acceptodds
Under review as a conference paper at ICLR 2027

DIGAA: Decoupling Information Gathering and Accurate Answering in Retrieval-Augmented Diffusion Language Models

Abstract

Diffusion language models (DLMs) have recently attracted attention in retrieval-augmented generation (RAG) owing to their ability to predict tokens at all masked positions in parallel at each denoising step. These parallel predictions can reveal bridge entities early (Jünger et al., 2026) and provide diverse information, thereby enabling in-depth and broad retrieval. However, this introduces noise along with additional information, making it harder for the model to generate accurate answers. The conflict here is that in-depth and broad retrieval of external knowledge introduces wide-ranging information, while answer generation requires accuracy. Therefore, we introduce Decoupled Information Gathering and Accurate Answering (DIGAA), which explicitly separates the roles of information gathering and answer generation to reduce this conflict. During information gathering, the model uses bridge entities generated during intermediate reasoning to explore external knowledge and collect relevant facts. During answer generation, the model needs to select and integrate this information to answer directly and accurately. Experiments on Dream-7B-Instruct and LLaDA-8B-Instruct across six factual QA benchmarks show that, relative to the strongest baseline for each benchmark and metric, DIGAA improves Dream's average EM and F1 by 4.10 and 5.16 percentage points, respectively, and LLaDA's by 12.39 and 15.61 percentage points. These results indicate that decoupling wide-ranging information gathering from accurate answering is a promising design principle when DLMs are augmented with dynamic retrieval.

Primary area:
foundation or frontier models, including LLMs
Keywords:
diffusion language models, retrieval-augmented generation, dynamic retrieval, factual question answering

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_KdRJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents DIGAA, a training-free framework that separates information gathering from answer generation in retrieval-augmented diffusion language models. During information gathering, the model uses parallel token predictions to guide retrieval and collects information through a second forward pass. A separate generation stage then produces the final answer from the collected information. Experiments on Dream-7B-Instruct and LLaDA-8B-Instruct across six factual QA benchmarks report improvements over nine baselines. The paper also evaluates retrieval settings, inference latency, and the effect of applying the answer-generation stage to existing methods.

Critical strengths
  1. Separating retrieval-oriented intermediate predictions from the final answer is a practical design choice, and the framework requires no additional training.
  2. The evaluation covers two diffusion models, six QA benchmarks, sparse and dense retrievers, and multiple retrieval depths.
  3. Applying the same answer-generation stage to the baselines is a useful control. DIGAA retains improvements over these strengthened baselines, suggesting that its information-gathering procedure contributes beyond answer refinement alone.
Critical weaknesses
  1. The information-gathering ablation does not isolate the proposed design choices. Although Table 4 controls the answer-generation stage, DIGAA changes both the documents used to construct retrieval queries and the number of forward passes per gathering step. It therefore remains unclear whether the gains arise from using fixed initial documents, additional computation, or their interaction. Separate component ablations would clarify this contribution. Retrieval recall or supporting-fact coverage would also help substantiate the claimed broader and deeper retrieval.

  2. The quality and latency comparison uses different decoding budgets across methods. DIGAA is evaluated across several confidence thresholds, while SARDI is fixed at 0.90 and most other baselines use 32 or 64 forward passes. The stronger SARDI + AG baseline is also absent from the latency comparison. Although the reported wall-clock measurements are useful, comparable threshold sweeps and computational budgets are needed to support the claim that DIGAA offers a superior balance between quality and latency.

  3. Several reported quantities are difficult to reconcile with the stated protocol. The no-retrieval Dream baseline scores 15.9 EM on NQ in Table 1, but 14.9 in Tables 6 and 7 and 14.8 in Table 8, despite the stated common decoding settings. Changing retrieval settings should not itself affect this baseline. Moreover, Algorithm 1 requires at least three forward passes and two retrieval calls per question, whereas Table 9 reports averages of 2.8 and 1.9 for Dream at τ = 0.10. These discrepancies require clarification before the experimental comparisons can be assessed confidently.

