acceptodds
Under review as a conference paper at ICLR 2027

AdaFleet: An Adaptive Orchestrator–Subagent Coevolution Framework for Efficient Quantitative Signal Discovery

Abstract

Self-improving multi-agent systems must balance task execution against changes to the prompts, policies and parameters that guide their work. In continual discovery, this choice is further complicated by a changing acceptance criterion, as each successful discovery alters which future candidates are admissible. We propose AdaFleet, an adaptive orchestrator–subagent coevolution framework that connects reflective prompt revision, dispatch-rule revision and joint fine-tuning through a shared discovery history. A pre-specified feedback rule triggers prompt revision and fine-tuning, while adoption tests compare proposed changes with incumbents. We instantiate the framework in quantitative signal discovery within an evaluation world whose six admission checks account for standalone quality and the growing pool of accepted signals. Under matched evaluator-call budgets, AdaFleet with a 4B backbone admits 6.2 times as many signals as an adapted AlphaSAGE baseline and 3.7 times as many as our re-implementation of NVIDIA's discovery pipeline. The full framework also achieves higher average admissions than adaptive dispatch across five seeds. These findings support coordinated adaptation for continual discovery under changing acceptance constraints.

Primary area:
learning on time series and dynamical systems
Keywords:
Self-improving agents, Meta-agent, Multi-agent systems, Adaptive control, Coevolution, Quantitative signal discovery

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_7SHVOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

AdaFleet is a system of LLM agents used to search quantitative signals for financial prediction. The goal is to find multiple signals that are useful for the prediction and are different from each other. AdaFleet essentially models the problem as a research problem and automates the research via an Orchestrator agent that allocates search and refinement tasks to sub-agents. Results show substantial increase in the number of signals found.

Critical strengths
  • The problem is well-motivated.
  • The Orchestrating system is thorough and is highly effective in maximizing the yield of signals.
Critical weaknesses
  • Most importantly: the main metric in the paper experiments is discovery yield. It indeed looks important to have multiple useful signals that are not too correlated in pairs; but this metric is not sufficient to guarantee that the whole set of signals is strong. The primary objective should be more closely related to the ultimate prediction accuracy.
  • An example to such a direct metric is Sharpe. I had to combine data from Table 2 and Appendix D3 in order to get explicit Sharpe values comparison between the methods, and found that AlphaSAGE beats AdaFleet on this metric.
  • If I understand correctly, the metric only penalizes similarity (a) in pairs, not over the whole set together, and (b) only through correlation which misses non-linear mutual information.
  • The intro is missing some context about the field, the problem and common methods.
Questions for authors
  • Why is discovery yield the primary metric? Can you justify the method via more explicit downstream metrics as well?
  • How would AdaFleet compare with incremental signal selection based on residual prediction?
Reviewer_2BMnOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies continual discovery of trading signals, where every admitted signal tightens the acceptance rule through a pool correlation check. An orchestrator and subagents share one frozen backbone with separate LoRA adapters, and gated prompt, dispatch-rule and adapter updates are fired by a pre-specified trigger. On a WRDS-based US equity world with 3,000 charged evaluator calls, the full system on Qwen3-4B admits 15.9 factors per 1,000 calls against 4.3 for an NVIDIA pipeline re-implementation and 2.6 for AlphaSAGE (Table 1). The authors report that their own pre-specified criterion against adaptive dispatch is not met (Section 6.5).

Critical strengths

The evaluation world is built with care. Point-in-time lags, delisting handling and charging repeats and rejected validations to the budget make it a stricter test than the IC thresholds of earlier alpha mining work. I also respect the reporting. The paper fixes a criterion in advance, lists what was seen before, and reports negative results such as zero adopted prompt revisions.

Critical weaknesses

The evidence for coevolution itself is thin. Most of the gain comes from memory with deduplication (2.1 to 10.0) and adaptive dispatch (12.7) in Table 2. The full configuration beats adaptive dispatch in only 3 of 5 seeds. Prompt revision was adopted 0 of 6 times, joint FF 3 of 7 times and only after call 2,400, and on all three extra backbones the full configuration admits fewer factors than adaptive dispatch (Table 3). Without intervals, the 16% gain looks like seed noise. Novelty is limited too. Joint prompt and weight tuning is BetterTogether (Soylu et al., 2024), the prompt step is GEPA (App. C.2), and AlphaAgent (Tang et al., 2025), which steers LLM alpha search away from an existing factor zoo, is not discussed. Family allocation under a moving target is a non-stationary bandit, so sliding-window UCB (Garivier and Moulines, 2011) would be a natural dispatch baseline.

