Table 3. I think this is a very important ablation study. Without these, it is not possible to tell how much of the gain comes from the warning versus the re-injection, or how much is lost to source-selection errors. This also correlates with the weakness above.

  • How does ALERT compare with (a) a linear probe on the final-prompt-position hidden state, and (b) a post-generation detector plus regeneration, at matched compute? As your features are designed, ideally, the performance should be better.

Reviewer_4fvJOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a method named ALERT which uses a series of features extracted from the LLM's attention profile to train a linear probe and predicts the generation failure. After identifying the model failure, the ALERT will re-inject the relevant evidence to improve the answer accuracy. The experiments cover four different benchmarks and ten different LLMs, but the baselines comparisons are missing in experiments.

Critical strengths

1.The paper studies a meaningful problem: detect model failure signals before generation, and propose intervention while the detect risk is high, which is important to a lot of real-world applications

Critical weaknesses

1.Baseline comparisons are missing in Experiments. This paper basically proposes a method to detect generation failure, and the authors do compare their method with other baselines in the Related Work section. However, in the experiment section, the authors only evaluate their method on different benchmarks and model backbones, without the direct experimental comparison with other baselines, which make it less convincing that the proposed method is truly useful. Other evidence such as the selective re-injection doesn't outperform always re-injection on BABILong further supports this claim.

2.The technical contribution and analysis insights are limited. The proposed pre-generation method can be understood as to train a linear probe with 14 different human-selected features on frozen LLMs to predict generation failures, while the linear probe is already a heavily studied method. In addition, the paper doesn't justify why the 14 features (as well as the 9 features in the detector are selected). The analysis part is also not deep enough. For example, the analysis of "longer context makes the answer accuracy drop down" has already been supported by literature like Lost in the Middle. And paper uses only one case to support the claim "Attending to the evidence does not guarantee a correct answer", which is not convincing enough.

Questions for authors

Please refer to the Weakness part.

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

Related papers

Discussion (0)

Sign in to comment.