acceptodds
Withdrawn from ICLR 2027

LVMH: A Meta-Optimization Harness for Long-Video Agent Memory

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Long-video agents must map hours or days of multimodal evidence into a bounded context while respecting token and latency budgets, making memory design a context-management challenge. Existing memory systems largely rely on hand-designed policies, leaving open how to jointly optimize memory design for performance and efficiency. We propose Long Video Meta-memory Harness (LVMH), an automated multi-agent framework that formulates memory-policy development as multi-objective code search around a fixed answer model. To guard against self-certification and development-set overfitting, LVMH separates implementation from independent review and enforces protected observation boundaries. Persistent meta-memory records past proposals, traces, costs, and failures to guide subsequent policy revisions. Through this harness, we discover Self-Evolving Evidence Memory (SEEM), a “text-first, vision-second” policy that answers from compact structured event memory and evidence specialists, consulting localized frames through a selective visual gate. The policy evolved on EgoLifeQA generalizes to unseen benchmarks, outperforming Meta-Harness on both Ego-R1 Bench and Video-MME (L) with GPT-5 mini. Across the three benchmarks, SEEM with GPT-5 achieves 65.7% average accuracy, surpassing the previous state of the art by 7.7 percentage points with approximately 32× lower average evaluation token use and 91× shorter average response time.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
long video understanding, agent memory, self-evolving agent

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_HTEuOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes LVMH, a multi-agent system of gpt-5.5. The worker agent iteratively proposes and improves memory policy code for long video QA and the orchestrator agent independently reviews these changes, evaluates the memory policy on heldout validation set, and manages the meta-memory of past evolving trajectory. LVMH is run for 50 iterations on a 100 question dev set of EgoLifeQA, resulting in SEEN, a “text-first, vision-second” memory policy that summaries memory mainly in text assisted with certain important frames. SEEM beats Meta-Harness and memory based baselines on EgoLifeQA, Ego-R1 Bench and Video-MME both on performance and token efficiency.

Critical strengths
  1. Strong accuracy efficiency balance. Table 2 and 9 show SEEN achieves better performance with less cost. This memory augmented method even cheaper than direct request without memory.
  2. Good meta system design. The paper decouples the evolving system (LVMH) and the policy (SEEN). This helps the meta system to be applied on boarder fields.
Critical weaknesses
  1. The evidence of reducing overfitting is weak. Figure 4 still shows a large dev to test gap (71.0 vs 47.5), which is very close to Meta-Harness's. Also the paper doesn't clearly explain whether Meta-Harness uses the validation set to gate the searching process, which may count for the improvement on valid set. Maybe it just use dev and valid set together during search? (In Related Work section the paper critiques Meta-Harness "allow full access to all benchmark examples and evaluation feedback during search"). Also the experiment is only run once, so it's not sure whether the small improvement (47.5 vs 46.0) in Figure 4 is random noise. The trigger and exclusion terms in Appendix B.3 also seem like overfitting on EgoLifeQA. So the claim of reducing overfitting is not well supported.
  2. The evidence of generalization is weak. Ego-R1 Bench uses the same EgoLife recordings and environment as EgoLifeQA, the training data, which means this benchmark measures generalization of question sets rather than domains. On the only out-of-domain benchmark, Video-MME (L), SEEN's improvement over WorldMM and base model is relatively small (GPT-5: 74.4 / 72.6 / 72.3; GPT-5 mini: 55.9 / 52.6 / 54.1). Although SEEN is more efficient, its performance improvement on OOD tests is not clear.
Questions for authors
  1. How is the valid set used in the Meta-Harness reproduction? And what is the total cost of gpt-5.5 in the search process of LVMH and Meta-Harness? Please provide more details of this reproduction process.
  2. Can you report multiple independent search runs (seeds) for LVMH and Meta-Harness?
  3. How much important is the trigger and exclusion terms in Appendix B.3? Can you add ablations to separate the contribution of memories and these seemly hand made terms?
  4. How better gpt-5.5 can do on this benchmarks simply with codex harness (no additional memory module, just 256k default context window and auto compaction)? If the vanilla deployment of gpt-5.5 in codex can achieve good performance, does it worth designing complicated harness and memory policy?
5 comments

We thank the reviewer for recognizing the strong accuracy–efficiency balance of SEEM and the value of separating the meta-optimization framework from the discovered memory policy. We would like to clarify the scope of our generalization results and the role of the discovered heuristics.

