acceptodds
Under review as a conference paper at ICLR 2027

Incentivizing Self-Learning in Large Language Models via Reinforcement Learning

Abstract

Large language models (LLMs) draw on parametric knowledge and retrieval-augmented generation (RAG) to solve knowledge-intensive tasks. Yet problem solving in unfamiliar domains calls for tightly coupled knowledge acquisition and reasoning, beyond factual retrieval and aggregation or learning solely from a preassembled context. We study self-learning—autonomously retrieving, studying, and applying knowledge from long-form reference materials on the fly. To investigate this capability, we formulate a textbook-based mathematical problem-solving task. LLMs must navigate textbook exposition, learn and apply relevant principles, and construct proofs for exercises, with reference solutions withheld. We introduce a BookNavigator interface to support this human-inspired study process and incentivize self-learning through reinforcement learning (RL) with fine-grained rewards. Experiments show that our method improves performance across in-domain and out-of-domain textbooks and subjects, raising Intern-S2-Preview-35B's overall proof accuracy from 51.5% to 79.5% and outperforming RL-trained textbook-free and fixed-retrieval baselines. These gains extend to complex long-context benchmarks involving different task types, with improvements over the base model of 1.7 and 2.3 percentage points on BrowseComp-Plus and CL-bench, respectively. Together, these findings suggest that RL can foster a domain-agnostic self-learning strategy, highlighting the potential of self-learning from long-form corpora to strengthen broader problem-solving capabilities on complex, long-context tasks.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large Language Models, Reinforcement Learning, Self-Learning

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_WAWLOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how to use RL to train LLMs to acquire and apply knowledge from textbooks when solving mathematical proof problems. The authors construct a textbook-based proof task and introduce Book Navigator, which allows models to search and read relevant textbook sections during reasoning.

The results show substantial improvements on mathematical proof tasks and some gains on two cross-task benchmarks.

Critical strengths

1. Good results on mathematical proof tasks.

The method achieves clear improvements over No-Book RL and fixed-retrieval baselines, with good performance on unseen textbooks and mathematical subjects.

2. Thorough experimental analysis.

The paper compares different textbook conditions and includes ablation studies on the exploration bonus and tool credit. These experiments help explain how textbook access and the different reward components contribute to the performance gains.

Critical weaknesses

1. Limited independent validation of the proof evaluation.

The paper uses GLM-5.2 to judge proof correctness, both for RL training rewards and final evaluation. Although the authors use eight independent judgments, I am still concerned about the reliability of LLM-as-Judge for informal mathematical proofs, especially without human evaluation or other independent validation.

2. Limited comparison with related methods.

The paper mainly compares against No-Book and fixed-retrieval baselines. While the authors include an ablation with correctness-only GRPO, they do not provide direct comparisons with existing interactive retrieval RL methods that use step-level rewards or credit assignment, such as StepSearch, SLATE, and IG-Search. As a result, it is difficult to assess how much the proposed reward design improves upon existing approaches.

Questions for authors

Q1. Reliability of Proof Evaluation

Have the authors evaluated the reliability of GLM-5.2 in judging informal mathematical proofs, for example through human evaluation? If so, could you report the agreement rate between the judge and human evaluators? Since the same judge is used for both training rewards and final evaluation, I am concerned about the potential risk of reward hacking.

Q2. Comparison with Existing Methods

Could the authors compare the proposed approach with existing methods that use step-level rewards or credit assignment under the same experimental setting? If direct comparisons are difficult, could you clarify how the proposed approach differs from these methods and what advantages it offers?

2 comments
AuthorsOpenReview ↗

Response to Reviewer WAWL: Independent Evaluation and Interactive Retrieval RL

Thank you for recognizing the mathematical results and the attribution analyses. We address independent validation and the comparison with interactive retrieval RL below.

