acceptodds
Under review as a conference paper at ICLR 2027

Pretrain Once, Route Anywhere: Towards a Foundation Model for LLM Routing

Abstract

Large language model (LLM) routing aims to assign each query to the most suitable model from a heterogeneous candidate pool, improving the quality–efficiency trade-off of LLM inference. Existing routers are typically learned through *local fitting*: a router is optimized for a particular query workload and candidate pool, and often requires additional supervision or retraining as the routing environment changes. We ask whether LLM routing can instead be approached from a **foundation-model** perspective, learning a reusable routing capability that generalizes across tasks, candidate models, and deployment conditions. To this end, we introduce **RouteFM**, which learns to characterize anonymous candidate models from behavioral context and infer their target-specific capabilities, rather than binding routing decisions to fixed model identities or a single environment. Through episodic pretraining across heterogeneous routing environments, this capability can be reused by a frozen router and adapted to new environments through context alone. Experiments demonstrate transfer across changes in domains, modalities, candidate pools, and context budgets, with the largest gains when behavioral evidence is limited. On MMR-Bench, which is excluded from pretraining, RouteFM outperforms the strongest baseline by 2.23 quality points with only eight observations per candidate. These results support moving LLM routing from repeated local fitting toward a *pretrain once, route anywhere* paradigm.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM Routing, Foundation Models, Model Selection, In-Context Capability Inference, Adaptive Inference

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

This paper addresses the limitations of existing LLM Routers, which typically rely on specific task distributions and candidate model pools, making them difficult to adapt to dynamic routing environments. To overcome these challenges, the paper proposes RouteFM, a general-purpose routing model based on cross-environment pretraining. RouteFM constructs capability representations of anonymous candidate models from their historical behavior, enabling the pretrained Router to adapt to new routing environments without parameter updates. Experimental results demonstrate that RouteFM achieves promising performance in both In-Domain Routing and Cross-Modal Transfer settings.

Critical strengths

S1. This paper investigates cross-environment generalization from the perspective of reusable routing capabilities, offering clear research value and practical significance.

S2. The design of RouteFM is well-motivated and reasonable, and the proposed method achieves competitive experimental results.

Critical weaknesses

S1. Insufficient disclosure and analysis of the candidate model pools. The paper lacks detailed statistics on the candidate models used during pretraining and testing. This may weaken the paper's claims regarding generalization across different candidate model pools.

S2. Lack of ablation studies on key components. The paper lacks necessary ablation studies on several critical design choices, such as the Capability Profile Branch and the loss function.

Questions for authors
  1. Could the authors provide more details about the candidate model pools?

(a) Which candidate models are used during pretraining? What is the size and diversity of the overall model pool in terms of model families, parameter scales, and the composition of open-source and closed-source models?

(b) To what extent do the candidate models used for testing overlap with those used for pretraining? How do their distributions differ in terms of model families and parameter scales?

(c) How is the candidate model pool constructed for each pretraining episode? Are models randomly sampled, or is a specific sampling strategy employed based on model capabilities, families, or other factors?

  1. Could the authors provide additional ablation studies on the following key components?

(a) Capability Profile Branch: Table 2 includes an ablation of Target-to-Context Attention, but lacks a corresponding ablation of the Capability Profile Branch. The authors are encouraged to evaluate a variant that entirely removes capability profiling and retains only the Raw Context Branch, to assess the independent contribution and necessity of the Capability Profile.

(b) Training Objectives: RouteFM employs three training objectives: Point-wise Regression, Pairwise Ranking, and Routing-aware Regret. The authors are encouraged to ablate each loss term individually to assess its contribution to the final routing performance.

3 comments
AuthorsOpenReview ↗

Response to Reviewer AEru

We sincerely thank the reviewer for recognizing the research value, practical significance, and effectiveness of RouteFM, as well as for the constructive suggestions. In response, we provide detailed statistics on candidate model pools and training–test overlap, clarify our episodic pretraining procedure, and conduct additional ablation studies on key architectural components and training objectives. We address these concerns below.


R1: Could the authors provide more details about the candidate model pools?

We thank the reviewer for this valuable suggestion. We have added detailed candidate pool statistics and training–test overlap analysis in the revised Appendix B.1, along with further clarification of the episodic pretraining procedure in Appendix B.2.