Weakness 1 & 2 Concerns of the generalization and performance of our method

1. Comparison with Meta-Harness

Meta-Harness uses the validation set to gate the searching process. Yet LVMH uses the designed gated review.

The API cost of complete 50-round evolution is extremely expensive which prevents us from multi-seed rerun. However, we provide the mutli-seed harness ablation in section 5.2 of the updated manuscript. We show that LVMH has robust generalization performance with better token efficiency than single agent harness.

We emphasize that our claim is to mitigate the risk of overfitting, rather than fundamentally eliminate it. We agree that the 1.5-point improvement on the 200-question EgoLifeQA test is modest. However, it should be used as the only decisive evidence to judge our method. For generalization performance, we believe it is important to look into the cross-benchmark transfer. LVMH searches memory policies on EgoLifeQA, yet the resulting SEEM policy achieves substantially higher accuracy than Meta-Harness on two additional benchmarks:

MethodAnswer ModelEgo-R1 BenchVideo-MME (L)Average
Meta-HarnessGPT-5 mini42.0%47.2%44.6%
LVMH (SEEM)GPT-5 mini53.0%55.9%54.5%
Improvement (pp)—+11.0+8.7+9.85

2. Performance concern

We politely argue that the generalization performance of our method is not weak, instead it is strong.

First, our method achieves overall best performance across the benchmarks with significantly better efficiency improvement. On Video-MME, though the performance improvement compared to the best baseline is modest, our method reduced token use by 40.6× and wall-clock time by 136× compared to agentic memory baseline WorldMM (CVPR 2026 Highlight). We respectfully suggest that evaluating the contribution primarily through absolute accuracy improvements may overlook an equally important dimension of our work: the joint optimization of accuracy, inference cost, and latency. This aligns with one of our core motivations that traditional memory systems typically rely on manually designed architectures, requiring researchers to balance accuracy, computational cost, and latency through human-driven experimentation. LVMH instead formulates these decisions as an automated search over executable memory policies, optimizing multiple metrics at the same time.

Secondly, another advantage of our method is that the memory designed by LVMH is dynamic and evolving, instead of a fixed architecture. We introduce SEEM to demonstrate that the memory policy searched by LVMH is well performed and competitive to human designed methods. We therefore believe the contribution of LVMH should be assessed not only through the magnitude of accuracy improvements over individual baselines, but also through its evolving potentials.

We thank you for the suggestion and will revise the manuscript to make this distinction clearer and to appropriately qualify our claims regarding overfitting mitigation and search robustness.


Q1 Meta-Harness reproduction

  • Valid set is used for the selection of the frontier harness (best so far during the development) in Meta-Harness.
  • The total token cost of the search process for LVMH is ~311M. For Meta-Harness, it is ~481M. The difference mainly comes from the gated review design of LVMH, in which this rule-based mechanism reduced repetitive evaluation on validation set for harness selection.
  • For Meta-Harness, we provided Meta-Harness codebase and the same development benchmark to Codex with GPT5.5 as the proposer.

Q2 Search variance.

We provided multi-seed experiments for LVMH and Meta-Harness in Section 5.2 of the updated manuscript. It can be seen that LVMH shows competitive dev gain and higher transfer gain with smaller variance. As discussed in Q1, benefiting from the rule-based gated review design, LVMH cost less tokens during the search process.


Q3 Trigger and exclusion rules.

The lexical trigger and exclusion rules in Appendix B.3 were not manually designed. They are part of the executable policy discovered by LVMH through autonomous policy search, i.e. designed by agent.

**Our primary contribution is not the individual optimality of these rules, but the framework's ability to discover such policies under accuracy and efficiency constraints.Nevertheless, we agree that the rules may exhibit dataset-dependent behavior and should not be assumed universally optimal. **

In table 3, our existing ablations show that removing gated visual verification reduces EgoLifeQA accuracy from 53.6% to 50.8%. In addition, the evolved gate achieves slightly higher accuracy than an LLM-based gate across all three benchmarks while reducing visual-verification calls from 450 to 112. These results support the effectiveness of the discovered gating policy as a whole.


