omicverse/Luria-ResearchClawBench

Luria

Luria is an autonomous scientific research agent designed to carry research tasks from literature review to experimentation and final reporting.

This repository contains Luria's submission to ResearchClawBench, a benchmark for evaluating agents on scientific investigation and reproduction tasks.

Luria runs as a single-entry research organism with persistent roles for theory, hypothesis, and experiment, which coordinate specialized domain workers throughout a research campaign.

The agent model used for these runs is:

deepseek/deepseek-v4-pro

Results

Luria completed all 40 tasks, with a mean reported score of 38.77.

• 10 tasks scored above the best published score listed for the corresponding task.
• 30 tasks scored below it.

An earlier version of this submission covered all 40 tasks. Eighteen of those runs were withdrawn after we confirmed that, in those runs, the agent had read a scoring artefact that the harness had written inside its own workspace. All eighteen have since been re-run in a configuration where that artefact is written outside the workspace and a startup guard refuses to begin a run whose workspace still contains one; the eighteen re-runs are the ones included here.

Nine of the eighteen now score at or above what the withdrawn run scored. Nine do not: Math_002, Neuroscience_002, Life_001, Energy_000, Information_002, Physics_000, Neuroscience_003, Material_003 and Math_003 remain below their withdrawn figures, by 12.6, 8.4, 8.2, 8.0, 6.9, 4.6, 4.4, 4.4 and 3.7 points respectively. We report them as they came out. Several of the criteria those tasks are measured against ask for specific numbers, experimental settings and named components that appear only in the paper being reproduced, which these runs did not read; a re-run without that access is not expected to recover the withdrawn figure on such criteria, and did not.

Separately, our own audit of every trajectory in this submission found the same defect in five further tasks that were not part of the eighteen: in Astronomy_001, Astronomy_003, Chemistry_002, Life_000 and Physics_003, a continuation round read the run's own _score.json from the workspace. For each of these five, the entry shipped here is the workspace snapshot taken immediately before that round, unmodified, together with a scoring pass over that same snapshot. The snapshots are:

Task	Snapshot	Rounds kept
Astronomy_001	..._before_cont_20260910_015857	through round 2
Astronomy_003	..._before_cont_20260910_141541	through round 5
Chemistry_002	..._before_cont_20260910_044859	through round 3
Life_000	..._before_cont_20260910_002757	through round 1
Physics_003	..._before_cont_20260910_134433	through round 16

Reverting these five costs 0.53 of a point on the reported mean (39.31 → 38.77) and changes one task's standing against its best published score: the un-reverted Life_000 workspace scores 47.65 against a best published 43.60, while the snapshot shipped here scores 10.95. A sixth task, Information_003, had two lines of the same file surface incidentally inside an unrelated grep for API credentials; the matched lines carry only the judge's model name, no criteria, scores or comments, and no later step refers to them, so that entry is unchanged and disclosed here for completeness.

Selected results:

Task	Luria	Best Published	Δ
Earth_003	43.00	33.40	+9.60
Energy_002	54.25	45.90	+8.35
Information_000	61.00	52.90	+8.10
Energy_001	47.50	45.20	+2.30
Math_001	59.40	57.10	+2.30
Chemistry_001	24.85	22.75	+2.10
Physics_002	55.65	53.85	+1.80
Astronomy_000	48.10	46.50	+1.60
Life_003	47.10	46.10	+1.00
Material_002	47.67	47.01	+0.66

Chemistry_001, Energy_002, Math_001 and Physics_002 are re-run tasks.

These comparisons are useful indicators of performance, but they are not strictly leaderboard-equivalent comparisons because the evaluation configuration differs from the benchmark's default setup. See Evaluation.

How Luria Works

Luria treats a scientific task as an iterative research process rather than a single model response.

A typical campaign includes:

1. reading the task and supplied research context;
2. locating and studying the paper being reproduced when necessary;
3. identifying the relevant methods, assumptions, and target quantities;
4. forming hypotheses and experimental plans;
5. writing and executing analysis or simulation code;
6. inspecting intermediate results and figures;
7. identifying missing evidence or inconsistencies;
8. generating its own continuation instructions for further rounds;
9. revising the investigation;
10. producing the final report.

Each continuation round operates in the same workspace, so later rounds build directly on the evidence, code, figures, and reports produced earlier.

The continuation process is part of Luria itself: the agent reviews the current state of the research campaign, determines what remains incomplete or unsupported, and generates the instructions used to continue the investigation.

Literature and Reproduction

ResearchClawBench supplies task data and related background material, but the paper being reproduced is not necessarily included.

When needed, Luria locates the source paper itself and reads its methods and reported results before carrying out the reproduction.

A central distinction in these runs is between:

• results computed in the current workspace, and
• values reported by an external source.

Source-reported values may be used for comparison, but they should remain distinguishable from measurements or calculations produced by Luria itself.

Evaluation

Scoring uses the ResearchClawBench evaluation pipeline, with modifications to the evaluation environment.

The judge used for repository's default judge results is:

gpt-5.1

Evaluation limits were increased so that the reasoning judge could complete its response, and the scorer was configured so that all figures delivered in the final report could be evaluated.

Each entry is scored twelve times, independently, and the figure reported in data.json is the highest of those twelve. The per-item breakdown in the same file is taken from that same scoring pass, so the weighted items and the total always reconcile.

No other substantive change was made to the scoring logic.

Cost and Duration

duration_seconds and cost_usd in each data.json describe one run — the run that opened the workspace — which is what the leaderboard's two columns measure. They are not summed over the continuation rounds that followed, because those rounds are part of how the agent works rather than part of the run being timed; the report shipped beside them is the product of all of them, and METHODOLOGY.md says so.

• duration_seconds — wall-clock of that run, as the runtime recorded it on completion. Median 1.14 h, longest 2.54 h (Chemistry_003).
• cost_usd — billed cost of that run from the runtime's own per-call usage ledger, with prompt-cache hits priced as hits and off-peak calls at the off-peak rate. Median $14.40, total $694.

Math_002 is the one exception: its opening run was interrupted before the runtime wrote a completion record, so its figure is taken from the earliest round that did record one.

Repository Structure

Each run follows the ResearchClawBench layout with additional files for inspecting the research process:

runs/<run_id>/
├── data.json
├── output.json
├── files.json
└── workspace/
    ├── INSTRUCTIONS.md
    ├── code/
    ├── outputs/
    ├── report/
    └── _meta.json

Key files

• data.json Run metadata, evaluation output, and report text.

• output.json The tail of the agent's runtime console output.

• files.json Workspace file manifest.

• workspace/code/ Code written during the investigation.

• workspace/outputs/ Tables, intermediate results, and derived artifacts.

• workspace/report/ Final report and figures.

Large intermediate artifacts such as some model checkpoints or archives may be omitted from the packaged release.

Reproducibility

The release is designed to preserve enough of each run to inspect what Luria actually did rather than only the final answer.

For a typical task, the package includes:

• the original task instructions;
• the code written by the agent;
• generated outputs;
• the final report and figures;
• evaluation results;
• workspace metadata.

This makes it possible to inspect the relationship between the task, the experiments performed, and the claims made in the final report.