acceptodds
Under review as a conference paper at ICLR 2027

Beyond Evidence Availability: Interference and Reasoning Diagnostics in Long-Context QA

Abstract

Longer contexts let models access more information, including potentially irrelevant or misleading text. We examine how background content affects question answering across context lengths and reasoning settings. Six models at two reasoning settings answer questions requiring information from several documents, with all annotated supporting passages retained in inputs of about 25,000 or 112,000 tokens. Stage 1 compares background sampled from other questions, related passages filtered to exclude answer mentions, and similar related background with a short passage supporting a designated wrong answer. The passage averages less than 0.2% of the input. We first measure the accuracy loss when sampled background is replaced by related text. We then measure the additional loss under similar background containing the misleading passage. This second loss is larger in nine of 12 model–reasoning combinations, with 95% confidence intervals for the difference above zero. Stage 2 examines cases that the same model answers incorrectly with little or no reasoning but correctly with higher reasoning. Keeping the full input, we provide the lower setting with document numbers from the successful higher-setting answer, numbers of the annotated supporting documents, or the first set plus guidance on using the evidence. On selected cases passing guidance checks with valid responses to all three prompts, the guidance-containing prompt exceeds the prompt using the same document numbers alone by 25.1–39.3 percentage points in five of six models, with positive 95% confidence intervals.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Long Context, Robustness

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.25 over 4 reviews, above 2% of the others.
Reviewer_QEEJOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper explores how random/neutral, relevant but hard and distracting documents affect the performance of LLMs on long-context question answering tasks during inference time. To evaluate this task the authors propose a 2 stage framework. For stage1, they compare the effects of neutral, relevant but hard and misleading documents for a question-answering task across multiple model and context length. For stage2, they compare against high and low reasoning modes of the LLMs and test whether providing hints can help the low reasoning model achieve better performance. Experiments show that the misleading documents have the capability to pull other wrong answers towards the misleading answer.
The paper is well written and easy to follow.

Critical strengths
  1. The paper is clearly written and mostly easy to follow. The motivation is also clear.
  2. The authors experiment with 6 LLMs on the MuSiQue dataset and the 2WikiMHQA dataset. The experiments are mostly well designed.
  3. The findings showing that wrong answers tend to shift towards the misleading document is a strength of this paper.
Critical weaknesses
  1. The biggest weakness is the design of experiments and claims on Stage2 of the experiment. Using direct hints from the high reasoning model and providing them to the low reasoning model is somewhat questionable. There is a chance that this would provide answer specific hints and is a reason why the gains on this setting is substantially higher.
  2. The use of misleading information and distracting prompts/documents have been studied previously. Hence the novelty of the work is limited.
Questions for authors
  1. Is it possible to setup some experiments to show that during Stage2, the low reasoning model is actually better at reasoning?
  2. Can the authors experiment and report the performance of the LLMs when the positioning of the misleading documents are at the start, middle and end?
  3. Would it matter how far the misleading document is from the gold-position document?
  4. Can the authors clearly differentiate their contributions from [1] and [2]?
  5. In the manuscript page 5 - line 225, there is a mention of Stage 3 in the experiment pipeline; can you clarify what is stage 3 here?

References:

  1. Bhuiya, Neeladri, Viktor Schlegel, and Stefan Winkler. "Seemingly Plausible Distractors in Multi-Hop Reasoning: Are Large Language Models Attentive Readers?." Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

  2. Lee, Seongyun, et al. "Lost in the noise: How reasoning models fail with contextual distractors." arXiv preprint arXiv:2601.07226 (2026).

2 comments

Thank you for recognizing the motivation and experimental design.

1. Stage-2 hints and whether low mode becomes better at reasoning.

Stage 2 asks what higher reasoning offers in long-context QA. We use information transfer within the same model as a diagnostic tool. The model parameters, low-reasoning setting, and full context stay fixed. We interpret the gains as more accurate answers with guidance, rather than an increase in the model's underlying reasoning ability.

On cases where high mode succeeds and low mode fails, we compare teacher IDs, annotated gold IDs, and teacher IDs plus a hint. The original successful high-mode response supplies the teacher IDs, indicating where to look. A separate high-mode call uses those fixed IDs to generate guidance on organizing and using the evidence. The generation prompt asks for guidance without stating the final answer, and we reject hints containing the canonical answer or the original high-mode answer after text normalization (Appendix B.2). The example in Section 3.3 illustrates this information transfer.