I also doubt the baseline ratios. Only evaluator calls are matched, while GEPA reflection, LoRA training and replay are free (App. C.5). FactorMiner loses 71% to 87% of calls to cached repeats. AlphaSAGE and NVIDIA's original code see only price and volume fields (App. C.4), while the LLM arms can use Compustat, I/B/E/S and OptionMetrics. FactorMiner is rebuilt without code on a 4B model. So the 3.7×3.7\times to 9.5×9.5\times gaps may mostly measure deduplication and, for AlphaSAGE, data access. Admission count is also a weak target. AlphaSAGE pools reach the 99th OOS percentile while the full configuration sits at 44 in 2024.

Questions for authors
  1. What are GPU hours and tokens per run for the full configuration versus adaptive dispatch, and does the ranking hold under matched compute?
  2. Can you run all arms on the same data fields, for example OHLCV only?
  3. Can you add seeds and a paired test for full configuration versus adaptive dispatch?
  4. Can you report an OOS measure comparable across systems, such as combined pool Sharpe net of costs?
Reviewer_bVvdOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In continual factor discovery, a candidate can pass every standalone quality check and still be rejected because its daily returns track a factor already in the pool, so each admission changes which later candidates can get in. AdaFleet runs an orchestrator and parallel subagents on one frozen backbone, with one LoRA adapter for the orchestrator and a shared one for the subagents, and logs every proposal and outcome in a common history. From this history it can revise the subagents' generation prompt, amend the orchestrator's dispatch rule, or fine-tune both adapters. A fixed state rule decides when prompt revision or fine-tuning is requested, and every proposed change has to pass an adoption test against the incumbent before it replaces it.

The evaluation world uses US equity data and holds every system to the same six admission checks and 3,000 charged evaluator calls. With Qwen3-4B, the full system admits 15.9 factors per 1,000 calls over seeds 0-2, against 12.7 for adaptive dispatch alone, 4.3 for a re-implementation of NVIDIA's pipeline and 1.7 for a rebuilt FactorMiner; AlphaSAGE, which keeps its own GFlowNet generator, admits 2.6. Over five seeds the full system admits 16% more factors than adaptive dispatch and leads in the final 500 calls of every seed. The authors also fixed a success criterion before the main runs completed: a higher admission rate than adaptive dispatch in each of seeds 0-2 and a mean out-of-sample percentile of at least 50 in both 2024 and 2025. It fails on both counts, and the paper says so.

Critical strengths

The evaluation world is the main contribution in my view. Every system faces the same six checks, the same charged-call budget and an admission rule that tightens with each acceptance, and calls spent validating rejected updates stay on the budget. Inside this world, the component study in Table 2 goes from fixed prompts to memory with deduplication, adaptive dispatch, dispatch with prompt revision and the full system, and it shows clearly where the gains come from. Together with the open reporting of the failed criterion and the rejected updates, this decomposition under a matched budget is a publishable result, and the large differences between systems show that the world can tell them apart.

Critical weaknesses

The title and abstract claim more than the runs show. The paper presents orchestrator-subagent coevolution as its contribution, but in the full-configuration runs no prompt revision was adopted, the adopted joint updates came after about 80% of the budget had been spent, and the 16% average over adaptive dispatch comes from one seed and does not appear on the other three backbones. Someone citing the abstract would attribute the result to the wrong mechanism. The defensible contribution is the evaluation world together with memory, deduplication and pool-conditioned dispatch. Coevolution should be presented as a guarded addition whose benefit shows up late in the budget and is not yet stable.

Questions for authors
  1. Would you consider reframing the paper around the evaluation world, memory with deduplication and dispatch, with prompt revision and fine-tuning as secondary mechanisms? If you want to keep the claim that coevolution explains the advantage, the pre-specified comparison against adaptive dispatch should be the primary endpoint, and at the moment it fails.
  2. How much of the gap to the published baselines is due to deduplication alone? For example, what do FactorMiner and the NVIDIA re-implementation admit if repeated proposals are resampled, as in your memory-with-deduplication arm?
  3. Can you report a portfolio result net of transaction costs? Admission counts on gross returns support a claim about search yield, not about tradable signals.
  4. Are more seeds feasible for the other backbones? With one seed each, it is hard to tell whether the update mechanisms just fail to pay for their validation cost there.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.