acceptodds
Under review as a conference paper at ICLR 2027

Can Computation from Earlier Problems Help LLMs Solve New Ones?

Abstract

Large language models often solve independent problems in the same conversation. Can computation from earlier problems help them solve new ones? To answer this question, we first conduct preliminary experiments showing that retained history can raise or lower later-turn accuracy, even within the same domain. To understand these effects, we use controlled replay to isolate internal state changes specific to each problem-history pairing. Across different histories, these changes preserve similar relationships among current problems. To improve reasoning under retained history, we introduce **STAIR** (**Stale-Token Attention for Inter-query Reuse**). STAIR captures keys and values from earlier response generation in a fixed bank. It learns to redirect current queries when they read this bank during prompt processing. The base model remains frozen; only 12,288 parameters are trained. Across three Qwen models and four benchmarks, STAIR improves average later-turn accuracy by up to 11.67 percentage points over the unmodified model with history. Code is provided in the supplementary material.

Primary area:
foundation or frontier models, including LLMs
Keywords:
large language models, multi-turn reasoning, attention state reuse, historical K/V states, parameter-efficient adaptation

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_xVvHOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2(+1)
Read review
Paper summary

This paper studies the setting of question ordering during prompting: their primary finding is that the model’s accuracy (pass@4) score on a question is affected by its position in a context (e.g., does that question come before or after other questions). In their experiments, they consider 4 different question permutations (T1/T2/T3/T4) across three different Qwen models across a variety of language modeling benchmarks (heavily math reasoning). They find that when a question is asked later in the context, the model places a disproportionately large amount of attention on earlier tokens in the context in earlier layers. They then design STAIR: a method to “re-address” how much attention is given to preceding tokens based on a new query; they show that in most settings STAIR raises QA performance compared to Native prompting. STAIR introduces a small set of parameters that need to be trained prior to use.

Critical strengths
  1. I think the experimental setting and metrics used are fair and relevant to the current literature surrounding this work. While STIAR does introduce a small computational overhead when training additional parameters, the rest of the evaluation matches the visible history, current problem, generation seed, and decoding budget (lines 376-377). This makes it easy to compare the performance of STAIR to vanilla and native decoding.

  2. The topic of how to manage the effect of a model’s prior context on later inference is pressing; many current solutions study this problem from the perspective of intervening on the context. This paper takes an interesting approach of intervening on the model activations (which may be relevant in settings where users are free-form interacting with a model outside of a harness which can better manage context).

Critical weaknesses
  1. I am comparing the results in Figure 1D and Figure 2D; from Figure 1D it seems that Vanilla is the best performing QA strategy (in ⅞ settings) and that STAIR basically matches its performance; since error bars are omitted I suspect there is no significant difference in performance between STAIR and vanilla? In which case, why not just clear the context history and use vanilla T1 prompting? I know the authors use Figure 2D to motivate that in some cases, prior steps aid later steps, but if you look at this graph temporally (T2 -> T4), it seems that T4 “wrong→correct” is lower than T2. Overall at each Tx the net change is negative and it gets worse over T2,3,4. It really does just make you wonder if the cheapest and simplest option (resetting the context) is really just the best one.

  2. Tables 1 and 2 are missing error bars; it doesn’t seem like the performance gain from STAIR is substantially better than Vanilla in most settings.

Questions for authors
  • Only a single model family is considered. I am wondering if STAIR (and the finding that the ordering of a question) generalizes to other open-weight model families? (e.g., LLama3)?

  • How does STAIR affect the model performance in a non-QA setting (e.g., longer form text completion)? Would you remove stair for non-QA tasks?

5 comments
AuthorsOpenReview ↗

Response to Reviewer xVvH (Part 1 / 3)

We thank you for your careful review and for recognizing the value of our controlled comparisons and activation-level approach to retained history. Your questions have helped us clarify the role of resetting, statistical evidence, and generalization.

In response, we have strengthened the Vanilla comparisons, added paired uncertainty estimates to Figure 1d and the appendix, extended evaluation to a Llama-based backbone, and evaluated cross-task transfer to code generation. We address your concerns below.

1. Why learn to use history instead of simply resetting it? (W1)

Your question about why one should not simply reset the context is fundamental to the motivation of this work. We agree that resetting is a reasonable and effective choice when previous context is unnecessary. However, our work addresses a complementary setting: when conversational continuity requires retaining history, can the model learn to use the computation produced by earlier turns more effectively?

In real interactions, users frequently ask independent questions, switch topics, pursue follow-up discussions, and later return to earlier tasks within the same conversation. Conversation continuity does not necessarily imply task continuity. Resetting after every task switch discards the preceding interaction, even when users may wish to revisit it. Existing summarization and selection methods manage what history to retain; STAIR instead investigates how current queries can access the historical computation that remains available.

We deliberately evaluate independent, self-contained problems to isolate history effects without direct semantic dependencies between tasks.

Regarding Figure 2d, your observation about the negative net trend is correct. For Qwen3-4B Instruct on MATH-500, the net effects of retained history at T2, T3, and T4 are −0.50, −0.95, and −1.25 percentage points, respectively.

However, the central observation in Figure 2d is that history changes answers in both directions, rather than simply causing uniform degradation. Among 2,000 paired responses at each turn, the numbers changing from wrong to correct are 37, 40, and 35, while those changing from correct to wrong are 47, 59, and 60, respectively.

Thus, positive answer changes persist at every turn, but negative changes increasingly outweigh them under Native as the conversation progresses. This is the motivation for studying selective historical readout: rather than assuming that retained history is uniformly helpful or harmful, we seek to adjust how its computational states influence the current problem.

Importantly, other model–task combinations show a net benefit from retaining history, as detailed in §2 below. Therefore, the negative trend in Figure 2d describes one controlled setting, not a universal property of retained conversations.

Historical computation also offers information beyond ordinary visible context. In thinking models such as Qwen3.5, earlier intermediate reasoning traces are not carried forward as visible text in subsequent turns; the ordinary conversation retains the post-thinking answer text. STAIR separately preserves K/V states produced during those earlier reasoning steps, allowing current queries to access that computation without restoring the intermediate reasoning text to the visible prompt.

This is precisely why we investigate activation-level historical access as a complement to context management, rather than treating Reset and STAIR as competing solutions to an identical problem.