W1/Q1: Human calibration and independent judges. We calibrated judge accuracy before finalizing the evaluation protocol. Four senior mathematics PhD researchers annotated 90 base-model proofs following the protocol of AdvancedMathBench (Kong et al., 2026). The sample contained 30 proofs each from ID-Book, OOD-Book, and OOD-Subject, yielding 49 correct and 41 incorrect proofs. GLM-5.2 judged each proof eight times with the reference solution. Against these human labels, TPR and TNR measure accuracy on correct and incorrect proofs, respectively; Bal. F1 denotes their harmonic mean:

AggregationTPR (%)TNR (%)Bal. F1 (%)
Average score93.385.389.1
Majority vote95.978.186.1
Pessimistic91.892.592.1

The pessimistic rule balances acceptance of valid proofs with rejection of invalid ones and achieves strong agreement with human experts (Bal. F1: 92.1%), supporting its use for supervision. We attribute this reliability to three factors: (1) the capacity of the 744B GLM-5.2 to evaluate textbook-level proofs generated by InternS2-35B; (2) reference solutions grounding its judgments; and (3) pessimistic aggregation improving TNR (Huang et al., 2025). Together, these results support GLM-5.2 as a reliable source of proof-correctness supervision.

Independent reevaluation supports our conclusion, with no clear sign of judge-specific reward hacking. We reevaluated the main results with GPT-6-astra and Claude-Opus-5.5 as independent judges. Overall accuracy (%):

Method (+ RL)GLM-5.2GPT-6-astraClaude-Opus-5.5
No-Book68.565.967.0
Top-k Sections73.170.571.6
Long-Context72.269.673.7
Oracle Section71.069.470.4
Self-Learning79.578.681.0

Across all Base/+RL entries, scores change by at most 3.4 points. Crucially, Self-Learning retains both the highest post-RL accuracy and the largest RL gain: +28.0/+28.6/+28.9 points under GLM/GPT/Claude. The main conclusion therefore persists beyond the training judge.

W2/Q2: Differences from related interactive retrieval RL. We are happy to clarify how our method differs from prior approaches, as suggested. The key differences lie in supervision and credit assignment:

MethodStep-level signalCredit mechanism
StepSearchGold support-document coverage and reference-query matchingDense retrieval rewards and redundancy penalties within PPO
SLATEJudge scores of candidate reasoning/query steps before retrievalStep-level GRPO over candidates sampled from a shared prefix
IG-SearchGold-answer likelihood under actual versus random retrievalInformation-gain adjustment of query-token advantages
OursActual tool observation and subsequent reasoning, judged by a frozen initial policyCentered round-level redistribution plus correctness-gated exploration

Our design targets sustained, useful exploration of long textbooks. It requires neither gold intermediate retrieval paths nor gold-answer likelihood calculations for tool credit, and evaluates how actual tool returns are used in subsequent reasoning. Centering preserves the trajectory's token-weighted mean advantage while redistributing credit across rounds. The correctness-gated bonus sustains exploration; conservative credit then discourages harmful or redundant calls while preserving reasonable search attempts. This accommodates multiple valid reading paths in open-ended proof construction.

We have added human calibration and independent-judge reevaluation in Appendix D.1–D.2 (Tables 7–8), and expanded the comparison with related interactive retrieval RL methods in Section 4.

Thank you for the detailed response.

Q1 (Proof evaluation). The human calibration study and the re-evaluation with GPT-6-astra and Claude-Opus-5.5 as independent judges are convincing, and they address my concerns about the reliability of the proof evaluation and the risk of reward hacking.

Q2 (Comparison with related methods). Thank you for clarifying the differences from StepSearch, SLATE, and IG-Search in terms of step-level signals and credit assignment. The explanation is reasonable, although direct empirical comparisons would further strengthen the paper.

Remaining concern. The improvements on BrowseComp-Plus and CL-bench over the base model are only 1.7 and 2.3 points, so the cross-task gains are limited, and the claim of a domain-agnostic self-learning strategy seems somewhat overstated.

Since my main concern has been addressed, I will raise my score.

Reviewer_dqukOpenReview ↗
Rating
3(+2)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies how reinforcement learning can train models to consult textbooks and apply the retrieved knowledge during proof construction. It introduces BookNavigator and combines correctness rewards with an exploration bonus and round-level tool credit. The method outperforms textbook-free and fixed-retrieval baselines, with gains on held-out textbooks, subjects, and two external benchmarks.