Q4 GPT-5.5 with a vanilla Codex-style harness.

Thank you for raising this important question. We deeply believe the specific memory harness design is critical for multimodal agents, especially in long-horizon tasks, such as long video understanding.

We conduct evaluation of this method on raw EgoLifeQA video (no caption and transcript as they are part of the memory). We set the maximum frames to be 100, which is the maximum image number allowed for a single Codex request. The final result is surprisingly low as the model failed to retrieve the key information from massive frames.

ConfigurationValue
Model and harnessGPT-5.5 + Codex harness
Evaluation benchmarkEgoLifeQA (full 500 questions)
Visual input budgetUp to 100 frames per question
Accuracy40.00% (200/500)
Input tokens18,902,831
Output tokens458,490
Total tokens19,361,321
Evaluation API cost$119.10

Therefore, Codex+GPT5.5 struggles to solve long video understanding tasks fundamentally.

Reviewer_HTEuOpenReview ↗

Follow up on authors' comments

Thank you for clarifying validation-based selection in Meta-Harness, reporting search costs, and adding multi-seed ablations. I appreciate SEEM's accuracy–efficiency trade-off and transfer gains. Two concerns remain.

1. Overfitting and generalization.

Agent-generated rules can still exploit dataset-specific patterns. Removing visual verification or replacing its gate with an LLM does not isolate the searched trigger/exclusion vocabulary. Could you replace these rules with generic, non-searched defaults while retaining the memory and verification pipeline, and report accuracy and verification frequency on each benchmark? Matching verification budgets and ablating question-form-dependent retrieval weights would help isolate the contribution of dataset-specific rules.

Validation-based selection resolves my fairness concern, but validation is consequently not an independent generalization measure for either method. The test difference remains 47.5% versus 46.0% on 200 questions. Three-seed, 15-round ablations do not establish the stability of the 50-round comparison. Is Figure 5's single-agent variant the Meta-Harness reproduction? Please define its uncertainty bands and provide available per-seed results, qualifying the overfitting claim if full-length repeats are infeasible.

The +8.7pp over Meta-Harness on Video-MME is meaningful. However, as Appendix A acknowledges, Ego-R1 measures transfer within the same EgoLife environment. Video-MME remains the only different-source benchmark, with modest accuracy gains over direct answering and WorldMM. Please report multi-seed transfer results separately for Ego-R1 and Video-MME and reflect this scope in the main claims.

2. Codex baseline configuration.

The new 40% result is useful, but its interpretation depends on the configuration. Please clarify:

  • The Codex client/version, prompt, frame-selection procedure, and tools: could Codex access the source video and extract or revisit frames across requests?
  • The context limit: default 256k, maximum 1M, or a custom setting? What was the auto-compaction threshold?
  • The source of the 100-image restriction: documentation or an observed error, and model endpoint, client, or evaluation setting? Does it apply per request or to the whole trajectory? Current OpenAI documentation lists up to 1,500 images/512 MB per request, subject to context and image-processing limits. This does not establish your endpoint's limit, but warrants clarifying why a multi-step agent was limited to 100 frames overall.
  • Total cache-read tokens, uncached input tokens, and output tokens separately; does the reported input count include cached tokens?
  • Were the 500 questions answered in independent sessions or one shared session? For independent sessions, please report mean context occupancy at the final answer, not cumulative input usage; for a shared session, report total auto-compactions.

I favor sequential evaluation with context retained across questions: this directly tests whether Codex's native long-context management and compaction already provide an effective memory system. This means a shared session, not merely serial execution of independent sessions, while preserving temporal causality and withholding gold answers/correctness feedback.

Finally, withholding captions/transcripts changes evidence access relative to SEEM, so the comparison also reflects preprocessing differences. Please qualify the conclusion as applying to the tested configuration rather than establishing a fundamental limitation of Codex + GPT-5.5.