AuthorsOpenReview ↗

Response to Reviewer xVvH (Part 2 / 3)

2. Statistical support and comparisons with Vanilla (W2)

Your emphasis on statistical uncertainty is important for interpreting comparisons with Vanilla. The revised Figure 1d now displays paired pointwise 95% intervals for STAIR relative to both Native and Vanilla. We also report complete per-turn Avg@4, Pass@4, and paired intervals across all three Qwen models and four benchmarks (Tables 28–30).

We estimate uncertainty using 4,000 paired bootstrap replicates, accounting for question and actual-session clustering while preserving matched conditions and each problem's four sampled responses.

The following representative comparisons illustrate both directions of historical influence:

Model / BenchmarkVanillaNativeSTAIRSTAIR − Native (95% CI)STAIR − Vanilla (95% CI)
Qwen3-4B Instruct / MATH-50095.0094.1094.63+0.53 [−0.41, 1.48]−0.37 [−1.48, 0.90]
Qwen3.5-4B / MATH-50096.0094.7896.27+1.48 [0.43, 2.68]+0.27 [−1.06, 1.79]
Qwen3.5-4B / AIME 202560.8340.5652.22+11.67 [3.16, 20.77]−8.61 [−18.72, 1.25]
Qwen3.5-4B / GPQA-Diamond63.6475.2575.38+0.13 [−2.62, 3.06]+11.74 [7.26, 16.84]
Qwen3.5-9B / GPQA-Diamond68.8179.8875.93−3.96 [−6.96, −1.14]+7.11 [3.16, 11.58]

All values are Avg@4 (%). Vanilla reports independent solving at T1; Native and STAIR report mean T2–T4 results.

These results establish two complementary findings.

First, STAIR can mitigate substantial history-induced interference. On AIME 2025/Qwen3.5-4B, Native drops from 60.83% (Vanilla) to 40.56%, while STAIR recovers 11.67 points [3.16, 20.77], approximately 57.5% of the observed loss.

Second, resetting is not universally optimal. On GPQA-Diamond, both Qwen3.5 models perform substantially better with retained history than with Vanilla. STAIR achieves +11.74 [7.26, 16.84] over Vanilla on Qwen3.5-4B and +7.11 [3.16, 11.58] on Qwen3.5-9B, even though the latter trails Native. These positive paired intervals show that clearing history is not always the best-performing choice.

We further strengthened the analysis underlying Table 1 by conditioning on the correctness of the actual first historical response. When T1 is incorrect, Native falls 19.72 points below Vanilla [−25.08, −14.38], while STAIR recovers +8.21 points [4.34, 12.22] over Native. When T1 is correct, STAIR improves over Native by +0.76 points [0.15, 1.40] (Table 33).

Finally, we expanded the fixed-history T4 comparisons across three decoding-seed groups. On Qwen3.5-4B/AIME 2025, STAIR achieves 68.33 ± 2.50% Avg@4, compared with 46.11 ± 3.37% for Native, with positive improvements in every group (Tables 4, 32).

Together, these analyses clarify the central distinction: Vanilla is a strong reference when history can be discarded, while STAIR improves how historical computation is used when conversational history is retained.

AuthorsOpenReview ↗

Response to Reviewer xVvH (Part 3 / 3)

3. Generalization to other model families and non-QA tasks (Questions)

(1) Other open-weight model families and the effect of question position.

Your question about whether both the history effect and STAIR generalize beyond Qwen is particularly important for establishing the scope of our findings. We extended STAIR beyond Qwen by independently training a controller on DeepSeek-R1-Distill-Llama-8B, using its own OpenThoughts response data.

Importantly, this experiment evaluates both the effect of retained history and STAIR's ability to improve historical readout.

Across three decoding-seed groups on AIME 2025:

ConditionMean Avg@4 (%)
Vanilla (T1)34.17 ± 0.83
Native (T2–T4)23.98 ± 0.32
STAIR (T2–T4)28.80 ± 0.58

Retaining history reduces accuracy by 10.19 percentage points relative to Vanilla, while STAIR recovers 4.81 points, with positive gains in every decoding-seed group (Table 24).

Thus, both the history-induced performance change associated with later turns and the benefit of learned historical readout extend to a Llama-based backbone.

We additionally compared SFT and RL checkpoints within the same 1.5B training lineage. At the SFT stage, Vanilla, Native, and STAIR achieve 25.56%, 14.72%, and 20.00%; at the RL stage, they achieve 28.89%, 20.83%, and 27.22%, respectively. These results further show that retained history affects subsequent reasoning across different post-training stages, with positive STAIR gains in all three decoding-seed groups at both stages (Tables 26–27).

(2) Non-QA tasks and long-form generation.

Your question about non-QA generation is important for understanding whether historical readout can benefit tasks beyond benchmark-style reasoning. To evaluate a task beyond conventional question answering, we applied Qwen3.5-9B's original math-trained controller directly to HumanEval+ code generation, without code-domain fine-tuning.

Across all 164 HumanEval+ tasks:

ConditionAvg@4 (%)
Vanilla (T1)87.50
Native (T2–T4)86.18
STAIR (T2–T4)87.91

STAIR improves over Native by +1.73 points [−0.54, 4.23], with positive Avg@4 gains at every later turn (Tables 25, 34).

At the point-estimate level, the unchanged math-trained controller fully recovers the performance loss caused by retaining history and slightly exceeds Vanilla. This provides encouraging evidence that learned historical readout can transfer beyond mathematical reasoning to structured code generation without task-specific fine-tuning.

STAIR itself is not restricted to QA outputs. It adjusts current-prompt representations during prefill, while leaving the backbone and subsequent decoding path unchanged. The same mechanism can therefore be applied to general text generation.

We would not categorically disable STAIR for non-QA tasks. Its usefulness depends on whether the task can benefit from retained historical computation. For open-ended long-form prose, task-specific evaluations of coherence, relevance, and instruction following would help determine when to enable the auxiliary readout. The STAIR branch can also be disabled at inference, recovering the Native computation whenever auxiliary historical access is not desired.

Closing

Thank you again for your thoughtful review. Your comments have helped us clarify an important distinction: resetting history is effective when prior context is unnecessary, whereas STAIR investigates how to improve reasoning when history is retained for conversational continuity.

The additional Vanilla comparisons, paired uncertainty estimates, history-quality analysis, Llama-based evaluation, and HumanEval+ transfer results provide a clearer picture of when historical computation helps, when it interferes, and how learned query-side readout can improve its use.

We hope these revisions provide a stronger basis for evaluating the motivation, reliability, and broader applicability of STAIR.