(a) RouteFM is not trained on a single fixed model pool. Instead, it combines task-specific candidate pools from LLMRouterBench, RouterBench, RouterEval, and MixInstruct. Together, these sources contain 182 unique models across 15 major model families, including Llama, Qwen, Mistral, DeepSeek, Gemma, GPT, and Claude. The pool includes 170 open-weight models and 12 proprietary API models, covering a wide range of parameter scales.

(b) None of the nine candidate models in MMR-Bench appears in pretraining (0/9 exact model-ID overlap), although some model families overlap (5/6 families). Importantly, RouteFM treats all candidate models as anonymous, without access to their names, identities, or provider information. Instead, it infers candidate capabilities solely from behavioral observations, enabling adaptation to unseen models without relying on fixed model identities.

(c) As detailed in Appendix B.2, each pretraining episode is constructed from a single source and task, with 6–24 candidate models and varying behavioral context sizes. Candidate ordering is randomly permuted to prevent dependence on fixed model positions. Target queries are sampled from three regimes: Natural (65%), Opportunity (20%), and Boundary (15%), encouraging the router to learn target-conditioned model comparison.


R2: Could the authors provide additional ablation studies on the following key components?

We thank the reviewer for these valuable suggestions. Following the reviewer's suggestions, we conduct additional ablation studies on MMR-Bench to evaluate the contributions of the Capability Profile Branch and the three training objectives. The new results have been incorporated into Table 2 of the revised manuscript.

MethodK=8K=16K=32K=64Large
RouteFM0.73230.74570.74870.75040.7614
w/o capability profiling0.73570.74220.74060.74430.7520
w/o point-wise regression0.71570.74430.74320.74590.7529
w/o pairwise ranking0.73630.74420.73070.74030.7397
w/o routing-aware regret0.72670.74450.75160.75030.7529

(a) Capability Profile Branch

Removing capability profiling slightly improves performance at K=8, but consistently reduces performance at larger context sizes, with a 0.94-point drop in the Large setting. This suggests that direct access to raw observations can be sufficient when context is extremely limited, while capability profiling becomes increasingly valuable as more behavioral evidence is available, providing a compact global summary that complements fine-grained context retrieval.

(b) Training Objectives

The three objectives make complementary contributions. Removing point-wise regression consistently degrades performance, particularly at K=8 (−1.66 points), highlighting its importance for accurate capability estimation. Pairwise ranking provides substantial gains at larger context sizes, especially K=32 (1.80 points) and Large (2.17 points), demonstrating the importance of explicitly learning relative candidate preferences. Routing-aware regret provides additional improvements at K=8 (0.56 points) and Large (0.85 points), although its contribution varies across context sizes.

Overall, these results clarify the distinct roles of capability profiling and the three training objectives, as well as how their contributions vary with the amount of available behavioral evidence.

Thanks for the authors' further clarification and ablation studies. This is an interesting attempt to bring foundation model pretraining to LLM routing. I therefore tend to keep my overall positive score.

AuthorsOpenReview ↗

Thank you for your positive feedback

We sincerely thank the reviewer for their time and thoughtful consideration. We greatly appreciate your positive feedback and continued support of our work.

Reviewer_ZyM9OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary
  • RouteFM uses a pre-trained router that generalizes across tasks, candidate models, and deployment conditions. It looks at a few past examples of how each model performed and determines routing in-context.
  • It compresses each model's performance history into a capability profile.
  • For unseen data, there are strong gains when only a few examples per model are available.
  • For seen data or large data sets, the gains are smaller.
Critical strengths
  • Novel approach by pre-training a router. This allows for a lot of flexibility to adapt to new situations.
  • Good performance for cold start scenarios. This potentially solves a practical problem with new model deployments.
Critical weaknesses
  • Unclear how this will work in new paradigms (e.g., other modalities like audio, long-thinking, etc.)
  • Minimal gains over cheaper models when there's many examples per model.
Questions for authors
  1. [Clarification] Does pre-training need to happen again to introduce audio support?
  2. [Clarification] Will RouteFM work for reasoning models (e.g., with different thinking budget options)?
4 comments
AuthorsOpenReview ↗

Response to Reviewer ZyM9 — Part (1/2)

We thank the reviewer for recognizing RouteFM's novelty and practical value, particularly its flexibility in cold-start scenarios. In response to the constructive suggestions, we conduct additional experiments on audio routing and long-thinking strategies, and clarify RouteFM's performance across different context sizes. We address these concerns in two parts below.