Critical strengths
  1. The study builds a useful environment with 158 textbooks and 5,183 proof problems, separating reference exposition from exercise solutions.
  2. The baselines also receive RL training, and the full method performs best across all three mathematical evaluation splits.
  3. Textbook substitution, source-section removal, and textbook-free controls provide useful evidence that the reference content contributes to performance.
  4. The reward ablation uses the same interface and training settings. The full method improves accuracy while using fewer tool calls than the exploration-bonus-only variant.
Critical weaknesses
  1. The rewards do not directly measure knowledge gain. A correct answer with one successful get_toc call earns the exploration bonus, while useful and neutral calls receive identical tool scores. Calls without demonstrated learning value can therefore still be encouraged.
  2. Proof accuracy lacks independent validation. GLM-5.2 supplies both training rewards and mathematical evaluation scores. Eight reviews from the same model do not rule out systematic errors, and no expert audit or independent-judge evaluation is reported.
  3. Learning unfamiliar knowledge is not clearly established. Held-out textbooks and subjects are unseen during task-specific RL, but their content may already be familiar to the base model. Cross-book near-duplicate checks are also not reported.
Questions for authors
  1. How often does the policy call only get_toc before answering? Can removing or replacing tool outputs test whether tool credit reflects actual information gain?
  2. Can the authors independently audit final proofs and tool-credit labels, and reevaluate the main results with a different judge?
  3. Can the authors evaluate on controlled unfamiliar materials or newly defined rules, and report multiple training seeds to assess the robustness of knowledge acquisition and transfer?
3 comments
AuthorsOpenReview ↗

Response to Reviewer dquk: Tool Credit, Judge Reliability, Knowledge Acquisition and Robustness

Thank you for the detailed assessment. We clarify the reward implementation and provide independent evaluation and training-seed statistics below.

W1/Q1: Exploration and tool credit. Calling get_toc alone does not earn the exploration bonus: it requires both a correct final proof and actual textbook-content access. We will clarify this condition in place of “at least one successful tool call.” Only 0.47% of evaluation trajectories make one get_toc call and then stop using tools.

Equal scores for useful and neutral calls are deliberate: reasonable exploration can rule out irrelevant chapters without immediately retrieving usable evidence. Tool credit preserves such attempts and penalizes harmful calls, including redundancy. It redistributes credit during RL and is disabled at evaluation. In Table 4 (Table 5 in the revised manuscript), adding it to the exploration-bonus variant raises accuracy from 77.0% to 79.5% while reducing mean tool calls from 5.9 to 3.0, demonstrating its training value without requiring every label to measure information gain exactly.

W2/Q2: Independent proof validation. We calibrated judge accuracy before finalizing the evaluation protocol. Four senior mathematics PhD researchers annotated 90 base-model proofs following AdvancedMathBench (Kong et al., 2026): 30 per split, with 49 correct and 41 incorrect proofs. GLM-5.2 judged each proof eight times using the reference solution. TPR/TNR measure accuracy on human-labeled correct/incorrect proofs; Bal. F1 denotes their harmonic mean:

AggregationTPR (%)TNR (%)Bal. F1 (%)
Average score93.385.389.1
Majority vote95.978.186.1
Pessimistic91.892.592.1

Pessimistic aggregation balances acceptance of valid proofs with rejection of invalid ones, achieving strong agreement with human experts (Bal. F1: 92.1%). We attribute this reliability to (1) the capacity of the 744B GLM-5.2 to evaluate textbook-level proofs generated by InternS2-35B; (2) reference solutions grounding its judgments; and (3) pessimistic aggregation improving TNR (Huang et al., 2025). These results support GLM-5.2 as a reliable source of proof-correctness supervision.

Reevaluation with different judges confirms our main conclusion. We reevaluated the main results with GPT-6-astra and Claude-Opus-5.5. Overall accuracy (%):

Method (+ RL)GLM-5.2GPT-6-astraClaude-Opus-5.5
No-Book68.565.967.0
Top-k Sections73.170.571.6
Long-Context72.269.673.7
Oracle Section71.069.470.4
Self-Learning79.578.681.0

Across all Base/+RL entries, scores change by at most 3.4 points. Self-Learning retains both the highest post-RL accuracy and the largest RL gain: +28.0/+28.6/+28.9 points under GLM/GPT/Claude. The main conclusion therefore persists beyond the training judge.

