- S-ATLAS
- Verified results
DeepResearch Bench 1 · verified 12 November 2026
S-ATLAS on DeepResearch Bench 1
The official results for the 100 reports S-ATLAS submitted, verified by the benchmark's own team.
100 PhD-level research tasks in 22 fields
The verified numbers
Official overall score
53.15
Verified by the DeepResearch Bench team. By this score S-ATLAS places 11th of 18 on the leaderboard.
Citation accuracy
73.56%
Measured by the DeepResearch Bench team with the benchmark's own FACT method.
Effective citations per report
160.81
Counted by the DeepResearch Bench team with the benchmark's own FACT method, across the reports we submitted.
Score breakdown
The overall score has four parts. Each one is the team's verified figure.
| Measure | Score |
|---|---|
| Comprehensiveness | 52.83 |
| Insight | 53.94 |
| Instruction following | 53.58 |
| Readability | 51.04 |
| Overall | 53.15 |
| Citation accuracy | 73.56% |
| Effective citations per report | 160.81 |
How these were measured: the overall score, the citation accuracy and the effective citations are the DeepResearch Bench team's verified results for the 100 reports we submitted. The team cleaned the reports with the benchmark's cleaning step, averaged three scoring runs, and checked citations with the benchmark's FACT method. The place, 11th of 18, is our verified overall score set against the 17 systems listed on the public leaderboard on 12 November 2026; the benchmark's team is adding our entry.
From first line to verified score
How it got here
- July 2026Hrithikesh Thulluri starts building S-ATLAS, one stage at a time.
- 5 November 2026The 100 S-ATLAS reports are submitted to the DeepResearch Bench team.
- 12 November 2026The team verifies the results: it cleans the reports, averages three scoring runs and checks the citations with the FACT method.
- NextThe benchmark's team is adding the S-ATLAS entry to the public leaderboard.
The benchmark
What DeepResearch Bench measures
DeepResearch Bench is a public benchmark for research AI: 100 PhD-level research tasks across 22 fields, half in English and half in Chinese. Reports are scored by an AI judge against expert reference reports, and their citations are checked separately.
The overall score has four parts: comprehensiveness, insight, instruction following and readability. The citation check, called FACT, measures two things: citation accuracy, and the number of effective citations a report carries.
- References that resolve
- All 4,649 references in the 100 submitted reports resolve to their sources
- Best single report
- 59.79, in our own run of the same evaluation before we submitted
- Languages
- English and Chinese, for both the question and the report
- Time per report
- About 15 minutes for one pass, on average, in the final run
- Engine version
- S-ATLAS Pro, version one
Questions
About the score, the place and how they were measured.
What did the DeepResearch Bench team verify?
The overall score of 53.15, the citation accuracy of 73.56% and the 160.81 effective citations per report, for the 100 reports S-ATLAS submitted. The team cleaned the reports with the benchmark's cleaning step, averaged three scoring runs, and checked citations with the benchmark's FACT method.
Where does 11th of 18 come from?
It is the verified overall score set against the 17 systems listed on the public leaderboard on 12 November 2026. The benchmark's team is adding the S-ATLAS entry.
Which numbers come from S-ATLAS's own run?
The best-single-report figure of 59.79 and the per-report numbers on the home page. They come from our own run of the same evaluation before we submitted, graded with the same judge model and the same official prompts the leaderboard uses.
What is DeepResearch Bench?
A public benchmark for research AI: 100 PhD-level research tasks across 22 fields, half in English and half in Chinese. Reports are scored by an AI judge against expert reference reports, and their citations are checked separately.
Who built S-ATLAS?
Hrithikesh Thulluri, 19, the AI architect of S-ATLAS and founder of S-IDEON. He is the first independent Indian student to enter a deep research engine on DeepResearch Bench 1.
What comes next?
S-ATLAS Max, version two, is in development. It is built to compete with the top five systems on the leaderboard and to raise citation accuracy toward 90%.
India's first independent student entry on DeepResearch Bench
Built by a 19-year-old Indian student
Hrithikesh Thulluri, the AI architect of S-ATLAS, built it in five months on $800 of his own money and a single small computer.