First, use [D94] to determine which state contains Borough House Plantation. Then, use [D93] to identify the oldest college located in that state.

On matched cases from questions accepted by both auditors, with accepted hints and valid answers to all three prompts, teacher+hint improves accuracy over teacher IDs alone by 25.1–39.3 percentage points in five models (Figure 3). Gains persist when the IDs exactly match the gold-document set (post-hoc analysis, Table 3). These results suggest considering both finding useful evidence and using it to answer when interpreting within-model reasoning gains.

2. Position of the misleading passage.

We found no common position ranking across the tested models. Accuracy point estimates are highest late for both Qwen models and early for DeepSeek, in each thinking setting (Tables R1–R2).

We use all 87 MuSiQue questions accepted by both auditors with gold evidence originally at the front. In approximately 25K-token inputs, we hold the gold prefix fixed and move the misleading passage to boundaries near 10%, 50%, or 90%. Contents, construction-tokenizer length, and other documents' order stay fixed; labels follow the new order.

Both experiments test the two Qwen models and DeepSeek V4.1-Flash-1017. Off disables thinking; High requests Qwen's 4,096-token thinking budget or DeepSeek's max reasoning effort with an 8,192-token total output cap.

Table R1. Accuracy by misleading-passage position.

ModelThinkingnEarly accuracy (%)Middle accuracy (%)Late accuracy (%)
Qwen3.5-27BOff8756.3256.3262.07
Qwen3.5-27BHigh8755.1755.1760.92
Qwen3.5-35B-A3BOff8739.0839.0842.53
Qwen3.5-35B-A3BHigh8647.6745.3550.00
DeepSeek V4.1 FlashOff8750.5744.8345.98
DeepSeek V4.1 FlashHigh7558.6757.3357.33

Table R2. All paired position differences in accuracy, using the cohorts in Table R1.

ModelThinkingMiddle − early (pp) [95% CI]Late − early (pp) [95% CI]Late − middle (pp) [95% CI]
Qwen3.5-27BOff0.00 [-6.90, 6.90]5.75 [-1.15, 13.79]5.75 [1.15, 11.49]
Qwen3.5-27BHigh0.00 [-5.75, 5.75]5.75 [0.00, 11.49]5.75 [0.00, 11.49]
Qwen3.5-35B-A3BOff0.00 [-8.05, 9.20]3.45 [-3.45, 10.34]3.45 [-4.60, 11.49]
Qwen3.5-35B-A3BHigh-2.33 [-10.47, 5.81]2.33 [-2.33, 6.98]4.65 [-2.33, 11.63]
DeepSeek V4.1 FlashOff-5.75 [-10.34, -1.15]-4.60 [-10.34, 0.00]1.15 [-3.45, 5.78]
DeepSeek V4.1 FlashHigh-1.33 [-8.00, 4.00]-1.33 [-8.00, 5.33]0.00 [-6.67, 6.67]

3. Distance between misleading and supporting documents.

We found no consistent pattern of lower accuracy when the misleading passage was closer to the supporting documents. Qwen3.5-27B High matches the designated wrong answer more often in near layouts (Table R3).

For each of the 87 questions, we create four versions of the same context, keeping all text and total length fixed. The supporting documents stay together as one block. Positions below refer to the centers of this block and the misleading passage.

  • Near: supporting documents at about 20% and the misleading passage at 30%, or supporting documents at 80% and the misleading passage at 70%.
  • Far: supporting documents at about 20% and the misleading passage at 70%, or supporting documents at 80% and the misleading passage at 30%.

We average the results of the two layouts in each group. Their mean center-to-center distances are approximately 2.5K tokens for Near and 12.5K for Far. Both groups use the same positions equally often and include the misleading passage both before and after the supporting documents.

Table R3. Near/far accuracy and paired differences. Target match means exact normalized equality with the designated wrong answer.