  4. The answer-generation analysis does not clearly distinguish improvements in factual correctness from changes in answer format. Section 4.2 attributes 61% of the observed improvements to locating answers already present in the first-stage output. The remaining 39% groups inference with word recombination, answer completion, and other edits. Since the analysis only examines improved examples, it also omits cases where refinement damages a correct answer. A separate breakdown of factual corrections, formatting changes, and regressions would better establish the contribution of this stage.

Questions for authors

Please refer to the summary of weaknesses.

Reviewer_nmQrOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2(+1)
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DIGAA, a training-free framework that decouples Information Gathering from Answer Generation in Retrieval-Augmented Generation (RAG) based on Diffusion Language Models (DLMs). Existing approaches utilize intermediate tokens generated during the denoising process for additional retrieval while simultaneously generating the final answer. The authors argue that this coupled process may introduce a conflict between the objective of obtaining diverse retrieval clues and that of generating accurate answers.

To address this issue, DIGAA employs a two-stage generation process. In the first stage, Information Gathering, each denoising iteration consists of a step that predicts retrieval-guiding tokens using a fixed set of initially retrieved documents, followed by a separate prediction step that updates the Knowledge Collection using newly retrieved documents obtained from those predicted tokens. Subsequently, in the Answer Generation stage, the model performs a separate denoising process to generate the final answer based on the completed Knowledge Collection.

Critical strengths

S1. Consistent performance improvements across diverse experimental settings. DIGAA achieves higher EM and F1 scores than all baselines across two DLM backbones and six QA benchmarks. Furthermore, these improvements are consistently observed across different retrieval depths and with both BM25 and dense retrievers. This is significant because it demonstrates that the proposed approach provides performance improvements that are not limited to a particular model or retrieval configuration. Additionally, its training-free nature offers a practical advantage, as it can be applied to existing DLMs without additional model training.

S2. Demonstration of the additional benefits of Information Gathering through comparisons controlling for Answer Generation. The authors conduct experiments in which the Answer Generation stage of DIGAA is applied to the outputs of existing baselines. The results show that introducing a separate Answer Generation stage improves performance, while DIGAA still outperforms the corresponding baselines augmented with Answer Generation (baseline + AG). This is an important experimental strength, as it demonstrates that the performance gains of DIGAA cannot be attributed solely to the additional Answer Generation stage and that the modifications to the Information Gathering process provide further improvements.

Critical weaknesses

W1. Limited methodological novelty. The central idea of this paper, decoupling Information Gathering from Answer Generation, has already been explored in prior RAG research. For example, iterative RAG approaches such as IRCoT perform multiple retrieval steps and subsequently generate a final answer based on the collected evidence. Moreover, the queries used for additional document retrieval in IRCoT serve a role similar to that of the retrieval-guiding tokens used in DIGAA's dynamic retrieval process. In addition, Dynamic Retrieval using intermediate predicted tokens from DLMs has already been introduced by an existing baseline, SARDI. Therefore, the main technical distinction of DIGAA appears to lie in incorporating an established AR LLM-based RAG framework into an existing dynamic retrieval approach, which raises concerns regarding its methodological novelty.

W2. Insufficient ablation studies of the key components. The individual contributions of the proposed components, such as the separation of Information Gathering and Answer Generation and the use of a Knowledge Collection, are not sufficiently analyzed. Although the paper demonstrates the effectiveness of the overall Information Gathering approach through comparisons with existing baselines augmented with Answer Generation (baseline + AG), it does not independently isolate the effect of explicitly decoupling the Answer Generation process. Consequently, the current experiments do not provide sufficient evidence to clearly determine which mechanisms underlying the proposed design are responsible for the observed performance improvements.

W3. Insufficient analysis of the conflict between the two objectives identified as the main concern. The paper identifies the conflict between Information Gathering and Accurate Answering as its central problem. However, it does not provide quantitative evidence demonstrating that this conflict occurs in existing approaches. Furthermore, the paper does not report retrieval-level metrics that would verify whether DIGAA retrieves more diverse or relevant documents than existing baselines. As a result, it remains unclear whether the proposed approach improves retrieval quality compared to existing methods.

Questions for authors

Q1. Clarification on methodological novelty. Could the authors provide a detailed comparison highlighting the substantive technical differences and methodological novelty of DIGAA relative to existing approaches, including AR LLM-based iterative RAG methods such as IRCoT and DLM-based dynamic retrieval methods such as SARDI?