Reviewer_TY3COpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses memory management for long-video question answering by proposing LVMH, a multi-agent meta-optimization framework that jointly optimizes memory strategies in terms of accuracy, token consumption, and latency through code search, independent review, and persistent meta-memory. SEEM, the strategy discovered by this framework, prioritizes text and uses visual evidence as a supplement, combining structured event retrieval, rule-based evidence computation, and selective visual verification. The authors evaluate the approach on EgoLifeQA, Ego-R1 Bench, and Video-MME, reporting improvements in accuracy and online inference efficiency over the baselines included in the paper. Experiments across different answering models and ablation studies further examine strategy transferability and the contributions of individual components.

Critical strengths
  1. LVMH formulates the design of memory strategies for long-video question answering as an automated search problem and improves the auditability of the search process through independent review, restrictions on information access, and history logging. Structured event retrieval, rule-based experts, and selective visual verification complement one another, enabling the system to preserve key evidence while compressing the context and reducing the number of model calls. Filtering based on event-window end times also helps maintain temporal causality constraints.
  2. The paper evaluates the approach on three long-video question-answering benchmarks and with multiple answering models. It also provides ablation studies of the framework, strategy components, and visual gating, offering evidence for the method’s effectiveness and a degree of transferability. The experiments show that SEEM substantially reduces token consumption and response latency under the main configurations. The appendix further reports offline database construction costs, the tail latency of the visual pathway, and efficiency limits under different configurations
Critical weaknesses
  1. The original WorldMM paper reports accuracies of 65.6% and 65.3% on EgoLifeQA and Ego-R1 Bench, respectively, whereas the results reproduced in this paper are 50.8% and 50.6%. Such large discrepancies require a clear explanation of any differences in inputs, model configurations, and evaluation protocols. Without this clarification, the claim of outperforming the previous state of the art and the associated efficiency comparisons are insufficiently substantiated.
  2. In Figure 4, LVMH’s accuracy on the final test set of 200 questions increases only from 46.0% to 47.5%, corresponding to a net gain of three correctly answered questions. The larger improvements are primarily observed on the validation set, which is repeatedly used for strategy selection. Although the authors clearly disclose the data split, the lack of independent repetitions of the search and uncertainty analysis leaves insufficient evidence to support robust improvements in generalization.
  3. The paper’s characterization of Meta-Harness is overly general. The original Meta-Harness paper already includes evaluation on an unseen test set and recommends that scoring be performed by a mechanism separate from the proposer. Moreover, using the same 50 iterations does not ensure equal numbers of candidates or equal computational budgets. The existing single-run ablation is insufficient to disentangle the contributions of multi-agent collaboration, validation gating, and additional search resources.
Questions for authors
  1. How do the authors explain the substantial gap between the reproduced WorldMM results and those reported in its original paper? What differences are there in the input evidence, model configurations, and evaluation protocols?
  2. Could the authors provide paired results on the final test set, together with an uncertainty analysis? Were the per-iteration results on the transfer benchmarks computed only after the strategy had been frozen, and were they ever used for candidate selection or rule adjustment?
  3. Were LVMH and Meta-Harness matched in terms of initialization, number of candidates, and total computational budget? Are there comparisons under equal computational budgets or independent repeated runs supporting the claim that the advantage stems from the framework design?
2 comments

We thank reviewer for the insightful suggestions and we would like to address your concerns below:

Weakness1 & Q1 Performance gap of WorldMM between reproduced results and its original paper results.

Thank you for raising this point. We have invested several hundred dollars API in extensive efforts to reproduce the results reported in the original WorldMM paper. We experimented with a wide range of open-source and proprietary models across the entire pipeline, from memory construction to question answering. We also explored stronger configurations and regenerated the video captions and transcripts using GPT-5.5 and OpenAI Whisper, respectively. Despite these efforts, we find it struggling to approach the reported results.

Furthermore, our examination of the issues raised in the official WorldMM GitHub repository suggests that other users have encountered similar reproducibility challenges. However, WorldMM is still a representative and strong baseline. We therefore report the strongest results we were able to reproduce based on the official configuration.


Weakness2/Q2 Generalization and statistical uncertainty.