R1: Generalization to new paradigms (Audio and Long-thinking)

We thank the reviewer for these valuable suggestions. Following the reviewer's suggestions, we conduct two additional experiments on audio routing and long-thinking strategy routing to further investigate RouteFM's generalization to new paradigms. In both settings, RouteFM is applied directly with frozen parameters, without additional pretraining or fine-tuning.

(1) Audio Modality: SEAME Dev MAN

We construct an audio routing benchmark from SEAME Dev MAN, containing 2,610 speech samples and 11 candidate audio models. Each sample is evaluated using recognition quality defined as 1 - clipped WER, producing a 2,610 × 11 query–model quality matrix. We randomly split each episode into 80% context pool and 20% target queries, evaluating K ∈ {8, 16, 32, 64} historical observations per candidate.

Since the current Qwen3-VL and BGE encoders do not support audio inputs, we use Whisper to transcribe speech into text for query representation. These transcriptions are generated independently of the reference labels, which are used only for evaluation.

MethodK=8K=16K=32K=64
Weighted kNN0.56700.57330.57610.5821
MLP0.56870.56690.56750.5451
Embedding-Ridge0.57650.57520.57740.5804
EmbedLLM0.56780.57530.56560.5372
RM-Softmax0.56770.57620.57650.5818
TRouter0.56770.57090.57250.5454
InferenceDyn.0.57570.57240.57570.5821
UniRoute0.49770.54430.57650.5821
ICL-Router0.55870.57490.57570.5804
Context-Mean0.57570.57240.57570.5821
RouteFM0.57650.57990.57740.5856

RouteFM achieves the best or tied-best performance across all context sizes. Although the gains are modest, these results demonstrate that RouteFM can transfer to audio routing through transcription-based query representations without any additional routing pretraining.

For future work, we plan to extend RouteFM to support native multimodal routing. One promising approach is to employ interleaved attention to compress variable-length modality embeddings and behavioral contexts into fixed-size representations, enabling direct integration of audio encoders and other modality-specific encoders into the routing framework.

(2) Long-thinking: GSM8K-Strategies

We further evaluate whether RouteFM can adapt to different inference-time reasoning strategies. We construct GSM8K-Strategies using 500 GSM8K problems, where Qwen2.5-14B-Instruct solves each problem using three candidate strategies: DirectAnswer (256-token limit), Chain-of-Thought (1024-token limit), and Self-Consistency (three sampled reasoning paths with majority voting). We treat these reasoning strategies as anonymous candidates and apply the frozen RouteFM to select the most suitable strategy for each query based solely on historical behavioral observations. Results are averaged over three independent context sampling seeds.

MethodK=8K=16K=32K=64
Random0.72220.72220.72220.7222
Weighted kNN0.95330.95330.94890.9556
Embedding-Ridge0.95330.95330.94670.9489
UniRoute0.94890.94890.94890.9444
EmbedLLM0.94890.94890.94890.9489
RM-Softmax0.95330.95330.94890.9489
Context-Mean0.95330.95330.94890.9489
RouteFM0.95330.95780.96000.9622

RouteFM achieves the best or tied-best performance across all context sizes, reaching 96.22% accuracy at K=64, close to the 96.67% oracle upper bound. This demonstrates that RouteFM can also generalize from model selection to inference-time reasoning strategy selection without retraining.

Together, these experiments provide additional evidence that RouteFM learns a reusable routing capability that extends beyond the modalities and routing configurations encountered during pretraining.

AuthorsOpenReview ↗

Response to Reviewer ZyM9 — Part (2/2)

R2: Minimal gains over cheaper models when there's many examples per model.

We acknowledge that RouteFM's gains become smaller as more behavioral observations are available. However, the primary contribution of RouteFM is learning a transferable routing capability through heterogeneous pretraining, allowing a single frozen router to generalize across tasks, candidate pools, and modalities without environment-specific retraining.

