Benchmark report

Evaluating AI Agents
on Alpha Research

Strategy origination and iterative improvement under controlled evaluation budgets.

ABSTRACT

Can a model do the work of a strategy researcher?

A useful research system must do more than propose an investment idea. It must turn the idea into code, test it, interpret a disappointing result and decide what to try next. FAROS evaluates that sequence of work: creating a strategy from scratch and improving an existing one, under the same evaluation budget.

This pilot compares four model–harness systems across independent research trajectories. The recorded histories show models changing hypotheses, abandoning unsuccessful approaches and improving the economics of their strategies. The most informative result is how progress develops: an early lead does not always persist, and a small implementation change can matter more than a more complex prediction model.

These findings come from repeated development backtests, not a live fund or an untouched test set. The opportunity they illustrate is a repeatable, inspectable research process that can be evaluated and improved.

01 / THE EXPERIMENT

What the systems actually do

CORAL is the software that runs the research experiment

CORAL is not a trading model. It provides persistent workspaces, called islands, where a researcher keeps strategy code and research notes. It records submitted versions and evaluation results and manages the research session. The model supplies the hypotheses, edits and next-step decisions. The native coding harness—Codex or Claude Code—provides the tools through which it works.

A separate grader turns the submitted signal into a portfolio and computes performance after the declared trading costs. The researcher cannot award itself a score. This separation makes a proposed improvement testable and leaves a record of what changed.

  1. Form a hypothesisUse market features and prior feedback.
  2. Write the strategyTurn the hypothesis into executable code.
  3. Submit a testThe grader evaluates the frozen version.
  4. Read the resultRecord the outcome and choose the next experiment.

Researchers can explore price trends, volatility, short-selling data and insider activity on a reconstructed equity panel. Public 2021–2022 data supports their analysis. Aggregate feedback on 2023–2024 guides the iterative search. No one supplies a preferred final strategy.

Origination: find an idea

Start without a supplied signal. Build and test strategies. Measure the best net return found, with cash available as the incumbent.

Refinement: improve an idea

Start from the same five-day reversal strategy. Change it using feedback. Measure improvement over that unchanged starting point.

The primary comparison uses sixteen evaluations for origination and sixty-four for refinement, with five origination repetitions per system. Refinement includes every completed 64-evaluation run: GPT-6 Astra / Codex (3), GPT-5.6 Terra / Codex (3), GPT-5.6 Luna / Codex (3), Sonnet 5 / Claude Code (5). Every system faces the same budget within a task. Failed submissions consume that budget. The refinement allowance was increased after evaluation sixteen; the appendix records the schedule and operational history. Models may reason and edit between tests; equal evaluation budgets do not imply equal tokens, time or cost.

The current experiments keep research histories separate. CORAL can support exchanges between islands, but collaboration is not measured in this release. “Self-improvement” here means better strategies and research decisions within a run, not changes to the model’s weights or demonstrated learning across independent runs.

02 / RESULTS

Model comparison

Origination is measured at 16 evaluations; refinement at 64 evaluations. Each system contributes five origination repetitions. Refinement uses all completed runs: GPT-6 Astra / Codex (3), GPT-5.6 Terra / Codex (3), GPT-5.6 Luna / Codex (3), Sonnet 5 / Claude Code (5). All five original Sonnet trajectories have completed 64 evaluations. Additional Sonnet repetitions and the corrected active-budget instructions were authorized after observing the initial confidence intervals; this operational history is documented in the appendix. These endpoints are used consistently in the table and comparison plots. The research curves below retain the same repetitions at every checkpoint.

Scores use fixed public-control references: 100 × net return ÷ 2.8854 points for origination, and 100 × improvement over the common seed ÷ 17.2697 points for refinement. Overall is the equal-weight mean of these two scores. References do not depend on which models participate; scores may exceed 100. Overall summarizes this task-budget combination, not equal compute or equal task difficulty.