Reviewer_5yYxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates whether computation from earlier independent problems can help LLMs solve new ones. This paper then introduces STAIR, which learns to read stored historical K/V states during prefill while keeping the backbone frozen, training only 12,288 parameters. Experiments on three Qwen models and four reasoning benchmarks show improvements over the unmodified models with retained history in several settings.

Critical strengths
  1. Interesting research question. The paper investigates whether computation from earlier independent problems can support subsequent reasoning in the same conversation. This is a meaningful question, and the observation that retained history can both help and hurt provides useful motivation.

  2. Compact and parameter-efficient design. STAIR learns how to access a fixed historical K/V bank while keeping the backbone frozen, requiring only 12,288 trainable parameters. The query-reflection and differential-readout mechanism provides a clearly specified approach to adapting historical-state access.

  3. Promising empirical results. STAIR improves reasoning accuracy on AIME 2025 across all three evaluated models, achieving gains of up to 11.67 percentage points in mean later-turn accuracy over the unmodified models with retained history. These results support the effectiveness of the proposed approach. Joint training with LoRA further improves performance over LoRA alone, suggesting that learned historical-state access can complement weight adaptation.

Critical weaknesses
  1. Training dependence and uncertain cross-domain generalization. Although parameter-efficient, STAIR requires backbone-specific controller training rather than being training-free. The math-trained controller shows inconsistent out-of-domain benefits: on GPQA-Diamond, Qwen3.5-4B gains little, while Qwen3.5-9B’s Avg@4 decreases from 79.88% to 75.93%. It remains unclear whether reliable gains in new domains would require domain-specific retraining, which could limit the method’s practical applicability.

  2. Limited evidence at larger model scales. The evaluation is restricted to 4B and 9B models within the Qwen family. It is unclear whether the history-induced degradation motivating this work remains substantial in larger models, and whether STAIR would provide comparable benefits at those scales. Comparing the no-history, retained-history, and STAIR conditions on larger models would strengthen the generality of both the motivating observations and the proposed method.

Questions for authors

See the weakness section.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 5yYx (Part 1 / 2)

We thank you for your constructive review and for recognizing our research question, STAIR's compact design, and its complementarity with LoRA. Your suggestions have helped us sharpen the evaluation of transfer and scalability.

In response, we have extended our evaluation beyond Qwen, investigated different SFT/RL training stages, added cross-domain transfer experiments and repeated decoding evaluations, and provided more complete Vanilla–Native–STAIR comparisons. We address your two concerns below.

1. Backbone-specific training and cross-domain generalization (W1)

Your concern about whether backbone-specific training limits practical cross-domain use is particularly important for assessing STAIR's applicability. STAIR requires backbone-specific training, but its supervision is deliberately lightweight: STAIR requires neither manually annotated historical relevance labels nor a correctness verifier. Training uses saved model-response trajectories—including correct, incorrect, and incomplete responses—with current-response prediction and a Native KL anchor. For the three main Qwen models, a single training epoch updates only 12,288 reflector parameters while keeping the backbone frozen.

Importantly, backbone-specific training does not necessarily imply domain-specific retraining. To investigate this distinction, we evaluate the same math-trained controllers directly on out-of-domain benchmarks without further adaptation.

Model / OOD benchmarkVanillaNativeSTAIRSTAIR − Native (95% CI)
Qwen3.5-4B / GPQA-Diamond63.6475.2575.38+0.13 [−2.62, 3.06]
Qwen3.5-9B / GPQA-Diamond68.8179.8875.93−3.96 [−6.96, −1.14]
Qwen3.5-9B / HumanEval+87.5086.1887.91+1.73 [−0.54, 4.23]

All values are Avg@4 (%). Vanilla denotes independent solving; Native and STAIR report mean T2–T4 performance.

Cross-domain transfer to code generation. We applied Qwen3.5-9B's original math-trained controller directly to all 164 HumanEval+ tasks, without code-domain fine-tuning. STAIR improves mean Avg@4 from 86.18% to 87.91%, with positive gains at every later turn (Tables 25, 34).

Notably, retained history reduces Native performance from Vanilla's 87.50% to 86.18%. At the point-estimate level, the unchanged STAIR controller fully recovers this history-induced loss and slightly exceeds the fresh-context reference (87.91% vs. 87.50%). This provides encouraging evidence that learned historical readout can transfer to a substantially different reasoning domain without task-specific retraining.

Scientific reasoning presents a more nuanced pattern. On GPQA-Diamond, Qwen3.5-4B already benefits substantially from retained history: Native achieves 75.25%, compared with 63.64% for Vanilla, and STAIR maintains this benefit at 75.38%. For Qwen3.5-9B, STAIR decreases from Native's 79.88% to 75.93%, although it remains +7.11 points [3.16, 11.58] above Vanilla.

These results suggest that cross-domain effectiveness depends on the interaction between the learned readout, backbone, and target task. The positive HumanEval+ transfer shows that domain-specific retraining is not always necessary to obtain a benefit, while the GPQA differences motivate broader training mixtures for more consistent transfer.

Furthermore, because STAIR learns from response trajectories without explicit correctness filtering or historical relevance annotations, adapting a lightweight controller to a new backbone or domain does not require constructing a separate labeled relevance dataset.

AuthorsOpenReview ↗

Response to Reviewer 5yYx (Part 2 / 2)

2. History effects and STAIR across model scales (W2)

We appreciate your suggestion to examine larger models and to compare the no-history, retained-history, and STAIR conditions directly.

The revised manuscript now presents Vanilla, Native, and STAIR side by side, and our additional experiments broaden the evaluated model range from 1.5B to 9B, covering Qwen, Llama, and two stages of the same SFT/RL training lineage.

The following table summarizes AIME 2025 results across these checkpoints.

Backbone / training stageEvaluationVanillaNativeSTAIRSTAIR − Native
R1-Distill-Qwen-1.5B (SFT)T225.5614.7220.00+5.28
DeepScaleR-1.5B (RL)T228.8920.8327.22+6.39
Qwen3-4B InstructT2–T447.5034.4438.06+3.61
Qwen3.5-4BT2–T460.8340.5652.22+11.67
R1-Distill-Llama-8BT2–T434.1723.9828.80+4.81
Qwen3.5-9BT2–T461.6743.3349.44+6.11

All values are Avg@4 (%). The 1.5B evaluations use shared-history T2 comparisons; the other models report mean T2–T4 results. Llama-8B and 1.5B results average three decoding-seed groups; corresponding standard deviations and group-level results are provided in Tables 24 and 26–27.

These experiments reveal two important observations.