W3/Q3: Knowledge acquisition and unfamiliar materials. Our evidence comes from controlled comparisons and does not depend on assuming that the textbooks were completely absent from pretraining. In Table 1, Self-Learning achieves 85.2%/86.7% on OOD-Book/OOD-Subject, outperforming No-Book RL by 6.8/12.6 percentage points. More directly, Table 2 holds the trained policy fixed: accuracy reaches 79.5% with textbook access versus 71.7% without it. These results show that Self-Learning strengthens reasoning while learning to use external material to improve proof construction.

OOD-Subject addresses cross-book similarity through subject-level separation. Separating training and evaluation by subject reduces the influence of similarities among textbooks within the same discipline. The 12.6-point gain over No-Book RL demonstrates benefits across subjects, beyond transfer among similar textbooks within one discipline.

W4/Q3: Controlled unfamiliar materials and robustness. CL-bench evaluates learning from expert-created or modified knowledge and rules, providing a controlled test of contextual knowledge acquisition. Our reported RL scores in Table 3 (Table 4 in the revised manuscript) average four independent training seeds. The paired seed comparison on CL-bench is:

QuantityResult
Self-Learning, mean ± sample SD (%)17.1 ± 0.79
No-Book RL, mean ± sample SD (%)16.2 ± 0.76
Mean paired gain (percentage points)+0.90
95% paired t confidence interval[+0.72, +1.08]

All four paired differences are positive, and the interval excludes zero, supporting a positive average gain beyond training-seed variation. The procedural-execution decrease is relative to Base; Self-Learning scores 15.9% versus 15.5% for No-Book RL, the strongest baseline overall. We will report the seed statistics and clarify these comparison points.

We have updated the reward clarifications in Sections 2.4 and 3.1, and added human calibration and reevaluation in Appendix D (Tables 7–8).

Thank you again for your review. We hope that our rebuttal could address your questions and concerns. As the discussion phase is nearing its end, we would be grateful to hear your feedback and wondered if you might still have any concerns we could address. It would be appreciated if you could raise your score on our paper if we address your concerns. We thank you again for your effort in reviewing our paper.

Thank you for the rebuttal and for addressing my questions. After considering your response, I raised my overall score.

Reviewer_yrKsOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper contributes in two ways. It formulates a textbook-based mathematical problem-solving task, extracted 5183 proof problems extracted from 158 textbooks, with exercise solutions strictly withheld. This is where models interact with via BOOKNAVIGATOR (a tool interface supporting get table of contents, search keywords, read full section). The paper also introduces a RL framework based on GRPO. There're two core mechanisms, adding an annealed exploration bonus gated on solution correctness to prevent the model from abandoning tool calls, and round-level tool credit to redistribute learning advantages based on intermediate tool utility. The framework shows ~28% improvement on mathematical proof accuracy base model, also outperforming fixed-retrieval baselines by ~6%.

Critical strengths
  1. Evaluation where the policy trained on textbook proof construction achieves gains on general deep research and in-context learning benchmarks under their original inference protocols provides strong evidence for the self-learning RL capability.
  2. The experiment includes detailed attribution interventions including topic-matched textbook subtitution, source section removal and neighborhood removal to isolate component gains from memorization of specific source texts, provining soundness of the paper.
  3. The paper addresses an important constraint in agentic RL and maximizes cross-domain generalization.
Critical weaknesses
  1. the introduction currently frames the paper primarily around "textbook-based mathematical proof solving", which buries its core technical contribution to solve the tool collapse failure mode in agentic RL. And the method for using correctness-gated exploration bonuses and centered round-level credit are brought up in later sections. Elevating this earlier would significantly clarify the paper's motivation and technical novelty and strength.
  2. While Table 4 tracks average tool calls, the paper lacks an explicit evaluation of token overhead, computational resource usage, or inference latency. Because the method expand context up to 64k tokens, analyzing the trade-off between compute cost and accuracy gains is necessary.
Questions for authors
  1. What is the average token footprint and inference cost/latency per trajectory for the self-learning RL approach compared to no-book or fixed retrieval baselines?
  2. What is the agreement rate between the pessimistic GLM-5.2 judge and human expert mathematicians on a sampled subset of math proofs?
2 comments
AuthorsOpenReview ↗

Response to Reviewer yrKs: RL Motivation, Inference Cost and Judge Reliability

Thank you for highlighting the technical contribution and the evidence from attribution interventions and cross-task transfer.