SystemOrigination scoreRefinement scoreOverall
GPT-6 Astra / Codex159.87152.43156.15
GPT-5.6 Terra / Codex155.60127.96141.78
GPT-5.6 Luna / Codex147.65118.02132.84
Sonnet 5 / Claude Code130.13126.83128.48

Green marks the highest observed mean. Intervals describe variation between research repetitions on this historical task. They remain wide, particularly for refinement; rankings are descriptive. Scores with confidence intervals ↗

Origination versus refinement
GPT-6 Astra / CodexGPT-5.6 Terra / CodexGPT-5.6 Luna / CodexSonnet 5 / Claude Code
Higher and farther right indicate stronger observed performance. Coordinates use the exact normalized scores in the leaderboard. Confidence intervals are hidden by default, with axes fitted to the means. Enable them to show shaded rectangles spanning the separate 95% intervals for each score; these are not joint 95% confidence regions or probability clouds. Axes expand to include every interval.
03 / RESEARCH FINDINGS

What changed during the research

1. Astra retains its lead as the research budget grows

Across all completed refinement runs, Astra's mean gain over the common seed is 23.20 percentage points at sixteen evaluations and 26.32 at sixty-four. Sonnet reaches 21.90 at sixty-four. Astra has the highest observed mean at sixteen, thirty-two and sixty-four evaluations; uncertainty remains too large for a definitive ordering.

Why it matters: an effective researcher must both find useful improvements and retain them while exploring further. Total performance counts all progress from the same starting strategy, so an early breakthrough remains part of the result.

2. One improvement came from trading the idea less aggressively

In Astra’s first clean refinement run, the researcher had combined price momentum with short-interest information. It then tested a specific hypothesis: daily changes in the combined signal could create costly trading noise. The next submission kept the signal construction and changed only its smoothing weight.

Before

Use 100% of today’s computed signal.

After

Use 10% of today’s signal and retain 90% of the previous smoothed value.

Recorded outcomeBefore · evaluation 10After · evaluation 11
Net profit$93$817
Gross profit$869$1,097
Traded notional$1,037,938$328,907
Declared trading costs$776$280

Traded notional fell by 68%. Net profit rose from $93 to $817 on the fixed $10,000 development book. Both higher gross profit and lower costs contributed. The hypothesis was recorded before the test, and the parent–child code comparison confirms the isolated smoothing change.

Why it matters: the researcher did more than search for a stronger predictor. It tested how an existing signal should be expressed over time. That is a concrete example of research on implementation and trading economics.

3. More complex models were not always the useful next step

In Astra’s first origination run, an early short-interest signal produced $171 net. Three subsequent public-trained tree-model submissions lost money after costs in development. The researcher returned to a simpler momentum signal, then combined it with short-interest, short-volume and insider-purchase information. The final submission reached $451 net.

Why it matters: the useful behavior was not a straight line of successful guesses. The history records an unsuccessful branch of research followed by a change of direction. The best-so-far curve hides that detour; the submitted strategies reveal it.

4. Some breakthroughs arrived late—and independent searches overlapped

In one Terra refinement run, the best result after ten evaluations was still a $1,263 loss. The first profitable submission arrived at evaluation 13, at $111 net, after switching to a signal based on short-interest days-to-cover. Stopping at ten tests would have missed that outcome.

Separately, Astra and Luna independently originated signals favoring stocks with lower days-to-cover, a measure of short-selling positions relative to trading volume. They used different formulas but both recorded $171 net. This suggests overlapping discoveries in this task; it does not prove identical portfolios or a general investment effect.

Why it matters: research needs enough room to change direction, but additional researchers do not automatically produce different ideas. Those observations motivate testing longer horizons and, later, whether sharing findings improves search efficiency.