We agree that the improvement from 46.0% to 47.5% on the 200-question EgoLifeQA test is modest. However, the EgoLifeQA test split is not our only evaluation of generalization. The policy discovered on EgoLifeQA substantially outperforms Meta-Harness on two additional benchmarks:

MethodAnswer ModelEgo-R1 BenchVideo-MME (L)Average
Meta-HarnessGPT-5 mini42.0%47.2%44.6%
LVMH (SEEM)GPT-5 mini53.0%55.9%54.5%
Improvement (pp)—+11.0+8.7+9.85

For uncertainty analysis, we provide the mutli-seed harness ablation in section 5.2 of the updated manuscript. We show that LVMH has robust generalization performance with better token efficiency than single agent harness. We are sorry for not able to provide multi-seed rerun of the complete main search experiments due to the API cost.

The per-iteration results on the transfer benchmarks are computed only after the strategy had been frozen, and they were never used for candidate selection or rule adjustment.


Weakness3 Comparison with Meta-Harness

We respectfully clarify that Meta-Harness does not consistently employ independent final testing across its experiments. Its Section 4.3 explicitly states that its TerminalBench-2 experiment performs search and final evaluation on the same 89 tasks. We agree that our Related Work statement was overly broad and will revise it accurately.

As discussed in weakness2, we provided the multi-seed ablation study in Section 5.2 to support the advantages of our harness design.


Q3 Configuration and token cost

LVMH and Meta-Harness share the same GPT5.5 as agent backbone and development set. The number of candidates rounds is the same as 50.

The total token cost of the search process for LVMH is ~311M. For Meta-Harness, it is ~481M. The difference mainly comes from the gated review design of LVMH, in which this rule-based mechanism reduced repetitive evaluation on validation set for harness selection.

Additional Clarification of our contributions

We would like to emphasize that the scientific objective of LVMH is not merely to achieve incremental accuracy improvements on a single benchmark, but to demonstrate that autonomous policy optimization can discover effective and transferable memory systems under multiple competing objectives.

The substantial cross-benchmark improvements, together with the significant reductions in token consumption and latency, provide complementary evidence of this capability.

While we acknowledge the limitations of single-run search evaluation, we believe that focusing primarily on the benchmark improvement does not fully reflect the broader empirical contributions of the proposed framework.

Reviewer_D5GXOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes LVMH, a multi-agent framework that searches over executable memory policies for long-video question answering while keeping the answer model fixed. Workers develop candidates, an Orchestrator reviews and promotes them, and persistent meta-memory records search outcomes. The resulting policy, SEEM, combines structured textual event memory, fielded retrieval, evidence specialists, and selective visual verification. Evaluation on EgoLifeQA, Ego-R1 Bench, and Video-MME (L) reports improved accuracy and favorable inference costs, with additional comparisons across answer models and ablations of both the harness and the discovered policy.

Critical strengths
  • Optimizing executable memory policies is a useful formulation that treats storage, retrieval, and evidence presentation as coupled design choices. The separation between policy evolution and answer-time execution gives the framework a clear technical scope.

  • The transfer results provide meaningful evidence beyond the development setting. With GPT-5 mini, SEEM achieves 55.9% on Video-MME (L), compared with 47.2% for Meta-Harness. Evaluation with GPT-5 and Qwen3-VL-8B also examines whether the discovered policy remains useful with different answer models.

  • The evaluation includes harness ablations, four SEEM component removals, a comparison against an LLM visual gate, and separate memory-construction costs. These analyses support the practical value of the combined system and reveal trade-offs beyond headline accuracy.

Critical weaknesses
  • The validation set repeatedly guides frontier promotion, while the untouched EgoLifeQA test advantage over Meta-Harness is 47.5% versus 46.0%, a net gain of three correct answers. The transfer results and harness ablations support the design in the reported runs, but they do not establish how stable the generalization advantage is across policy-search variability.

  • This paper states that prior harnesses expose all benchmark examples and feedback during search, but Meta-Harness explicitly withholds test results from its proposer and evaluates its math harness on unseen problems. Protected evaluation is therefore not, by itself, an established methodological distinction of LVMH.

  • The gate can reject visual questions because of their wording, and even an opened gate selects frames by ranking caption windows. Caption omissions can therefore also obstruct access to corrective visual evidence. The positive ablation and LLM-gate comparison demonstrate average benefits, but do not establish reliable recovery from the information loss motivating visual escalation.