First, increased model capacity does not automatically eliminate history-induced degradation. Within the Qwen3.5 family, Vanilla AIME accuracy increases slightly from 60.83% at 4B to 61.67% at 9B. Yet Native still suffers a substantial history-induced loss: −20.27 percentage points at 4B and −18.34 points at 9B.

STAIR recovers +11.67 points at 4B and +6.11 points at 9B. Thus, the motivating history-interference phenomenon remains substantial at the largest Qwen scale we evaluate.

Second, the effect extends beyond this single model family and training recipe. DeepSeek-R1-Distill-Llama-8B also shows history-induced degradation, with Native averaging 23.98% against Vanilla's 34.17%; STAIR improves it to 28.80%. Across three decoding-seed groups, the gain is positive in every group.

Similarly, at the 1.5B scale, both the SFT checkpoint and its RL continuation experience degradation under retained history, and both benefit from STAIR. Their Avg@4 improvements average +5.28 and +6.39 points, respectively, with positive gains across all three decoding-seed groups (Tables 26–27).

These results extend the evidence across a sixfold parameter range and several distinct training regimes, while showing that the magnitude of history interference and recovery remains model-dependent.

Regarding models beyond 9B, our available computing resources and the limited rebuttal period constrained the additional controller training and multi-turn, long-reasoning evaluations we could complete at larger scales. We therefore prioritized broadening the evaluation across six checkpoints spanning 1.5B–9B, two model families, and both SFT and RL stages.

These new results provide direct evidence that substantial history effects persist through the 9B scale, while establishing a broader empirical basis for investigating historical computation in larger models.

Closing

Thank you again for your valuable suggestions. Your feedback has encouraged us to distinguish backbone-specific training from domain-specific generalization and to examine how history effects vary across architectures, training stages, and model scales.

Our additional experiments show that a math-trained historical readout can transfer to a different task without domain-specific fine-tuning, while the expanded AIME evaluations demonstrate that history-induced degradation is not confined to a single architecture or model size.

We hope these results and clarifications provide a more complete basis for evaluating both the practical applicability and broader significance of STAIR.

Dear Reviewer,

Thank you again for your time and thoughtful review.

As the discussion period draws to a close, we would like to kindly follow up on our responses and revised manuscript, which includes additional experiments and clarifications addressing your concerns. We would greatly appreciate any further feedback and would be happy to address any remaining questions.

Thank you again for your time and consideration.

Best regards, The Authors

Dear Reviewer,

Just a brief follow-up as the discussion period is nearing its end. We would greatly appreciate it if you could take a moment to review our responses and updated manuscript when convenient.

We would be happy to address any remaining questions or concerns.

Thank you again for your time and consideration!

Best regards, The Authors

Reviewer_Q4v6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies continuous multi-problem sessions, where an LLM solves several independent problems in the same conversation, and investigates whether computation from earlier problems can help solve new ones. The authors first show that retaining history can improve or degrade performance depending on the model and task. Through a controlled replay analysis, they identify a recurring problem-dependent component in history-conditioned representations. Finally, the authors introduce STAIR, a lightweight method to improve reasoning under retained history. STAIR stores historical attention keys and values in a separate bank and learns to redirect current queries using Householder reflections. The backbone remains frozen, with only 12,288 parameters trained. The authors report improvements of up to 11.67 percentage points over the native model across three Qwen models and four benchmarks.

Critical strengths
  • The submission studies a practical and relatively under-explored setting, i.e., solving independent problems within the same conversation, and demonstrates that retaining history can significantly affect model performance.

  • Introduced a lightweight STAIR method that learns to reuse historical attention states while keeping the backbone frozen, requiring only 12,288 trainable parameters.

  • Well furnished details in appendices for reproducibility, along with several useful ablations to understand the role of learned reflections and historical K–V states.

Critical weaknesses
  • It is not clear how much the findings generalize beyond the Qwen family. The effect of retained history likely depends on the pre-training and post-training recipe of the model (e.g., sequence packing, multi-turn SFT, reasoning RL). Further, Qwen3.5 uses a hybrid architecture with Gated DeltaNet layers, while the representation analysis in Section 2 is conducted only on Qwen3-4B Instruct. Given that the largest gains are reported on Qwen3.5, it is unclear whether the analysis sufficiently explains the improvements observed on these models.

  • It is not clear how much of STAIR's improvement comes from reusing historical computation versus simply modifying the current model's internal representations. In Table 3, a random reflector achieves +15.56 points compared to STAIR's +26.67, while a separately trained Bank-free controller achieves +15.00 without reading the historical K–V bank. Although the K–V permutation results suggest that historical content contributes, a substantial fraction of the gain can be achieved without learned historical addressing.

  • The experimental results and claims could be better supported. On AIME 2025, where the largest improvements are reported, STAIR still performs below Vanilla (without history) for all three models. STAIR also degrades Qwen3.5-9B performance on GPQA-Diamond compared to native (79.88 to 75.93). Further, evaluations on relatively small benchmarks (30 AIME and 34 AMC23 problems), without confidence intervals or multiple training seeds, make it difficult to assess the reliability of smaller gains.

Questions for authors
  • Have the authors considered evaluating the effect of history and STAIR across different training stages (e.g., base, SFT, RL checkpoints) or a model family other than Qwen? This would help understand whether the observed behavior is a general property of LLMs or specific to certain training recipes.
5 comments
AuthorsOpenReview ↗

Response to Reviewer Q4v6 (Part 1 / 3)

We thank you for your thoughtful review and for recognizing the importance of the continuous-session setting, STAIR's lightweight design, and our reproducibility efforts. Your questions have guided the additional analyses below.

In response, we have extended our representation analysis to the hybrid Qwen3.5 architecture, evaluated STAIR beyond the Qwen family and across SFT/RL training stages, strengthened our matched mechanism controls, and added three decoding-seed groups to several experiments alongside paired uncertainty estimates. We address your concerns below.

1. Generalization beyond Qwen and across training stages (W1)

Your concern that history effects may depend on model architecture and training recipe is important for assessing generalization. We therefore added three complementary analyses.

(1) Representation analysis on hybrid Qwen3.5. We extended the controlled replay experiment in §2 to Qwen3.5-4B, using 8,192 problem–history replays across two disjoint groups. At layer 19, cross-history neighborhood overlap reaches 48.22–50.66%, compared with approximately 12% under identity shuffling. The matched overlap also exceeds the 97.5th percentile of the shuffle control across all 32 layers (Figure 2, Appendix A.4).