The code examples are selected after observing the runs. Each describes one adaptive trajectory or comparison; independent 95% confidence intervals are not estimable for these individual changes. Aggregate means and intervals remain the primary model comparison. Full trajectories, hypotheses and source code ↗

04 / RESEARCH PROGRESS

Performance over the evaluation budget

Each curve shows the mean best strategy retained after a given number of evaluations. Where enabled, shading represents pointwise 95% confidence intervals across research repetitions. Best-so-far curves cannot decrease: a plateau may contain many unsuccessful tests. Each model’s refinement curve uses its own fixed cohort of completed runs across the entire 1–64 range; origination uses the same five across 1–16.

Origination: finding a strategy
GPT-6 Astra / CodexGPT-5.6 Terra / CodexGPT-5.6 Luna / CodexSonnet 5 / Claude Code
Five independent repetitions per system. Mean normalized origination score. Confidence intervals can be enabled using the chart control.
Refinement: improving the same starting strategy
-12.232.076.2120.4164.61163264Mean refinement scoreCharged evaluations per researcher

GPT-6 Astra / CodexGPT-5.6 Terra / CodexGPT-5.6 Luna / CodexSonnet 5 / Claude Code
Normalized refinement score, using the same fixed reference as the leaderboard. Each model retains the same completed-run cohort at every checkpoint. The full trajectory is shown by default. Confidence intervals can be enabled using the chart controls.

What happens after an apparent plateau?

Astra's first matched trajectory did not beat its evaluation-29 incumbent during evaluations 30–60. A revised smoothing rule at evaluation 61 raised recorded net profit from $1,000 to $1,024. Terra found a larger late change in the same repetition: adding long-term momentum to its short-volume signal increased net profit from $337 at evaluation 46 to $495 at evaluation 56.

Not every edit adds useful information. Luna changed positive regime multipliers but obtained identical profit because the downstream rankings did not change. These code-verified examples show why longer research needs both exploration and diagnosis. They are selected cases, not estimates of average model effects. Inspect the underlying changes ↗

05 / IMPLICATIONS

From generated ideas to a research process

The compelling part of this pilot is the combination of behaviors: models propose hypotheses, implement them, encounter failures, revise their approach and sometimes improve the economics of an existing idea. The resulting code and research records make that work inspectable. A model comparison can therefore examine how a system researches, not just the profit of its final submission.

The aggregate results also distinguish capabilities. Origination means are closely grouped and their 95% intervals overlap. Refinement shows a wider range of observed means, with Astra highest in observed mean. A single score should not obscure those differences. This release does not report multiplicity-adjusted pairwise significance tests.

The next scientific question is whether these behaviors repeat across new tasks and untouched periods, and whether a better research environment makes them more reliable. CORAL supplies a common environment for measuring that progression. Collaboration, transfer and research-framework comparisons are planned follow-up experiments and are not scored in this release.

What this pilot does not yet establish

All repetitions reuse the same historical equity task. The 64-evaluation refinement comparison includes all completed runs per model. All five original Sonnet trajectories have completed 64 evaluations. Unequal sample sizes are disclosed above; the appendix retains a comparison restricted to repetitions completed by every model. Additional Sonnet repetitions and the corrected active-budget instructions were authorized after observing the initial confidence intervals; this operational history is documented in the appendix. The budget was increased after the initial sixteen evaluations, and operational recovery changed continuation prompts partway through the experiment. The researchers receive repeated development feedback, so better backtests are not evidence of out-of-sample alpha. Reconstructed data, assumed execution costs and variations in native harness versions also limit interpretation. Confidence intervals describe variation between research runs, not future market returns. Fixed public-control normalization stabilizes the scoring scale when models are added; it does not calibrate equal task difficulty.

These limits leave a concrete result to build on: a working environment for comparing autonomous research, with examples of hypotheses, failures and improvements that can be traced to executable strategies.

RESOURCES

Technical appendix and evidence

The evidence archive also preserves earlier and deferred experiments; they are not included in this release’s Overall score.