This capability is particularly valuable in low-context settings, where conventional routers have insufficient supervision to learn effective routing policies. As context size increases, conventional methods benefit from richer target-environment supervision and naturally narrow the performance gap. Nevertheless, RouteFM maintains competitive or superior performance without any parameter updates. Thus, the diminishing gains are expected and further highlight RouteFM's value in enabling effective and flexible deployment across changing routing environments.

Thank you for the additional studies on audio and long-thinking. Results look good and I'm maintaining my positive-leaning score.

AuthorsOpenReview ↗

Thank you for your positive feedback

We greatly appreciate your positive feedback and continued support of our work. Thank you again for your time, careful review, and constructive suggestions.

Reviewer_8pP1OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper tries to study how to choose which language model to answer which query. Many routers are trained for a particular setting, but RoutesFM is designed to take the approach of pre-training a foundation model such that it can learn in context instead of requiring retraining for new domains. I heard you. The experiments cover held-out router queries, and the paper shows that RouteFM is able to achieve a high reported quality that transfers to a different benchmark.

Critical strengths

The set of experiments is relatively comprehensive and careful, covering both architecture and data ablations, as well as reproductions of many other baselines on a held-out benchmark. They also report that their gains are mostly in the few-shot regime and diminish with more examples.

Critical weaknesses

The empirical gains they measure are small and the baseline set seem to be weak, especially since some of the methods have very similar results. Furthermore, it is unclear how to quantify the gains of this method relative to the total amount of possible headroom. The evidence of generalization relies on one benchmark, which is this MMR bench, and the model transfer benefits from the multimodal encoder it starts with. So it is unclear how much of this is borrowed. Most importantly, the idea of pre-training to generalize across tasks is not novel.

Questions for authors

How reliable are the gains? And what is the degree of uncertainty there? Can you put the gains in context relative to an article or some other baselines? Why do some of the learned baselines behave identically? Can you test transfer on a broader suite of independent benchmarks?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 8pP1 — Part (1/3)

We sincerely thank the reviewer for the careful evaluation and constructive feedback. In response, we have conducted additional experiments and analyses to assess the reliability of our improvements, quantify the gains relative to oracle headroom, and validate generalization on additional benchmarks and modalities. We also clarify the behavior of certain baselines, the role of the multimodal encoder, and the novelty and practical value of RouteFM. We address each concern in detail below.


R1:Novelty of the Proposed Routing Paradigm

We agree that pretraining for cross-task generalization is not novel in itself. However, LLM routing presents a distinct challenge: both query distributions and candidate model pools vary across environments, and newly introduced models must be continuously incorporated into existing routing systems. Conventional routing policies tied to fixed model identities often require retraining or adaptation when the candidate pool changes.

Our novelty lies in introducing a foundation-model paradigm for LLM routing that learns a reusable routing capability rather than an environment-specific policy. Through heterogeneous episodic pretraining, RouteFM learns to infer the capabilities of anonymous candidate models from behavioral observations and perform target-conditioned model comparison. This enables a single frozen router to generalize across tasks, candidate pools, and modalities, while also incorporating newly introduced models into existing pools with only a few behavioral observations, without retraining.

In practice, this shifts routing adaptation and new-model incorporation from repeated parameter optimization to context construction, reducing deployment and maintenance costs in dynamic routing environments.


R2:Empirical gains are small. How reliable are the gains?

R2.1: Empirical gains are small

We acknowledge that RouteFM's performance gains are relatively modest in large-context settings. However, our primary contribution lies in learning a transferable routing capability through heterogeneous pretraining, enabling a single frozen router to generalize across tasks and candidate pools, and incorporate newly introduced models with only a few behavioral observations, without environment-specific retraining.

This pretrained capability provides a particularly strong advantage in low-context settings, where conventional routers struggle to learn effective routing policies from limited supervision. As context size increases, conventional routers benefit from more sufficient training data and naturally narrow the performance gap. Nevertheless, RouteFM remains competitive or superior without any parameter updates.

Therefore, the diminishing gains are expected and consistent with our design goal: learning a reusable routing capability that generalizes to new environments with minimal behavioral observations, rather than requiring repeated environment-specific training.

R2.2: How reliable are the gains? And what is the degree of uncertainty there?

To assess the reliability of our improvements, we additionally conducted a statistical uncertainty analysis using 10,000 paired cluster bootstrap resamples. We report the performance improvements over the strongest baseline, together with 95% confidence interval margins.