Importantly, the three full-attention layers controlled by STAIR, [3, 11, 19], exhibit 48.22–57.70% matched neighborhood overlap versus 11.51–12.64% under shuffling (Table 7). These layers provide direct, content-addressable access to historical K/V states, while the Gated DeltaNet layers retain their original computation. Thus, STAIR intervenes precisely at the historical readout interface where our diagnostics reveal recurring problem-dependent structure. On AIME 2025, Qwen3.5-4B also exhibits substantial history-induced degradation (60.83% Vanilla versus 40.56% Native), making selective historical re-addressing particularly relevant.

(2) Evaluation beyond the Qwen family. We independently trained STAIR on DeepSeek-R1-Distill-Llama-8B, using its own OpenThoughts response data. Across three decoding-seed groups on AIME 2025, STAIR raises mean T2–T4 Avg@4 from 23.98 ± 0.32% to 28.80 ± 0.58%, an average improvement of +4.81 points, with positive gains of +4.17 to +5.28 points across the groups (Table 24).

(3) Generalization across SFT and RL stages. We further compared DeepSeek-R1-Distill-Qwen-1.5B (SFT) with its RL continuation, DeepScaleR-1.5B, under shared T1 histories. Both use stage-specific controllers trained under matched supervision and update budgets.

Training stageNative Avg@4 (%)STAIR Avg@4 (%)Gain (pp)
SFT14.72 ± 2.6820.00 ± 1.67+5.28
RL20.83 ± 1.4427.22 ± 1.73+6.39

These are means and sample standard deviations across three decoding-seed groups. STAIR improves over Native in every group at both stages (Tables 26–27).

We also tested cross-task transfer by applying Qwen3.5-9B's math-trained controller directly to all 164 HumanEval+ tasks without code-domain fine-tuning. Mean T2–T4 Avg@4 improves from 86.18% to 87.91%, a paired difference of +1.73 points [−0.54, 4.23], with positive gains at every later turn (Tables 25, 34).

Together, these experiments extend the evidence beyond a single model family, architecture, training stage, or reasoning domain, while showing how the magnitude of improvement depends on the evaluated setting.

AuthorsOpenReview ↗

Response to Reviewer Q4v6 (Part 2 / 3)

2. Separating historical computation reuse from generic representation modification (W2)

We appreciate your observation that Random-reflector and Bank-free also achieve substantial gains. This is informative because it separates the effect of modifying current computation from the additional contribution of learned historical addressing.

In the Shared-T1 T2 experiment, Native reaches only 21.67% Avg@4, indicating substantial difficulty under retained history. Random reflection is not an inactive control: even without learned directions, it redistributes attention over the historical K/V bank, producing a nonzero differential readout relative to the original query. Bank-free instead learns channel-wise adjustments to the current attention output without accessing the auxiliary bank.

The key question is therefore whether learned, content-dependent historical addressing provides further gains beyond these interventions.

MethodAvg@4 (%)Gain over Native (pp)
Native21.67–
Random-reflector37.22+15.56
Bank-free36.67+15.00
STAIR48.33+26.67

Bank-free matches STAIR's 12,288 trainable parameters, training data, objective, and update budget. Random-reflector retains the same historical bank but replaces learned reflection directions with random ones.

STAIR exceeds Bank-free by +11.67 points [2.50, 20.83] and Random-reflector by +11.11 points [2.78, 19.72], with paired 95% intervals excluding zero (Table 31).

More importantly, we directly test whether this learned advantage depends on historical K–V correspondence:

  • With intact K–V pairs, learned directions outperform random directions by +11.11 points.

  • After permuting historical Values while preserving Keys and positions, the learned-minus-random difference becomes −3.89 points.

  • The resulting interaction is +15.00 points [6.39, 24.44].

  • Training and evaluating a separate controller with permuted K–V pairs further reduces Avg@4 to 31.67%, compared with 48.33% for intact STAIR.

These interventions link the additional advantage of learned reflections to historical key–value correspondence, beyond generic representation adjustment or arbitrary attention redistribution (Tables 3, 14, 31).

Additional matched readout comparison. We also trained a Reference-read gate that accesses the same historical bank using unreflected queries and learnable channel-wise scaling, with matched parameter count, data, objective, and optimization budget.

In the new fixed-history T4 evaluation, repeated across three decoding-seed groups, STAIR achieves 68.33 ± 2.50% Avg@4, compared with 57.22 ± 2.10% for the Reference-read gate. STAIR outperforms the gate in every group and leads all standalone methods in mean Avg@4 and Pass@4 (Tables 4, 32).

Furthermore, STAIR remains complementary to other adaptation mechanisms. On Qwen3-4B Instruct/AIME 2025, jointly training LoRA + STAIR improves Avg@4 from 39.44% (LoRA alone) to 43.33% (Table 5).

Together, these results show that while generic representation modification can itself be beneficial, learning how to address correctly paired historical K/V states provides an additional, experimentally distinguishable contribution.

AuthorsOpenReview ↗

Response to Reviewer Q4v6 (Part 3 / 3)

3. Vanilla comparison, reliability, and uncertainty estimates (W3)

Your emphasis on the Vanilla reference and uncertainty estimates is important for interpreting gains under retained history. We agree that Vanilla is an essential baseline for independent problems. Our question is whether historical computation can be used more effectively when history remains part of a continuous conversation.

We now report complete Vanilla, Native, and STAIR results for all three models and four benchmarks, including per-turn Avg@4, Pass@4, and paired 95% uncertainty intervals (Tables 2, 19–22, 28–30).

Several representative comparisons illustrate both the benefits and limitations:

Model / BenchmarkVanillaNativeSTAIRSTAIR − Native (95% CI)
Qwen3.5-4B / AIME 202560.8340.5652.22+11.67 [3.16, 20.77]
Qwen3.5-4B / MATH-50096.0094.7896.27+1.48 [0.43, 2.68]
Qwen3.5-4B / GPQA-Diamond63.6475.2575.38+0.13 [−2.62, 3.06]
Qwen3.5-9B / GPQA-Diamond68.8179.8875.93−3.96 [−6.96, −1.14]

On AIME 2025, STAIR remains below Vanilla for all three main models. For Qwen3.5-4B, however, it recovers approximately 57.5% of the accuracy loss associated with retaining history, improving over Native by 11.67 points.