Q2. Additional ablation studies. Unlike existing baselines, DIGAA introduces two distinct components in its Information Gathering process. Could the authors independently evaluate the contributions of these components? For example, I would like to see ablation studies that remove the use of fixed initially retrieved documents for retrieval query generation, or that generate the final answer directly from the retrieved documents without constructing a separate Knowledge Collection.

Q3. Quantitative validation of the central hypothesis. The paper identifies the conflict between Information Gathering and Accurate Answering as its central motivation. Is there quantitative evidence demonstrating that this conflict occurs in existing approaches? Furthermore, could the authors evaluate whether DIGAA retrieves more diverse or relevant documents than existing baselines using metrics such as retrieval recall and evidence coverage?

Q4. Comparison with AR LLM-based RAG. One of the key characteristics of Diffusion Language Models (DLMs) is their ability to predict multiple tokens in parallel, which offers potential advantages in inference efficiency over conventional Autoregressive (AR) LLMs. However, DIGAA performs two forward passes in each denoising iteration and introduces an additional Answer Generation stage, making it unclear whether these potential advantages are preserved in practice. Although the paper compares answer accuracy and inference latency among DLM-based RAG methods, it does not provide comparisons with AR LLM-based RAG approaches. To substantiate the practical benefits of adopting DLM-based RAG, could the authors compare DIGAA with AR LLM-based RAG methods of comparable model sizes in terms of answer accuracy, end-to-end inference latency, and computational cost (e.g., FLOPs or GPU time)?

Q5. Clarification on experimental consistency and statistical significance. The Official Direct baseline, which does not involve retrieval, reports different performance scores across Tables 1, 6, 7, and 8, despite using the same model checkpoints, evaluation datasets, and decoding settings. Since the retrieval depth and retriever configuration should not affect a retrieval-free baseline, could the authors clarify the reason for these discrepancies? Furthermore, could the authors report the mean and standard deviation across multiple independent runs and provide statistical significance tests (e.g., p-values) for the performance improvements of DIGAA over the strongest baselines?

10 comments
AuthorsOpenReview ↗

Response to Weakness 1.1 & Question 1.1: Technical differences and methodological novelty of DIGAA relative to IRCoT

Thank you for the comment. DIGAA's novelty lies in organizing relevant facts while retrieving documents using model outputs, separating predictions used to construct queries from those used to organize facts, and finally generating a direct, accurate answer independently from the organized facts. This allows the model to generate intermediate entities for broad exploration without worrying that this content will be mistakenly retained. Because facts are organized during exploration, the model can generate the answer directly without facing the large amount of irrelevant information in raw documents. This is an advantage unique to DIGAA. Therefore, describing DIGAA as "incorporating an established AR LLM-based RAG framework into an existing dynamic retrieval approach" is inaccurate.

W1.1&Q1.1: Technical differences and methodological novelty of DIGAA relative to IRCoT

Methodological novelty

Relative to IRCoT, DIGAA's specific contribution is to extend "retrieving documents" to "organizing relevant facts while retrieving documents." This allows the model to generate answers directly without facing large amounts of irrelevant information in raw documents or having to select and organize useful information.

Technical differences

  • How retrieved information is collected and organized. IRCoT retrieves documents after each reasoning sentence and uses the document set retrieved throughout reasoning as the collected information. In each round, DIGAA predicts all positions in the knowledge collection in parallel, retrieves documents, and then organizes relevant facts from the newly retrieved documents into the knowledge collection. It uses this knowledge collection as the collected information.
  • Whether answers are generated from documents or organized facts. IRCoT outputs the set of documents retrieved during reasoning. DIGAA collects relevant facts into the knowledge collection after each retrieval. DIGAA therefore passes facts collected and organized by the model, rather than raw retrieved documents.

We replaced the knowledge collection read by DIGAA's answer-generation stage with the retrieved document set, adopting an answer-input format similar to IRCoT's to test the benefit of organizing relevant facts while retrieving documents:

Input to answer generationDream EM / F1LLaDA EM / F1
Reading the document set (IRCoT-style document input)31.17 / 40.001.54 / 14.97
Reading the knowledge collection (DIGAA)33.00 / 43.2730.96 / 42.87

Reading the knowledge collection achieves higher EM and F1 on both models, supporting the benefits of collecting and organizing relevant facts before answer generation.