ModelThinkingnNear accuracy (%)Far accuracy (%)Near − far accuracy (pp) [95% CI]Near − far target match (pp) [95% CI]
Qwen3.5-27BOff8744.2545.40-1.15 [-4.60, 1.74]2.87 [-0.57, 6.32]
Qwen3.5-27BHigh8642.4446.51-4.07 [-8.72, 0.58]4.65 [0.58, 8.72]
Qwen3.5-35B-A3BOff8733.3332.760.57 [-4.02, 5.17]-0.57 [-4.60, 3.45]
Qwen3.5-35B-A3BHigh8736.2133.332.87 [-1.15, 6.90]0.00 [-4.02, 4.02]
DeepSeek V4.1 FlashOff8744.8342.532.30 [-0.57, 5.75]-1.15 [-4.02, 1.16]
DeepSeek V4.1 FlashHigh7055.7155.710.00 [-5.00, 5.00]3.57 [-0.71, 7.86]

Table notes: n counts questions with valid answers at every layout. DeepSeek High has fewer such questions because some runs reached the output limit before answering. Scoring and statistical details are in Appendix I (Tables 42–44).

The supplementary experiments show no common position or distance pattern in accuracy across the tested models, whereas several background-quality and evidence-use findings recur in the main experiments. Together, these results illustrate the value of systematic comparisons for distinguishing recurring patterns from findings that vary across settings.

4. Differences from Bhuiya et al. [1] and Lee et al. [2].

Our aim is to understand how background quality affects answers when all supporting evidence is available, and what higher reasoning contributes on those tasks. Table 1 compares the experimental settings.

Existing findingsWhat our comparisons add
Plausible but incorrect reasoning paths can impair multi-hop QA even when the supporting paragraphs are present. Bhuiya et al. (2024) [1].The accuracy loss from related background to background containing a short misleading passage was larger than the loss from sampled to related background in nine of twelve model–setting combinations (Section 4.1, Figure 2).
Under hard-negative distraction, longer sampled reasoning trajectories were associated with lower accuracy. Lee et al. (2026), Appendix B.7 [2].We probe what higher reasoning offers by passing document selections and evidence-use guidance to the same model in low mode. Even when the document IDs exactly match the correct supporting documents, five models answer more accurately with guidance than with IDs alone (Table 3). These comparisons support interpreting reasoning gains through both finding useful evidence and organizing and using it to answer.

These measurements of input quality and evidence use in long-context QA provide a reference for future research on data construction and training.

5. Meaning of “Stage 3.”

“Stage 3” tests models using only the documents that support the correct answer, without added background passages (Section 4.2, Table 28). We now call it the “gold-only reference” in the revised main text.

References

  1. Bhuiya, Neeladri, Viktor Schlegel, and Stefan Winkler. "Seemingly Plausible Distractors in Multi-Hop Reasoning: Are Large Language Models Attentive Readers?" Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.
  2. Lee, Seongyun, Yongrae Jo, Minju Seo, Moontae Lee, and Minjoon Seo. "Lost in the Noise: How Reasoning Models Fail with Contextual Distractors." arXiv preprint arXiv:2601.07226. 2026.
Reviewer_BUKUOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates how irrelevant and misleading background text affects LLM performance in long-context QA. Experiments demonstrate that inserting a minute misleading passage significantly degrades accuracy across multiple models and context lengths. Furthermore, the authors show that augmenting lower-reasoning models with explicit document references and reasoning guidance effectively mitigates these distractor effects, substantially recovering QA accuracy.

Critical strengths
  1. The paper provides comprehensive evaluations across multiple scoring rules, including EM, F1, and Containment, alongside audit subset sensitivity checks and adversarial missing-value upper bound analyses, rendering its statistical inference both rigorous and credible.

  2. Rather than merely measuring macro accuracy drops, the paper thoroughly examines the fine-grained transitions in model predictions. The authors innovatively decompose errors into shifts from correct answers to targeted errors (C to T) and from other incorrect predictions to targeted errors (O to T).

Critical weaknesses
  1. The paper remains strictly at the level of empirical diagnosis without proposing or evaluating any algorithmic solutions or training strategies for distractor filtering, long-context alignment, or robustness enhancement. Consequently, the findings do little more than quantify known phenomena, offering limited actionable value for architectural designs in long-context models or practical RAG engineering.