Conversely, history can also be beneficial. On GPQA-Diamond/Qwen3.5-4B, STAIR exceeds Vanilla by +11.74 points [7.26, 16.84]. Even where STAIR trails Native on Qwen3.5-9B/GPQA-Diamond, it remains above Vanilla by +7.11 points [3.16, 11.58]. Interestingly, the same math-trained 9B controller also shows positive transfer to HumanEval+ code generation (+1.73 points without fine-tuning; Table 25), suggesting that cross-domain effects depend on the target task rather than following a uniform pattern.

Reliability and repeated evaluation. We have strengthened the evaluation in two ways.

First, we added paired pointwise 95% intervals using 4,000 bootstrap replicates with question and session clustering, preserving the four samples per problem and matched conditions. The larger MATH-500 benchmark (500 problems) also shows a positive STAIR–Native interval for Qwen3.5-4B: +1.48 [0.43, 2.68]. For AMC23, we explicitly exclude six training-overlap problems, leaving 34 evaluation problems.

Second, we expanded the fixed-history T4 comparisons, the non-Qwen Llama evaluation, and the SFT/RL comparisons to three matched decoding-seed groups, reporting group-level results and mean ± standard deviation (Tables 4, 24, 26–27, 32). These evaluations show consistent positive Avg@4 gains over the relevant Native baselines across the groups.

In addition, STAIR controllers have now been trained separately for six model checkpoints: three main Qwen models, DeepSeek-R1-Distill-Llama-8B, and the 1.5B SFT/RL pair. Their respective AIME evaluations all show positive STAIR–Native Avg@4 point estimates. Together with the repeated decoding evaluations, this provides complementary evidence for the applicability of the learning procedure across backbones and post-training stages.

We also provide an analysis conditioned on the correctness of the actual T1 historical answer. When T1 is incorrect, Native falls 19.72 points below Vanilla [−25.08, −14.38], while STAIR recovers +8.21 points [4.34, 12.22] over Native (Table 33). This further supports the importance of adapting historical readout to the content of retained history.

These additional comparisons and uncertainty analyses help distinguish reproducible positive trends from model- and task-dependent outcomes, while making the scope of our findings more transparent.

Closing

Thank you again for your careful review and valuable suggestions. Your feedback has motivated us to extend the analysis across architectures, model families, and training stages, strengthen mechanism attribution through matched controls and K–V interventions, and provide more comprehensive uncertainty estimates.

Our central finding is that historical computation can affect later reasoning in structured, problem-dependent ways, and learning how current queries access retained historical states provides an additional capability beyond generic representation adjustment.

We hope these new experiments and clarifications offer a stronger and more complete basis for evaluating the contribution of STAIR.

Reviewer_Mn9HOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies how earlier, independent problems affect later reasoning and proposes STAIR, which learns query reflections to read a fixed historical K/V bank during prefill. Across three Qwen models and four benchmarks, the authors report gains of up to 11.67 percentage points in mean later-turn accuracy over the unmodified model with retained history, While still lower than remove past inforamation

Critical strengths
  • s1. Controlled analysis of history effects. The experiments hold the checkpoint and sampling settings fixed and report both accuracy gains and losses from retained history. The representation analysis separates average problem and history effects and includes a shuffled-identity control.
  • s2. Adaptation with a frozen backbone. STAIR trains 12,288 controller parameters while keeping the backbone and historical K/V entries fixed. It modifies queries only during prefill and preserves the original decoding procedure.
Critical weaknesses
  • w1. Unclear training procedure and rationale for STAIR. A natural rationale for reusing past computation is that earlier solutions contain information or reasoning patterns relevant to the current problem. However, retaining history can also reduce accuracy. STAIR learns query reflections that modify attention to a fixed historical K/V bank, but the training procedure and the rationale for this modification are unclear. How are training examples and supervision targets constructed, and what loss is used to optimize the controller? In particular, are useful historical states explicitly labeled, or is their relevance learned indirectly from supervision on the current response? The paper should also explain how this objective guides the query reflections and why the resulting attention changes are expected to improve reasoning.

  • w2. The contribution of query reflections needs to be separated from other ways of managing history. The excerpt does not describe comparisons with prompting the model to summarize earlier questions and responses, selecting relevant turns, or applying sparse attention to historical tokens. Such comparisons would test whether STAIR offers benefits beyond compressing or filtering history. Using the same K/V bank with an identity query transformation or a simpler learned gate would help isolate the reflection controller's contribution; trainable alternatives should use the same training data and comparable optimization budgets. Testing STAIR together with sparse attention would also clarify whether query modification remains useful when access to historical tokens is already restricted.

  • w3. Improvement over Native does not establish a benefit over a fresh context. Since each problem is independent and self-contained, solving it without previous conversation history is a directly applicable baseline. The excerpt reports STAIR's gains over Native but does not provide corresponding comparisons with Vanilla. Please report Vanilla, Native, and STAIR accuracy for each model and benchmark, including per-turn results and uncertainty estimates. This would establish whether STAIR exceeds fresh-context accuracy or primarily recovers losses caused by history retention. Report latency, prefill cost, peak memory, and bank storage for the same sequence of problems, accounting for any cache reuse. The parameter count alone does not establish inference efficiency.

Questions for authors

refer to weakness

6 comments
AuthorsOpenReview ↗

Response to Reviewer Mn9H (Part 1 / 4)

We thank you for your careful review and for recognizing our controlled analysis of history effects (s1) and lightweight frozen-backbone design (s2). Your suggestions have guided the clarifications and additional experiments below.

In response, we have clarified the training procedure and gradient flow (§3, Figure 8), added controlled comparisons with alternative history-management and readout methods (Tables 3–5), extended the history-management evaluation across three decoding-seed groups (Tables 4 and 32), provided per-turn results and uncertainty estimates across all models and benchmarks (Tables 19–22, 28–30), and expanded our inference-cost analysis (Tables 16–17). We address your concerns below.

1. Motivation, design rationale, and training procedure (W1)

Your questions about how Query Reflection is supervised and why differential readout may help are central to understanding STAIR. We have therefore clarified both its training procedure and design rationale.

Why learn to read history rather than simply reset it?

In real conversations, users often ask different questions within the same session, switch topics, and later return to earlier discussions. Conversation continuity does not necessarily imply task continuity. Resetting is reasonable when history is unnecessary, but it may also discard information users wish to retain. Meanwhile, retained history can either help or interfere with subsequent reasoning. Existing summarization and selection methods primarily address what history to retain; we investigate a complementary question: when history is retained, can models learn how to access its previous computation more effectively?