AuthorsOpenReview ↗

Response to Weakness 1.2 & Question 1.2: Technical differences and methodological novelty of DIGAA relative to SARDI

W1.2&Q1.2: Technical differences and methodological novelty of DIGAA relative to SARDI

Constructing queries from model outputs is precisely how dynamic retrieval methods that use generation to drive retrieval operate. This mechanism is shared by multiple methods, including IRCoT, and is not unique to SARDI. It only shows that these methods belong to the same category of retrieval approaches and does not establish that DIGAA is a simple integration of existing dynamic retrieval methods.

Methodological novelty and advantages

Relative to SARDI, DIGAA's specific contribution is to fully exploit the DLM's parallel prediction capability to support retrieval while preventing the model from retaining intermediate entities used for early exploration in the output.

This is reflected both in the explicit use of two predictions with different responsibilities during information gathering and in the separate final answer-generation step. The model can use intermediate entities for broad exploration, decide which facts to retain based on new evidence, and finally generate a direct, accurate answer independently.

Technical differences

  • Whether exploratory predictions are directly retained in the output. SARDI performs only one forward pass in each round, and its predictions enter both the query and the output, causing intermediate entities used for early exploration to be mistakenly retained. Moreover, the answer may already be fixed while generation and retrieval continue. DIGAA separates the predictions used to construct queries from those used to organize facts. It explicitly uses two predictions with different responsibilities, allowing broad exploration through model outputs while preventing intermediate information from mistakenly entering the output. Finally, it separately generates a direct, accurate answer from the organized facts, further removing information irrelevant to the answer.
  • Which documents are used to generate queries and collect facts. SARDI makes predictions using documents retrieved in the previous round. DIGAA generates retrieval clues using the fixed initial documents and collects facts using newly retrieved documents.

We separately examine the effects of using two predictions with different responsibilities and using fixed documents for the first prediction:

ModelSettingEMF1Retrieval recallFull supporting-document coverage
DreamRolling + single pass (SARDI + AG)30.439.756.027.9
DreamFixed + single pass30.941.156.428.1
DreamRolling + double pass31.841.256.728.6
DreamFixed + double pass (DIGAA)33.043.359.129.5
LLaDARolling + single pass (SARDI + AG)23.532.563.237.5
LLaDAFixed + single pass26.937.863.938.2
LLaDARolling + double pass28.138.264.638.3
LLaDAFixed + double pass (DIGAA)31.042.966.439.9

Using fixed initial documents improves EM and F1 under both the single-pass and double-pass settings. Adding a second prediction improves EM and F1 under both the fixed-document and rolling-document settings. Both designs also improve retrieval recall and full supporting-document coverage on both models, showing that both contribute to the performance gains.

AuthorsOpenReview ↗

Response to Weakness 2 & Question 2: Additional ablation studies

W2&Q2: Additional ablation studies

Following the request, we provide ablation studies that remove the use of fixed initially retrieved documents for retrieval query generation, or that generate the final answer directly from the retrieved documents. The results are shown below:

SettingDream EM / F1LLaDA EM / F1
Without fixed documents31.8 / 41.228.1 / 38.2
Answer generation from documents31.17 / 40.001.54 / 14.97
DIGAA33.00 / 43.2730.96 / 42.87

DIGAA still achieves the best results, supporting the benefits of using fixed initial documents for query generation and organizing facts while retrieving documents during the information-gathering stage.

We further provide component ablations that isolate the two design choices in the information-gathering stage, along with retrieval recall and full supporting-document coverage. The results show that both design choices contribute to the performance gains.

Each round of the information-gathering stage includes the following steps:

  1. Given the question and the documents initially retrieved using the question, the model predicts keywords in parallel and uses them to construct retrieval queries.
  2. These queries retrieve new documents.
  3. The model performs a second forward pass using the question and the newly retrieved documents to collect relevant facts into the knowledge collection.
  4. This process repeats until the knowledge collection is complete.