  2. The central finding of Stage 2 is that extracting document indices and reasoning guidance from a high-reasoning model that answered correctly, and feeding them to a weaker model, yields substantial performance gains. However, this setup is unfeasible in real-world deployments because it essentially provides pre-selected key evidence and step-by-step solutions to the weaker model, making the observed improvements largely trivial and predictable.

  3. The distractors introduced in Stage 1 are synthetic pieces of counter-evidence generated by an LLM to deliberately target the final multi-hop relation. This setup closely resembles retrieval poisoning or adversarial perturbation rather than the genuine information conflicts, temporal shifts, or organic redundancy that long-context models naturally encounter in practice. As a result, there is a clear distribution gap that limits the generalizability of these findings to realistic applications.

Questions for authors

see Critical Weaknesses

1 comment

Thank you for recognizing the statistical checks and analysis of answer changes.

1. Contribution and practical relevance of an empirical diagnostic study.

Our aim is to identify empirical patterns in how models respond to background information and find and use evidence.

With complete support retained, DeepSeek-high loses 2.3 percentage points from sampled to related background, versus 23.8 from related background to one containing a short misleading passage. These estimates combine 25K/112K results (Figure 2). In Stage 2, high-mode guidance can improve the same model's low-mode answers even when the supplied document IDs exactly match the gold-document set (Table 3).

From this diagnostic perspective, the absence of a new algorithm does not, by itself, negate the scientific value of systematic empirical evidence. Basic questions about information quality and evidence processing require quantitative investigation, even when some outcomes are intuitive. We view building this empirical foundation as a contribution in its own right. As models and training data improve, specific mitigation strategies may become less necessary. Our longer-term objective is to identify which difficulties persist across these changes and how models find and use reliable information. These questions also motivate future investigation in multimodal and robotic systems.

2. Purpose of the high-to-low information transfer.

Stage 2 uses within-model information transfer as a diagnostic tool. On low-wrong/high-correct cases across all three backgrounds, the full context and low-mode setting remain fixed. Original high-mode document IDs indicate where to look; a separate high-mode call supplies guidance on using the evidence. These interventions are not proposed as a deployment method (Section 3.3).

On matched cases from questions accepted by both auditors, with accepted hints and valid answers to all three prompts, teacher+hint improves accuracy over teacher labels by 25.1–39.3 percentage points in five models (Figure 3). Gains persist when teacher IDs exactly match the gold-document set (post-hoc analysis, Table 3).

3. Synthetic distractors and generalization.

We seek generalizable understanding of model behavior through controlled experiments. Comparisons across models, reasoning settings, and context lengths examine whether patterns recur beyond one configuration. The 2Wiki replication also shows accuracy loss and increased adoption of the misleading target in all three tested models (Section 4.3). We leave testing whether these patterns extend to naturally occurring conflicts in retrieved documents to future work.

Reviewer_8U8TOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies long-context multi-hop QA when all annotated supporting documents remain in the input. It compares neutral background, topically related background, and related background containing a short passage supporting a wrong answer. It then examines cases where a model answers incorrectly with low reasoning but correctly with high reasoning. For those cases, it tests whether document labels or a generated hint about using the evidence help the low-reasoning setting. The paper reports that the targeted misleading passage often causes a larger accuracy loss than replacing neutral background with related text, and that labels plus a hint outperform labels alone for five of the six primary models on the selected cases.

Critical strengths

The paper keeps the annotated supporting documents in the input while varying the surrounding context. This is a useful way to examine failures that cannot be explained simply by missing evidence.

Critical weaknesses
  1. It is unclear what new insight the results establish relative to existing work. Prior studies have already shown that hard negative or misleading context can substantially reduce QA accuracy. The paper needs to explain more precisely what we learn from the comparison between related background and a short targeted passage that was not established by earlier distractor studies.
  2. The reasoning result also needs a sharper interpretation. Finding the correct documents and combining their contents to answer a multi-hop question requires reasoning. Therefore, the observation that an evidence-use hint helps beyond document labels is plausible, but it does not by itself establish a new account of why high-reasoning models succeed.
Questions for authors
  1. What is the main new insight beyond prior findings that hard negatives and misleading passages harm QA accuracy?
  2. What does the hint experiment establish about reasoning that cannot be explained as transferring the output of an additional high-reasoning call?
1 comment

Thank you for recognizing the value of retaining supporting documents. Longer inputs give models more information, but its usefulness and reliability vary. This motivates studying both background quality and what higher reasoning contributes when complete supporting evidence is available.