We use independent, self-contained problems to isolate historical effects without direct cross-task semantic dependencies. Our controlled replay (§2) reveals recurring problem-dependent structure in history-induced representation changes: cross-history neighborhood overlap reaches 46.50–48.06% for Qwen3-4B Instruct, compared with approximately 12% under identity shuffling. Similar patterns appear in Qwen3.5-4B. These findings motivate learning query-conditioned historical readout rather than assuming that all historical information is beneficial.

Why differential readout? STAIR keeps the backbone and historical K/V bank fixed, learning only norm-preserving Householder reflections that redirect current queries toward different historical keys. The original and reflected queries read the same bank, and their difference is injected into the current computation. Importantly, the two reads differ only in their attention weights, whose difference sums to zero. Thus, differential readout isolates the change induced by re-addressing, rather than injecting another full historical read; shared components cancel, while the learning objective determines which redistributions help predict the current response.

How is the controller trained? For each Qwen backbone, we reorganize its own DAPO-Math responses into four-turn sessions of distinct problems. T1 supplies the initial history; T2–T4 use saved responses under teacher forcing, supervising both reasoning and final-answer tokens. Correct, incorrect, and incomplete responses are retained without correctness-based filtering. No historical states are explicitly labeled as useful or useless.

The objective is:

L=CE⁡(y,pSTAIR)+DKL(pNative∥pSTAIR).\mathcal L = \operatorname{CE}(y,p_{\mathrm{STAIR}}) + D_{\mathrm{KL}}(p_{\mathrm{Native}}\Vert p_{\mathrm{STAIR}}).

The Native teacher receives the same visible history, current question, and response prefix, with auxiliary readout disabled. CE learns historical readout through current-response prediction, while KL anchors the output distribution to Native. Although reflection is applied only during prefill, response losses backpropagate through the frozen backbone's current-prefix computation graph to the 12,288 reflector parameters. Previously captured historical bank entries remain detached.

We have clarified this training and gradient-flow process in revised §3.3, Appendices B.4–B.5, and Figure 8.

AuthorsOpenReview ↗

Response to Reviewer Mn9H (Part 2 / 4)

2. Isolating the contribution of Query Reflection (W2)

Your suggested comparisons with history-management and simpler readout methods are particularly useful for isolating the contribution of learned Query Reflection. Following your suggestion, we added fixed-history T4 comparisons on Qwen3.5-4B/AIME 2025, now repeated across three decoding-seed groups. All ten conditions share the same 30 problems, original Native T1–T3 histories, and matched generation settings, with four samples per problem in each group. Table 4 reports the mean and sample standard deviation across groups.

MethodAvg@4 (%)Pass@4 (%)
Native46.11 ± 3.3776.67 ± 5.77
Summarization67.50 ± 4.4178.89 ± 1.92
Turn selection36.39 ± 1.7365.56 ± 5.09
Sparse-25%49.72 ± 3.4778.89 ± 1.92
Sparse-50%51.11 ± 3.9476.67 ± 6.67
Reference-read gate57.22 ± 2.1081.11 ± 5.09
STAIR68.33 ± 2.5082.22 ± 6.94
STAIR + Summarization68.61 ± 3.1580.00 ± 0.00
STAIR + Sparse-25%70.28 ± 2.5581.11 ± 1.92
STAIR + Sparse-50%69.72 ± 2.9382.22 ± 5.09

Matched readout controls. The identity transformation reduces exactly to Native because the two reads cancel. We additionally train a 12,288-parameter Reference-read gate using the same bank, unreflected queries, and learnable channel-wise scaling, matching STAIR's training data, objective, and update budget.

Across the three decoding-seed groups, STAIR leads all standalone methods in both mean Avg@4 and Pass@4. It exceeds the Reference-read gate by +11.11 Avg@4 points on average and outperforms it in every group (Table 32). Compared with summarization, STAIR has a +0.83-point higher mean Avg@4, although the ordering varies across groups. These comparisons use fixed trained controllers and histories, so the repeated groups measure decoding variability under matched contexts.

Complementarity with history management. STAIR + Sparse-25% improves over Sparse-25% by +20.56 Avg@4 points on average, while STAIR + Sparse-50% improves over Sparse-50% by +18.61 points. Both improvements are positive in all three decoding-seed groups, and both combinations also increase mean Pass@4. In the original T4 evaluation group, Native and STAIR produce nearly identical response-token counts (53,498 vs. 53,402; Table 18). Thus, the observed improvement is not simply explained by generating longer responses.

Additional controls. In the separate Shared-T1 T2 experiment, STAIR achieves 48.33% Avg@4 versus 36.67% for a parameter- and training-matched Bank-free controller: +11.67 [2.50, 20.83]. Learned reflections also outperform random directions by +11.11 [2.78, 19.72] points with intact K–V pairing, but by −3.89 points after permuting historical values, yielding an interaction of +15.00 [6.39, 24.44]. Training and testing with permuted K–V pairs further reduces accuracy to 31.67% (Tables 3, 14, 31).

We also added two complementary ablations. First, restricting the readable bank to its first 2K, 8K, or 32K entries yields 35.00%, 43.33%, and 39.17% Avg@4, respectively, versus 48.33% with the full bank. Second, on Qwen3-4B Instruct/AIME 2025, joint LoRA + STAIR achieves 43.33% Avg@4, improving over LoRA alone (39.44%) by another 3.89 points. LoRA uses 336 times as many trainable parameters as STAIR, while the data, objective, and update count are matched (Tables 5 and 15).

Together, these results suggest that learned historical readout is distinct from simple gating or generic parameter adaptation, and can complement both history compression and weight adaptation.

AuthorsOpenReview ↗

Response to Reviewer Mn9H (Part 3 / 4)

3. Vanilla comparison and the effects of retained history (W3)

We fully agree that Vanilla is an essential reference for independent problems. We do not claim that retaining history universally outperforms resetting it. Rather, we study whether models can make better use of historical computation when history remains part of a continuous conversation.

The following table reports Vanilla, Native, and STAIR Avg@4, together with the paired STAIR–Native differences and 95% intervals.