We vary the following two design choices independently:

  • Documents used to construct retrieval queries: fixed vs. rolling. In step 1, the fixed setting uses the documents initially retrieved using the question. The rolling setting uses the documents retrieved in step 2 of the previous round.
  • Number of forward passes per round of the information-gathering stage: single vs. double. The single-pass setting updates the knowledge collection using predictions from the first forward pass and omits the second pass. The double-pass setting performs another forward pass after retrieving new documents and uses predictions from this second pass to update the knowledge collection.

The table below reports the results of this 2 × 2 ablation. All metrics are expressed as percentages.

ModelSettingEMF1Retrieval recallFull supporting-document coverage
DreamRolling + single pass30.439.756.027.9
DreamFixed + single pass30.941.156.428.1
DreamRolling + double pass31.841.256.728.6
DreamFixed + double pass (DIGAA)33.043.359.129.5
LLaDARolling + single pass23.532.563.237.5
LLaDAFixed + single pass26.937.863.938.2
LLaDARolling + double pass28.138.264.638.3
LLaDAFixed + double pass (DIGAA)31.042.966.439.9

Retrieval recall measures how many gold supporting documents are covered by the union of all documents retrieved throughout the retrieval process. Full supporting-document coverage measures the proportion of questions for which all gold supporting documents are retrieved.

Using fixed initial documents improves EM and F1 under both the single-pass and double-pass settings. Adding a second forward pass improves EM and F1 under both the fixed and rolling settings. Both design choices also improve retrieval recall and full supporting-document coverage on both models.

These results show that both design choices contribute to the performance gains.

We have added the above content to the paper (lines 1188–1231).

Reviewer_7L4MOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper motivates that coupling retrieval and answering in diffusion language models trades off retrieval breadth with answer accuracy. The authors propose DIGAA, a training-free dynamic retrieval method, that separates them into two consecutive stages. The "information gathering" stage iterates between evidence retrieval and text generation across denoising steps, building a reasoning trace. The "answer generation" stage then produces a final answer conditioned on the generated reasoning trace. Comprehensive experiments on single- and multi-hop question answering benchmarks show that DIGAA dominates the accuracy-speed tradeoff among diffusion RAG baselines.

Critical strengths

The idea of decoupling query generation from reasoning generation in the Information Gathering stage is very timely. It allows more aggressive retrieval exploration without committing distracting tokens, naturally extending the dynamic retrieval framework from prior work (Jünger et al.) which couples the two.

Critical weaknesses
  1. I am not convinced by the necessity of the separate answer-generation stage. It is common practice in prior work to direct the model via its prompt to append the answer after the reasoning trace in the format "The final answer is: <final answer>" (Trivedi et al., 2023, Wei et al. 2022). This already decouples reasoning and final-answer generation (the model usually generates the answer after the reasoning trace is completed) and addresses all the failure modes shown in Table 3, while being computationally cheaper.

  2. While the idea of enabling more aggressive retrieval exploration by decoupling query generation from reasoning trace generation is appealing, the execution in the paper does not seem to exploit this opportunity: both the query-generation and fact-collection stage use the prompt from ARAM (Kim & Ye, 2026), simply directing the model to produce an answer. Therefore, the two phases only differ in the provided evidence (question-retrieved documents vs. newly-retrieved documents), offering no advantage in retrieval exploration. Moreover, while the authors claim that this anchoring in question-retrieved documents reduces self-reinforcement during retrieval, it is not ablated in isolation whether this justifies twice as many forward passes.

Questions for authors

Clarification requests:

With "Answer Generation" enabled for both models and confidence thresholds swept for each, does DIGAA still dominate the accuracy-NFE Pareto curve over SARDI?

Suggestions for future work:

I really like the idea of more explorative keyword-based retrieval without worrying about accidentally committing distracting tokens. Possibly, one could explicitly prompt the model in the query-generation phase to directly generate useful keywords for retrieval (instead of just telling the model to answer the question).

3 comments

Thank you for the detailed response and the additional experiment showing the benefit of the separated answering stage.

Thanks for trying out my suggested keyword-prompting idea. Even though the observed accuracy gains are only moderate, I encourage you to motivate your method around this idea. I believe this is the central convincing argument of why it makes sense to separate the "aggressive query generation" from the "reasoning trace generation" phase.