1. What do we learn beyond the finding that distractors harm QA?

Jin et al. (ICLR 2025) [1] identify hard negatives as a contributor to degradation with larger retrieval sets. Gao et al. (First Drop, ICML 2026) [2] show that a small proportion of hard distractors can cause a sharp performance drop at fixed length. Bhuiya et al. (EMNLP 2024) [3] show that plausible incorrect reasoning paths impair multi-hop performance.

Building on these quantitative studies, we compare two background costs within the same long-context tasks, with complete support retained (Table 1). A model can lose little accuracy with extensive related background but lose more when the background contains a short misleading passage. DeepSeek-high loses 2.3 percentage points from sampled to related background, versus 23.8 from related background to one containing a short misleading passage. These estimates combine the 25K and 112K results for questions accepted by both auditors; Figure 2 shows how the relative costs vary across model settings (Section 4.1). This characterizes how far robustness to one kind of background extends to another.

Stage 2 complements these background comparisons by examining within-model reasoning differences through document cues and evidence-use guidance. Gains persist even when the document cues are correct, supporting the examination of both document localization and evidence use.

2. What does Stage 2 establish about reasoning and information transfer?

We agree that the hint transfers problem-solving information from an additional high-mode call. Information transfer is our diagnostic intervention: document IDs from the original successful high-mode response indicate where to look, and hints explain how to organize and use the evidence. Both steps can require reasoning.

On the selected same-model low-wrong/high-correct cases, we retain the full context and low-mode setting (Section 3.3). For questions accepted by both auditors, on matched cases with accepted hints and valid answers to all three prompts, teacher+hint exceeds teacher-label accuracy by 25.1–39.3 percentage points in five models (Figure 3). Gains persist when teacher IDs exactly match the gold-document set, showing room for guidance after the correct documents are specified (post-hoc analysis, Table 3). The hint replaces a citation reminder at fixed IDs in these prompt comparisons (Appendix B.2).

Together, the two stages provide a systematic empirical study of input quality and evidence processing in long-context QA. The measured background costs and responses to evidence interventions provide reference points for future data construction and training research. Intuition can suggest an effect's direction, but its magnitude and consistency across models and conditions require measurement.

References

  1. Jin, Bowen, Jinsung Yoon, Jiawei Han, and Sercan Arik. "Long-Context LLMs Meet RAG: Overcoming Challenges for Long Inputs in RAG." International Conference on Learning Representations. 2025.
  2. Gao, Muhan, Zih-Ching Chen, and Kuan-Hao Huang. "The First Drop of Ink: Nonlinear Impact of Distracting Information in Long-Context Reasoning." Proceedings of the 43rd International Conference on Machine Learning. 2026.
  3. Bhuiya, Neeladri, Viktor Schlegel, and Stefan Winkler. "Seemingly Plausible Distractors in Multi-Hop Reasoning: Are Large Language Models Attentive Readers?" Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.
Reviewer_RoRhOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper investigates how different types of background information affect the question-answering accuracy of LLMs in long-context settings, as well as performance differences across reasoning levels. The study consists of two stages. Stage 1 compares neutral background information, question-related background information, and background information designed to induce specific incorrect answers, while preserving all supporting evidence for the correct answer. Stage 2 examines performance changes by providing models operating under lower reasoning settings with document information and evidence-use guidance generated by the same models under higher reasoning settings. Based on these findings, the authors conclude that not only the availability of supporting evidence but also the quality of background information and how evidence is used play important roles in long-context question answering.

Critical strengths

This study systematically evaluates its hypothesis across various models, reasoning settings, context lengths, and background conditions. In particular, the authors compare performance changes across different types of background information while preserving the supporting evidence, aiming to maintain consistency in the experimental setup. Additionally, the authors conduct further analyses and robustness checks to examine their findings from multiple perspectives. These extensive experiments strengthen the reliability of the results and provide evidence that the observed effects are not limited to a specific model or experimental setting.