SettingK=8K=16K=32K=64
In-Domain Routing+0.0084 ± 0.0045+0.0003 ± 0.0015+0.0026 ± 0.0017+0.0010 ± 0.0006
Cross-Modal Transfer+0.0223 ± 0.0106+0.0119 ± 0.0053+0.0054 ± 0.0080+0.0064 ± 0.0042

The results demonstrate that RouteFM's improvements are reliable, with relatively small uncertainty.


R3:Why do some learned baselines behave identically?

Our baselines cover both established and recent LLM routing methods, implemented following the ORBIT toolbox and their original repositories under a unified evaluation protocol.

The identical performance of certain methods, such as TRouter and InferenceDynamics, is caused by routing collapse [1, 2], a known phenomenon in which learned routers degenerate into selecting the locally optimal candidate based on overall context performance, rather than making query-specific routing decisions. We further verified that these methods select the same candidate as Context-Mean, explaining their identical results. We will clarify this phenomenon in the final version of the paper.

[1] When Routing Collapses: On the Degenerate Convergence of LLM Routers. CoRR 2026.

[2] SkillOrchestra: Learning to Route Agents via Skill Transfer. CoRR 2026.

AuthorsOpenReview ↗

Response to Reviewer 8pP1 — Part (2/3)

R4:How much possible headroom is left?

We thank the reviewer for this valuable suggestion. We additionally evaluate the fraction of oracle headroom captured by each method, defined as: Closed Headroom=Qmethod−QrandomQoracle−Qrandom,\text{Closed Headroom} = \frac{Q_{\text{method}} - Q_{\text{random}}}{Q_{\text{oracle}} - Q_{\text{random}}}, where QrandomQ_{\text{random}} and QoracleQ_{\text{oracle}}​ denote the routing quality of random selection and oracle routing, respectively.

In-Domain Routing

MethodK=8K=16K=32K=64
Weighted kNN36.8%43.1%46.6%47.1%
MLP35.4%45.4%51.8%52.3%
Embedding-Ridge38.1%45.3%52.0%50.5%
EmbedLLM35.4%47.8%52.0%53.0%
RM-Softmax37.8%46.6%53.0%52.5%
TRouter36.7%45.1%51.7%52.0%
InferenceDyn.36.7%45.1%51.7%52.0%
UniRoute37.6%47.3%51.4%52.5%
Context-Mean36.7%45.1%51.7%52.0%
RouteFM40.6%47.9%53.8%53.3%

Cross-Modal Transfer

MethodK=8K=16K=32K=64
Weighted kNN30.0%34.7%35.9%37.0%
MLP31.6%39.9%41.5%43.6%
Embedding-Ridge32.6%40.2%43.3%43.9%
EmbedLLM34.0%41.6%42.8%43.0%
RM-Softmax31.4%39.9%43.8%44.7%
TRouter32.9%40.8%44.7%44.9%
InferenceDyn.32.9%40.8%44.7%44.9%
UniRoute30.4%39.2%42.2%44.6%
Context-Mean32.9%40.8%44.7%44.9%
RouteFM41.2%45.5%46.4%47.0%

The results show that RouteFM consistently achieves the highest closed headroom across all context sizes in both in-domain routing and cross-modal transfer, making its performance advantage more evident than absolute quality gains alone.


R5:Generalization only relies on one benchmark

We thank the reviewer for this valuable suggestion. We additionally evaluate RouteFM on VL-RouterBench [3], an independent vision-language routing benchmark. Following its official data split, we randomly sample a total of K ∈ {8, 16, 32, 64} behavioral context observations for each test query and compare RouteFM against existing baselines.

MethodK=8K=16K=32K=64
Weighted kNN0.72250.75250.73330.7302
MLP0.76540.78720.71830.7241
Embedding-Ridge0.72270.76000.75190.7373
EmbedLLM0.73180.76710.78720.7830
RM-Softmax0.71850.69010.66310.7157
TRouter0.71850.74790.78720.7183
InferenceDyn.0.74480.78220.78720.7863
UniRoute0.72250.74790.78720.7185
ICL-Router0.72250.71130.70510.6991
Context-Mean0.72250.78720.78720.7872
RouteFM0.77220.78720.79560.7978

RouteFM achieves the best or tied-best performance across all context sizes, further validating its generalization capability on an independent benchmark.