Here is one additional idea: denoising keywords in parallel in masked DLMs is ill-posed, because the keyword-token-distribution will be identical at all token positions. Under greedy decoding, this can lead to the same keywords being sampled at all positions. To improve retrieval diversity, I suggest to not sample these keywords independently, but rather sample them sequentially from their token distributions without replacement.

Overall, I am now convinced that the proposed architecture adds value over the prior works, and in light of the new results I have decided to raise my score.

AuthorsOpenReview ↗

Thank you for raising your score, as well as for recognizing the contributions and value of our paper

Thank you very much for carefully reading our responses and for your more positive assessment of the paper in light of the additional results. We also thank you for recognizing the contributions and value of our paper. We sincerely appreciate the time and effort you have devoted to this discussion.

We agree that enabling broad query exploration while preventing exploratory content from being directly retained in the reasoning trace provides a clearer motivation for this separation. Your keyword-prompting suggestion helped us demonstrate this benefit more concretely. I will follow your suggestion and motivate our method around this idea. The additional experiments have also strengthened the paper.

Thank you also for suggesting sequential sampling without replacement. We appreciate your point about potential keyword repetition under parallel greedy decoding. Your suggestion provides a concrete direction for improving keyword diversity and further exploring the potential of the query-generation stage. I am currently implementing this idea and will incorporate it into the paper shortly.

We are glad that the additional results have addressed your concerns. Thank you again for your constructive suggestions and thoughtful engagement, which have helped us improve both the paper and our understanding of the method.

Thank you again for suggesting sequential keyword selection without replacement. Following our previous response, we have now completed the implementation and obtained the initial experimental results. Your suggestion has improved performance on both models, and we sincerely appreciate your contribution to improving our method.

Due to time constraints, we have so far completed evaluation on 1,000 examples per benchmark. The full-benchmark experiments are still running. The results below compare the original DIGAA with the variant using keyword selection without replacement.

ModelMethodEM ↑F1 ↑Mean latency (seconds/question) ↓
DreamOriginal DIGAA32.4042.793.414
DreamSelection without replacement33.1243.774.122
LLaDAOriginal DIGAA30.3842.4310.703
LLaDASelection without replacement31.3043.4311.214

For Dream, EM and F1 improve by 0.72 and 0.98 percentage points, respectively, while mean latency increases by approximately 20.7%. For LLaDA, EM and F1 improve by 0.92 and 1.00 percentage points, respectively, while mean latency increases by approximately 4.8%.

These results show that your suggested strategy improves answer accuracy on both models, with additional time cost. These gains provide encouraging empirical evidence for this direction and support further investigation. Thank you again for this concrete and constructive suggestion, which has helped us improve the query-generation stage and further explore the benefits of separating retrieval exploration from fact collection. Given the limited time available, we have not yet been able to share an updated manuscript incorporating this suggestion, but we are continuing to work on this revision.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of reviewers feedback and authors response

Sincere thanks to the reviewers for the time devoted to evaluating our work and for the constructive comments. We sincerely appreciate the recognition of the architecture's value in decoupling information gathering from answer generation and organizing facts during broad exploration for DLMs augmented with dynamic retrieval, together with the acknowledgment of the method's consistent performance improvements. We also appreciate the feedback confirming that the additional evidence addressed the corresponding concerns. The comments have been invaluable in improving our manuscript.

New sections added to the revised manuscript:

  • Appendix H: Separate Answer Generation versus Prompted Continuation. Answer quality, forward-pass counts, and observed latency were compared. The separate stage achieves higher EM and F1 with lower observed latency on both models.
  • Appendix I: Prompting the Model to Generate Useful Retrieval Keywords. Explicit keyword prompting was evaluated. It increases the number of distinct retrieved documents and improves answer accuracy, with additional computational cost.
  • Appendix J: Separate Component Ablations. The ablation independently varied fixed versus rolling documents and single versus double forward passes. EM, F1, retrieval recall, and full supporting-document coverage were reported to examine each design's contribution.
  • Appendix K: Breakdown of Factual Corrections and Answer Adjustments in the Answer-Generation Stage. Improvements and regressions were reported separately for both categories, providing a more complete account of the stage's benefits and negative effects.

Sincere thanks again to all reviewers for helping improve the paper. The substantial revisions made in response to the comments have further strengthened 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.