W1: Foregrounding the RL contribution. We agree. The revised abstract and introduction will introduce tool collapse first and replace the phrase “with fine-grained rewards” with the two mechanisms: correctness-gated exploration bonuses sustain textbook access, and centered round-level credit improves the allocation of learning signals across interactions. The textbook task provides the environment in which this training problem is studied and measured.

W2/Q1: Accuracy–cost trade-off. We report average token usage per trajectory alongside Table 1's Overall accuracy. Cumulative input and output tokens are summed across all interaction rounds; trajectory length counts the trajectory content once.

Base

MethodAccuracy (%)Input tokensOutput tokensTrajectory tokens
No-Book55.72839,86910,152
Top-k Sections62.237,0697,71644,785
Long-Context56.753,2277,48360,711
Oracle Section60.915,4368,46123,898
Self-Learning51.523,1355,81512,131

+ RL

MethodAccuracy (%)Input tokensOutput tokensTrajectory tokens
No-Book68.528318,60718,890
Top-k Sections73.137,06917,38054,449
Long-Context72.253,22718,63771,865
Oracle Section71.015,43617,03832,474
Self-Learning79.537,19721,14324,025

Self-Learning provides a competitive accuracy–token trade-off. Relative to the strongest fixed-retrieval baseline, Top-k Sections RL, it achieves +6.4 percentage points with nearly identical cumulative input (+0.35%), 21.7% more output, and a 55.9% shorter trajectory. Relative to Long-Context RL, it achieves +7.3 points with 30.1% fewer input tokens. This supports selective, interactive textbook access: additional generation and interaction yield higher accuracy at comparable input volume and with a more compact trajectory.

Q2: Agreement with human experts. We calibrated judge accuracy before finalizing the evaluation protocol. Four senior mathematics PhD researchers annotated 90 base-model proofs following the protocol of AdvancedMathBench (Kong et al., 2026). The sample contained 30 proofs each from ID-Book, OOD-Book, and OOD-Subject, yielding 49 correct and 41 incorrect proofs. GLM-5.2 judged each proof eight times with the reference solution. Against these human labels, TPR and TNR measure accuracy on correct and incorrect proofs, respectively; Bal. F1 denotes their harmonic mean:

AggregationTPR (%)TNR (%)Bal. F1 (%)
Average score93.385.389.1
Majority vote95.978.186.1
Pessimistic91.892.592.1

The pessimistic rule balances acceptance of valid proofs with rejection of invalid ones and achieves strong agreement with human experts (Bal. F1: 92.1%), supporting its use for supervision. We attribute this reliability to three factors: (1) the stronger 744B GLM-5.2 can effectively evaluate textbook-level proofs generated by InternS2-35B; (2) reference solutions ground its judgments; and (3) pessimistic aggregation improves TNR (Huang et al., 2025). Together, these results support GLM-5.2 as a reliable source of proof-correctness supervision.

We have updated the abstract to foreground the RL contribution, added the accuracy–token analysis in Section 3.3 (Table 3), and documented human calibration of the judge in Appendix D.1 (Table 7).

Thank you for the thorough and constructive rebuttal.

W1: Foregrounding the challenge and reward mechanisms in the revised Introduction significantly improves the paper's positioning and clarity. W2/Q1: The new token comparison in Table 3 convincingly addresses my concern. Demonstrating that Self-Learning + RL outperforms while maintaining comparable or lower cumulative input tokens and a much more compact trajectory makes the efficiency case clear. Q2: The human validation study with mathematics PhD annotators provides strong empirical support for the reliability of the pessimistic GLM-5.2 verifier.

All of my primary concerns have been fully addressed. I view this as a solid submission and the score already reflects this.