ModelBenchmarkVanillaNativeSTAIRSTAIR − Native (95% CI)
Qwen3-4B InstructMATH-50095.0094.1094.63+0.53 [−0.41, 1.48]
AIME 202547.5034.4438.06+3.61 [−2.93, 12.65]
AMC23†92.6593.3893.38+0.00 [−3.19, 3.03]
GPQA-Diamond58.5958.0058.75+0.76 [−2.53, 4.13]
Qwen3.5-4BMATH-50096.0094.7896.27+1.48 [0.43, 2.68]
AIME 202560.8340.5652.22+11.67 [3.16, 20.77]
AMC23†76.4768.1474.51+6.37 [−2.18, 15.89]
GPQA-Diamond63.6475.2575.38+0.13 [−2.62, 3.06]
Qwen3.5-9BMATH-50096.6596.3896.78+0.40 [−0.66, 1.46]
AIME 202561.6743.3349.44+6.11 [−4.16, 16.46]
AMC23†77.9472.5575.74+3.19 [−4.91, 10.69]
GPQA-Diamond68.8179.8875.93−3.96 [−6.96, −1.14]

† AMC23 excludes problems overlapping with the training set, leaving 34 problems.

Vanilla reports independent solving at T1; Native and STAIR report mean T2–T4 results. Complete per-turn Avg@4, Pass@4, and paired uncertainty estimates are reported in Tables 19–22 and 28–30.

Three findings are particularly informative:

  • History can be harmful. On AIME 2025/Qwen3.5-4B, Native drops from 60.83% (Vanilla) to 40.56%. STAIR recovers 11.67 points [3.16, 20.77], approximately 57.5% of the observed loss, although it remains below Vanilla.

  • History can be beneficial. On GPQA-Diamond/Qwen3.5-4B, STAIR reaches 75.38% versus Vanilla's 63.64%, with a paired STAIR–Vanilla difference of +11.74 [7.26, 16.84]. On MATH-500, STAIR also improves over Native by +1.48 [0.43, 2.68], reaching 96.27% versus Vanilla's 96.00%.

  • The effect is not universal. On GPQA-Diamond/Qwen3.5-9B, STAIR trails Native by 3.96 points [−6.96, −1.14], but still exceeds Vanilla by +7.11 [3.16, 11.58].

An additional analysis stratifies sessions by the actual correctness of their T1 historical response, pooled over three Qwen models and three mathematical benchmarks. When T1 is incorrect, Native falls 19.72 points below Vanilla [−25.08, −14.38], while STAIR recovers +8.21 points [4.34, 12.22] over Native. When T1 is correct, STAIR still improves over Native by +0.76 points [0.15, 1.40] (Table 33). This further illustrates that the impact of retained history depends on its content.

These results support our intended positioning: resetting is reasonable when history is unnecessary, whereas STAIR investigates how to mitigate interference while retaining potentially useful historical computation.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We sincerely thank all four reviewers for your thoughtful and constructive feedback. We especially appreciate your recognition of the practical continuous-session setting (Q4v6, 5yYx, xVvH), our controlled analysis and evaluation (Mn9H, xVvH), and STAIR's lightweight frozen-backbone design (Mn9H, Q4v6, 5yYx).

Your suggestions have helped us clarify the motivation and mechanism and substantially expand our evaluation. Below we summarize the central contribution and major revisions; our individual responses address each concern in detail. Major revisions in the updated manuscript are highlighted in blue.

Our central research question

In real conversations, users frequently switch between independent tasks while retaining earlier interactions. Conversation continuity does not necessarily imply task continuity. Resetting the context is a reasonable choice when history is unnecessary, but it does not address how to make effective use of historical computation when the conversation must remain continuous.

Our work investigates this complementary problem. Through controlled replay, we identify recurring problem-dependent structure in history-conditioned representations. STAIR then learns how current queries access a fixed bank of historical K/V states, without modifying the backbone or stored historical computation.

Major improvements in the revised manuscript

1. Clearer method motivation and training procedure. We have strengthened the connection between our representation analysis and Query Reflection, including additional diagnostics on the hybrid Qwen3.5 architecture. We also clarified the training-data construction, supervision targets, CE + Native KL objective, and gradient flow from response prediction to the reflector parameters. The revised §3 and new Figure 8 explicitly illustrate the training process and historical bank accumulation.

2. Stronger mechanism isolation and comparisons with alternative approaches. We added controlled comparisons with history summarization, turn selection, sparse access, and a parameter-matched Reference-read gate (Table 4). Across three decoding-seed groups, STAIR achieves 68.33 ± 2.50% Avg@4, compared with 57.22 ± 2.10% for the Reference-read gate, outperforming it in every group. Additional matched controls show that STAIR improves over Bank-free by +11.67 points [2.50, 20.83], while K–V pairing interventions yield a +15.00-point interaction [6.39, 24.44] (Tables 3, 14, 31). We also demonstrate that STAIR can complement sparse historical access and LoRA-based adaptation.

3. More complete Vanilla comparisons and statistical evidence. We now provide Vanilla, Native, and STAIR results across all three main models and four benchmarks, with per-turn Avg@4, Pass@4, and paired 95% uncertainty intervals (Tables 2, 19–22, 28–30). For example, on Qwen3.5-4B/AIME 2025, Native falls from 60.83% (Vanilla) to 40.56%, while STAIR recovers +11.67 points [3.16, 20.77]. Conversely, on GPQA-Diamond, STAIR exceeds Vanilla by +11.74 points [7.26, 16.84], illustrating that retained history can also be beneficial. We further analyze how the correctness of earlier responses affects later performance (Table 33).

4. Broader generalization and practical evaluation. We extended our evaluation to six separately trained model checkpoints spanning 1.5B–9B, including a Llama-based backbone and both SFT and RL stages. On DeepSeek-R1-Distill-Llama-8B, STAIR improves Avg@4 from 23.98% to 28.80% across three decoding-seed groups (Table 24), and improves AIME 2025 Avg@4 by +5.28 and +6.39 points across matched 1.5B SFT and RL stages, respectively (Tables 26–27). We also demonstrate cross-task transfer: the unchanged math-trained Qwen3.5-9B controller improves HumanEval+ code-generation Avg@4 from 86.18% to 87.91% without domain-specific fine-tuning (Table 25). Finally, we added detailed measurements of end-to-end inference time, prefill overhead, peak memory, and historical bank storage (Tables 16–17).

Closing remarks

Taken together, these revisions provide a more complete understanding of when retained history helps or interferes with reasoning, what learned historical readout contributes beyond simpler interventions, and how the approach behaves across different models and tasks.

Our central contribution is not a claim that retaining history always outperforms resetting it. Rather, we demonstrate that computation produced by earlier interactions can be treated as a resource with a learnable readout, complementing existing approaches that retain, compress, select, or discard conversational context.

We sincerely thank all reviewers again for their valuable feedback. We hope the additional experiments, methodological clarifications, and more comprehensive evaluation provide a stronger and clearer basis for assessing the contribution of our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.