Additionally, as detailed in our response to Reviewer ZyM9, we conduct experiments in two further settings: (1) SEAME Dev MAN, demonstrating transfer to the audio modality, and (2) GSM8K-Strategies, where RouteFM achieves substantial improvements on long-thinking tasks.

Together, these additional evaluations provide broader evidence of RouteFM's generalization across modalities, benchmarks, and task types.

[3] VL-RouterBench: A Benchmark for Vision–Language Model Routing. CVPR 2026.

AuthorsOpenReview ↗

Response to Reviewer 8pP1 — Part (3/3)

R6:Does the multimodal encoder explain the gain?

We agree that the multimodal encoder contributes to query representations, but RouteFM's gains cannot be attributed solely to the encoder.

As shown in Table 2(a), RouteFM with a text-only BGE encoder still outperforms the strongest baseline using Qwen3-VL embeddings across all context sizes, achieving 0.7308 vs. 0.7100 at K=8. Moreover, all baselines except ICL-Router share the same Qwen3-VL embeddings as RouteFM, so the encoder alone cannot explain the performance differences.

EncoderK=8K=16K=32K=64
RouteFM / BGE0.73080.73840.74680.7475
Best Baseline0.71000.73380.74330.7440

Importantly, RouteFM is pretrained entirely on text-only routing data, without any image inputs. The multimodal encoder enables RouteFM to represent multimodal queries, while the transferable routing capability is learned through heterogeneous routing pretraining. This distinction supports our claim that RouteFM learns a reusable routing mechanism that can generalize beyond the modalities seen during routing pretraining.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response to Reviewers' Comments and Summary of Revisions

Dear Area Chairs and Reviewers,

We sincerely thank all reviewers for their careful evaluations and constructive feedback. We have submitted our detailed rebuttal responses and updated the manuscript accordingly. Below, we summarize the strengths recognized by the reviewers, the main concerns and our responses, and the key revisions made to the manuscript.


Strengths Recognized by Reviewers

We appreciate the reviewers' recognition of RouteFM's novel perspective on LLM routing (ZyM9), its research significance and practical value, particularly its flexibility and effectiveness in cold-start scenarios (ZyM9, AEru), and its comprehensive and carefully designed experiments, including extensive baseline comparisons and ablation studies (8pP1).


Main Concerns and Our Responses

We appreciate the reviewers' constructive suggestions, which have helped us strengthen the empirical evidence and clarify our contributions.

  1. Statistical Reliability and Oracle Headroom (Reviewer 8pP1, R2 and R4). We conducted bootstrap uncertainty analysis and evaluated the fraction of oracle headroom captured by each method, confirming generally reliable improvements and strong performance relative to the oracle upper bound.
  2. Additional Benchmarks and Encoder Analysis (Reviewer 8pP1, R5 and R6). We evaluated RouteFM on VL-RouterBench and conducted a text-only encoder ablation, further validating its generalization and demonstrating that its gains cannot be attributed solely to the multimodal encoder.
  3. Generalization to Audio and Reasoning Strategies (Reviewer ZyM9, R1). We conducted additional experiments on SEAME Dev MAN and GSM8K-Strategies, demonstrating RouteFM's ability to generalize to audio routing via transcription-based query representations and to inference-time reasoning strategy selection, without retraining.
  4. Candidate Pool Analysis and Ablation Studies (Reviewer AEru, R1 and R2). We provided detailed candidate pool statistics and training–test overlap analysis, and conducted additional architectural and training-objective ablations to clarify the contributions of RouteFM's key components.

Summary of Manuscript Revisions

We have incorporated the following updates into the revised manuscript:

  1. Appendix B.1–B.2: Added candidate pool statistics, model diversity and training–test overlap analysis, and detailed episodic pretraining procedures.
  2. Section 5.3 and Table 2: Incorporated additional architectural and training-objective ablations and corresponding analyses.

The remaining additional experiments and analyses are presented in our detailed rebuttal responses. We will also incorporate these additional findings and discussions into the final version of the paper to ensure that all reviewers' concerns are fully addressed.


We sincerely appreciate the reviewers' thoughtful feedback and hope that our responses and revisions adequately address their concerns.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.