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

The authors formulate a textbook-based mathematical proof task to study self-learning. The task asks the LLMs to study the texbook, try to solve the problems with the knowledge learned from the text book. An judge LLM is then used to evaluate the solutions and provide training signals for reinforcement learning. The authors develop a dedicated interaction system called BOOKNAVIGATOR for tool-calling learning process about the textbooks, and adopt fine-grained rewards. The experimental results show a decent improvement with RL from the baseline model.

Critical strengths
  1. The paper is well written and easy to read. The methodology is well explained, and an extensive experimental result is provided.
  2. The authors contribute a new task for self-learning, alongside with a good framework to study it. The task makes sense to me and could be helpful to the research community. I think the task is relatively more important in this paper since the methodology part is less complicated.
Critical weaknesses
  1. Even though I acknowledge the efforts authors have made in building the self-learning system, the methodology part remains primitive to me. It would be helpful if authors could provide more context on the previous study of this specific field, i.e., self-learning for mathematical textbooks.
  2. Is there any more baselines from previous papers? Right now I can see an ablation study in table 4 showing the contributions of the minor methodologies, but there is no RL baseline. Also, I believe the "RL" refers to the RL method authors proposed in section 2, is it possible to have some alternatives of it?
Questions for authors
  1. See critical weakness 2
  2. I would like the authors to clarify their contribution emphasis: is this a task-contribution paper or a methodology paper? The main content looks more like a task-contribution paper, but the title does not. I am also worried that the title is overclaiming, because the authors study in this specific area of self-learning mathematical textbooks, so I don't think the paper justifies any generalizability beyond.
2 comments
AuthorsOpenReview ↗

Response to Reviewer cSUw: Methodological Novelty, RL Baselines and Generalization

Thank you for recognizing the value of the task and the extensive evaluation. Our contribution combines a textbook-learning environment with an RL method that sustains useful knowledge acquisition.

W1: Context and methodological novelty. We are happy to provide more context. Prior work provides a foundation for knowledge-assisted mathematical proving: NaturalProofs (Welleck et al., 2021) studies retrieval of mathematical references; NaturalProver (Welleck et al., 2022) generates natural-language proofs conditioned on retrieved or supplied references; LeanDojo/ReProver (Yang et al., 2023) retrieves formal premises for Lean proofs. We will expand this context in the revision. Our focus is learning an interactive policy that identifies a knowledge gap, navigates long textbooks, and applies the retrieved material within proof construction. Unlike conditioning on a fixed reference set, textbook access is part of the learned decision process.

W2/Q1: Clarifying “RL”. Thank you for pointing out this ambiguity. In Table 4 (Table 5 in the revised manuscript), the “RL” label should read Correctness-only GRPO. It uses the same Self-Learning interface and training setup, providing a standard RL baseline against which the reward components are isolated. As summarized below, Table 4 (Table 5 in the revised manuscript) already compares tool credit and the exploration bonus against this RL baseline. In Tables 1–3 (Tables 1, 2, and 4 in the revised manuscript), “+RL” denotes the full method: correctness-only GRPO + exploration bonus + tool credit. We will correct these labels.

Training objectiveOverall accuracy (%)
Correctness-only GRPO74.9
+ Tool credit76.2
+ Exploration bonus77.0
+ Both (full method)79.5

Q2: Contribution and scope. Both the task and the method are central: the task makes textbook learning trainable and measurable, and the method addresses the resulting exploration failure. Generalization beyond textbook proofs is directly tested in Table 3 (Table 4 in the revised manuscript), using the external benchmarks' original inference protocols without BookNavigator. Self-Learning reaches 51.9% on BrowseComp-Plus versus 49.9% for the strongest RL baseline, and 17.1% on CL-bench versus 16.2%. These results demonstrate that Self-Learning generalizes to retrieval-augmented reasoning and in-context learning tasks.

We have updated the manuscript with expanded related-work context (Section 4) and clearer RL baseline definitions (Section 3.1 and Table 5). Cross-task generalization results are now in Section 3.4, Table 4.

Thank you again for your review. We hope that our rebuttal could address your questions and concerns. As the discussion phase is nearing its end, we would be grateful to hear your feedback and wondered if you might still have any concerns we could address. It would be appreciated if you could raise your score on our paper if we address your concerns. We thank you again for your effort in reviewing our paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.