- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 3(+1)
Read review
Paper summary
EEVEE adapts black-box language models to heterogeneous tasks without changing their weights. It maintains several specialized prompts and optimizes a prompt-based router to select the appropriate prompt for each input. The router and prompts are co-evolved using correctness feedback. Experiments on four benchmarks and two models show improvements over GEPA and ACE, with additional analyses of transfer, online adaptation, and computational cost.
Critical strengths
-
Co-evolution addresses the dependency between routing and prompt learning. Alternating updates allow input assignments and specialized prompt capabilities to develop together.
-
Coverage-based initialization and Pareto-front selection preserve prompts with different strengths. This accommodates heterogeneous tasks while reducing the need to combine potentially conflicting instructions in one prompt.
-
EEVEE improves average performance over GEPA and ACE across four benchmarks and two models. Routing and co-evolution ablations further support the contributions of its core components.
Critical weaknesses
-
Requires task-specific data and feedback. Adapting to new tasks depends on representative examples and reliable correctness feedback, which may be costly or unavailable, especially for open-ended tasks.
-
Fixed prompt capacity. EEVEE refines a fixed set of prompts without automatically adding specialists for new task types, potentially limiting its flexibility as task diversity grows.
-
Early routing errors may reinforce poor specialization. Prompts learn from their currently assigned examples, so unsuitable initial assignments can shape subsequent capabilities and reinforce the same partition.
-
Online updates remain challenging. This paper does not clearly explain when to update and how to maintain data during online learning.
Questions for authors
Please refer to the Weaknesses section.
5 comments
Response to Reviewer kQyX (2/2)
W3. Early routing errors
Thanks for the insightful comment. EEVEE does not lock in early assignments: each co-evolution cycle re-routes all adaptation data, and router evolution learns from cases where the assigned slot fails but another slot succeeds. To verify this, we additionally start EEVEE from a deliberately bad partition, where each slot receives an equal share of all four datasets and its prompt is first evolved on this partition until convergence. All other settings were kept unchanged, and we report the average over the same 3 runs as the main experiments.
| Setting | GPQA | Formula | TheoremQA | HumanEval | Avg. |
|---|---|---|---|---|---|
| No adaptation | 56.00 | 45.22 | 14.79 | 49.46 | 41.37 |
| No Co-evolution | 39.84 | 47.39 | 24.81 | 59.49 | 42.88 |
| EEVEE, bad partition | 55.56 | 53.17 | 22.81 | 70.33 | 50.47 ± 2.06 |
| EEVEE, original | 54.55 | 54.55 | 25.27 | 72.63 | 51.75 ± 1.62 |
Starting from a bad partition, EEVEE reaches 50.47, close to the original 51.75 and clearly above No Co-evolution (42.88). This shows that co-evolution remains effective even with poor initial assignments. We will add these results to the revised paper.
W4. Online update strategy
Thanks for the comment. In our online setting (Appendix B.4), the stream is processed window by window: EEVEE answers all queries in a window with its current router and prompts, and then updates once before the next window. For data maintenance, we compare two strategies: cumulative, which updates on all observed examples, and new-window, which updates only on the latest window. Both continue from the previous router and prompts, and a new prompt is accepted only if it outperforms the empty prompt on validation data. We additionally evaluate the unadapted model on the same stream to verify the actual improvement of EEVEE.
| Method | GPQA | Formula | TheoremQA | HumanEval | Avg. |
|---|---|---|---|---|---|
| No adaptation | 57.74 | 43.07 | 15.92 | 42.48 | 39.80 |
| ACE online | 29.12 | 42.23 | 16.92 | 28.86 | 29.28 |
| EEVEE cumulative | 51.01 | 50.67 | 24.71 | 47.97 | 43.59 |
| EEVEE new-window | 50.00 | 52.95 | 23.44 | 49.49 | 43.97 |
Both strategies improve over the unadapted model by 3.79 and 4.17 points, respectively, while ACE online falls 10.52 points below it. We will add a detailed description of the online procedure to the revised paper.
Follow-up on our responses
Dear Reviewer kQyX,
Thank you again for your time and constructive feedback. In our responses, we added experiments on 8-task mixtures with only 4 prompt slots, on starting EEVEE from a deliberately bad partition, and on the ALFWorld agent benchmark with verifier-based feedback. As the discussion period is coming to a close, we would like to check whether our responses have addressed your concerns. We would be happy to provide any further clarification.
Best regards,
The Authors
Response to Reviewer kQyX (1/2)
We thank reviewer kQyX for recognizing our co-evolution design, the diverse prompt selection, and the consistent improvements over GEPA and ACE.
W1. Requirement of task-specific data and feedback
Thanks for the comment. Learning from task examples with correctness feedback is adopted by most prior work in prompt optimization and context adaptation, such as OPRO (Yang et al., 2024), TextGrad (Yuksekgonul et al., 2024), GEPA (Agrawal et al., 2026) and ACE (Zhang et al., 2026). EEVEE follows this setting and does not require additional supervision such as dataset labels. It also requires only a small amount of data: with 200 adaptation examples (50 per benchmark), EEVEE still reaches 48.59, compared with 41.37 for the unadapted model (Appendix B.2).
We also study the case where feedback is sparse or noisy (Appendix B.3), where Sparse-k keeps feedback on only k% of the examples and Noisy-k flips k% of the feedback.
EEVEE retains more than 85% of its clean-feedback performance in all conditions and still improves over the unadapted model.
Moreover, many tasks, such as common agent tasks, do not have reference answers but provide a verifier of task success. We additionally evaluate EEVEE in this setting on ALFWorld (Shridhar et al., 2021), a text-based household environment where the agent completes a goal by issuing text actions (e.g., "go to countertop 1", "take apple 1") for up to 50 steps, and the only feedback is whether the goal is completed. We use Qwen3-4B-Instruct as in our main experiments, with an adaptation pool of 40 episodes per task type and 4 prompt slots for four task types: Look (find an object and examine it under a desk lamp), Clean (clean an object in the sink and place it), Heat (heat an object with the microwave and place it) and Cool (cool an object with the fridge and place it).
EEVEE improves the average success rate from 15.43 to 40.25, showing that it also works with verifier-based feedback. We will clarify the feedback requirements and add these results to the revised paper.
W2. Fixed prompt capacity
Thanks for the comment. EEVEE is not limited by a fixed number of slots as the number of tasks grows, since the number of slots bounds the number of specialized behaviors rather than task types, and a new task can reuse an existing slot when it shares behaviors with other tasks. To verify this, we additionally scale the mixture to 8 tasks by adding IFBench (Pyatkin et al., 2025), CRUXEval (Gu et al., 2024), FinQA (Chen et al., 2021), and BBH (Suzgun et al., 2023) or LegalBench (Guha et al., 2023), while keeping only 4 slots.
With only 4 slots, EEVEE improves the 8-task average by 7.80 and 5.12 points, with large gains on Formula and CRUXEval. We also compare with 8 slots, one per task, on the same mixtures.
The gains are similar with 4 and 8 slots, showing that EEVEE maintains strong performance over a wide range of task numbers without a clear drop as the number of tasks grows. We believe that 8 tasks cover a substantial range, since prompt optimization typically targets a specific deployment workload rather than a general-purpose model. We will add these results to the revised paper.