  • Each matching 30-second record contributes one count, so a prolonged activity can outweigh several shorter occurrences even with accurate captions. Since these counts can trigger direct answers, the grouped specialist ablation does not establish that the computation preserves the meaning of recurring-behavior questions.

Questions for authors

See weakness

2 comments

We thank the reviewer for the detailed assessment and for recognizing the value of executable memory-policy optimization, cross-benchmark transfer, and our accuracy–efficiency analyses. We would like to clarify three points.

Weakness1 Generalization robustness

We agree that the 1.5-point improvement on the 200-question EgoLifeQA test is modest. However, for generalization performance, we believe it is important to look into the cross-benchmark transfer. LVMH searches memory policies on EgoLifeQA, yet the resulting SEEM policy achieves substantially higher accuracy than Meta-Harness on two additional benchmarks:

MethodAnswer ModelEgo-R1 BenchVideo-MME (L)Average
Meta-HarnessGPT-5 mini42.0%47.2%44.6%
LVMH (SEEM)GPT-5 mini53.0%55.9%54.5%
Improvement (pp)—+11.0+8.7+9.85

In particular, Video-MME evaluates transfer to a different video domain and question distribution. These results provide complementary evidence of the discovered policy's transferability, beyond the modest improvement on the EgoLifeQA test split.

We agree that search repetitions would strengthen conclusions about the robustness of LVMH. We provide the mutli-seed harness ablation in section 5.2 of the updated manuscript. We show that LVMH has robust generalization performance with better token efficiency.


Weakness2 Meta-Harness and protected evaluation

We respectfully clarify that Meta-Harness does not consistently employ independent final testing across its experiments. Its Section 4.3 explicitly states that its TerminalBench-2 experiment performs search and final evaluation on the same 89 tasks. Its proposer can also inspect previous candidates' execution traces and evaluation scores, as illustrated in Figure 2.

Our intended methodological distinction is not the invention of held-out evaluation. Instead, LVMH explicitly separates policy implementation from independent Orchestrator review, restricts Workers from accessing validation examples and evaluation traces, and requires gated validation checks before frontier promotion.

We agree that our Related Work statement was overly broad and will revise it accurately.


Weakness3 & 4 Failure modes in the components of SEEM

We appreciate the reviewer's identification of potential failure modes in visual verification and frequency aggregation.We would like to emphasize the distinction between LVMH, the proposed meta-optimization framework, and SEEM, the specific executable policy discovered by that framework.

Our objective is not to manually design individually optimal visual gates or frequency-counting algorithms. Instead, LVMH automatically explores memory representations, retrieval mechanisms, and evidence-processing operations under accuracy and efficiency constraints. Consequently, individual heuristics in SEEM are not claimed to be universally optimal. Their contribution should be understood in the context of the overall discovered policy and its empirical accuracy–efficiency trade-off.

For visual verification, we agree that lexical gating and caption-based localization cannot guarantee recovery of all omitted visual evidence. The gate is instead a cost-aware selective verification mechanism. Our existing ablation shows that removing gated verification reduces accuracy from 53.6% to 50.8%, while Table 7 shows that the evolved gate achieves comparable or better accuracy than an LLM-based gate with substantially fewer visual calls (112 vs. 450).

For frequency aggregation, we agree that counts over fixed temporal windows are not necessarily equivalent to counts of distinct real-world event occurrences, especially for prolonged activities. We will clarify this semantic limitation rather than claim that window-level evidence support is an exact occurrence estimator.

Overall, we believe the central contribution should be evaluated through whether LVMH discovers effective, efficient, and transferable executable memory policies, while acknowledging the limitations of any particular discovered policy.

Additional Clarification of our contributions

We would like to emphasize that the scientific objective of LVMH is not merely to achieve incremental accuracy improvements on a single benchmark, but to demonstrate that autonomous policy optimization can discover effective and transferable memory systems under multiple competing objectives.

The substantial cross-benchmark improvements, together with the significant reductions in token consumption and latency, provide complementary evidence of this capability.

While we acknowledge the limitations of single-run search evaluation, we believe that focusing primarily on the benchmark improvement does not fully reflect the broader empirical contributions of the proposed framework.

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

Related papers

Discussion (0)

Sign in to comment.