Critical weaknesses

This study investigates how relevant or misleading background information can reduce model accuracy, even when all supporting evidence for the correct answer is preserved. However, this phenomenon has already been widely explored in prior research on long-context reasoning, RAG, distractors, and knowledge conflicts. In particular, the finding that passages intentionally designed to support an incorrect answer can lead models toward that answer is somewhat expected. Although the study provides a systematic quantitative analysis of this phenomenon across various experimental settings, it remains unclear what new insights it offers beyond existing research. Therefore, the paper would benefit from a clearer explanation of what it reveals beyond simply measuring the magnitude of accuracy degradation, how it differs from prior work, and why these findings represent a meaningful academic contribution.

Questions for authors

Prior studies on distractors and knowledge conflicts have already explored how misleading information can negatively affect model accuracy. While this study systematically analyzes this phenomenon across various experimental settings, it remains somewhat unclear what new insights it provides beyond existing research. Could the authors clarify the key differences from prior work and explain what they consider to be the most significant novel contributions of this study?

3 comments

Thank you for recognizing the systematic comparisons and reliability checks. Longer contexts give models access to more information, but its usefulness and reliability vary. Research on training-data quality motivates our question about inference-time inputs: how do different backgrounds affect answering when complete supporting evidence is present, and what does higher reasoning offer? Understanding both matters for effective use of long-context models.

1. Relationship to existing findings.

Gao et al. (First Drop) [1] quantify how a small proportion of related, answer-free distractors can sharply reduce accuracy at fixed context length. Bhuiya et al. [2] show that plausible alternative reasoning chains impair multi-hop answering even with the correct supporting paragraphs retained. Lee et al. [3] compare noise types and analyze their effects on reasoning behavior.

We build on this evidence by jointly studying background costs and within-model reasoning differences in long-context QA. Table 1 compares the studies.

2. What our comparisons add.

Stage 1 measures the relative costs of related background and targeted interference. At approximately 25K and 112K tokens, with complete support retained, we compare two losses: sampled-to-related background, and related background versus nearly matched background containing a short passage supporting a designated wrong answer. Combining both lengths for questions accepted by both auditors, the second loss exceeds the first in nine of twelve model–reasoning combinations. For DeepSeek-high, the losses are 2.3 and 23.8 percentage points (Section 4.1, Figure 2). This quantifies how the relative costs vary across model settings.

Stage 2 probes what higher reasoning offers. On cases the same model answers correctly in high mode but incorrectly in low mode, we compare teacher document IDs, annotated gold IDs, and teacher IDs plus guidance, keeping the full context and low-mode setting fixed. Teacher IDs come from successful high-mode responses; separate high-mode calls generate guidance on organizing and using those documents (Section 3.3).

On the matched consensus subset with accepted hints and valid answers to all three prompts, teacher+hint improves accuracy over teacher IDs alone by 25.1–39.3 percentage points in five models (Figure 3). Gains persist when teacher IDs exactly match the gold-document set (post-hoc analysis, Table 3). These diagnostic comparisons support considering both document localization and evidence use when interpreting reasoning gains.

These comparisons provide a systematic empirical reference for studying input quality and evidence processing. Measuring how effects vary across models and conditions gives subsequent research on data construction and training concrete patterns to explain and compare against.

References

  1. Gao, Muhan, Zih-Ching Chen, and Kuan-Hao Huang. "The First Drop of Ink: Nonlinear Impact of Distracting Information in Long-Context Reasoning." Proceedings of the 43rd International Conference on Machine Learning. 2026.
  2. Bhuiya, Neeladri, Viktor Schlegel, and Stefan Winkler. "Seemingly Plausible Distractors in Multi-Hop Reasoning: Are Large Language Models Attentive Readers?" Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.
  3. Lee, Seongyun, Yongrae Jo, Minju Seo, Moontae Lee, and Minjoon Seo. "Lost in the Noise: How Reasoning Models Fail with Contextual Distractors." arXiv preprint arXiv:2601.07226. 2026.

Thank you for the detailed response and clarifications. I appreciate the authors' efforts to better explain the paper's contributions and distinguish them from prior work.

While the response helps clarify the value of the empirical comparisons, my main concern regarding the limited novelty and incremental contribution remains.

Therefore, I will maintain my original rating.

Thank you for considering our response and for clarifying your assessment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.