Trinity Labs

Building reasoning systems for science and engineering. Creators of TrinitySM.

[Trinity Labs @trinitylabs_ai]

Method, evaluation details and limitations in our original announcement:
https://t.co/DUdpLBQfQq

[Trinity Labs @trinitylabs_ai]

How far can 27B-31B models go on olympiad proofs without extra training?

TrinitySM scored 29/42 on IMO 2026, matching the gold cutoff under automated grading, using Gemma 4 31B + Qwen3.6 27B.

Code, data & evaluation: https://t.co/sz10tp7eky

[Trinity Labs @trinitylabs_ai]

How far can small language models go in mathematical reasoning with a well-designed harness, without extra training?

TrinitySM reports 29/42 on IMO 2026 under automated grading.

https://t.co/sz10tp7eky

[Trinity Labs @trinitylabs_ai]

@AiBreakfast Sharing TrinitySM for your research community: a math reasoning harness using Gemma 4 31B + Qwen3.6 27B achieving an IMO 2026 gold.

[Trinity Labs @trinitylabs_ai]

How far can 27B-31B models go on olympiad proofs without extra training?

TrinitySM scored 29/42 on IMO 2026, matching the gold cutoff under automated grading, using Gemma 4 31B + Qwen3.6 27B.

Code, data & evaluation: https://t.co/sz10tp7eky

[Trinity Labs @trinitylabs_ai]

@dair_ai Sharing TrinitySM for your research community: a math reasoning harness using Gemma 4 31B + Qwen3.6 27B achieving an IMO 2026 gold.

[Trinity Labs @trinitylabs_ai]

How far can 27B-31B models go on olympiad proofs without extra training?

TrinitySM scored 29/42 on IMO 2026, matching the gold cutoff under automated grading, using Gemma 4 31B + Qwen3.6 27B.

Code, data & evaluation: https://t.co/sz10tp7eky

[Trinity Labs @trinitylabs_ai]

We'd value feedback on proof quality, evaluation and reproducibility.

Code is available under CC BY-NC 4.0, with third-party notices. Saved evidence can be inspected without a GPU.

https://t.co/sz10tp7eky

[Trinity Labs @trinitylabs_ai]

How far can 27B-31B models go on olympiad proofs without extra training?

TrinitySM scored 29/42 on IMO 2026, matching the gold cutoff under automated grading, using Gemma 4 31B + Qwen3.6 27B.

Code, data & evaluation: https://t.co/sz10tp7eky

[Trinity Labs @trinitylabs_ai]

IMO-ProofBench Selector@1:
Basic (30 problems): 82.38%
Advanced (30 problems): 50.00%

Saved proofs receive two automated grades. The repository includes rubrics, per-proof scores and selection records.

[Trinity Labs @trinitylabs_ai]

The 29/42 is Selector@1: one selected proof per problem. Selection does not use external grades.

[Trinity Labs @trinitylabs_ai]

The idea: give existing models more opportunities to develop and examine a proof.

Four proof lanes, extended reasoning, structured reviews and three refinement passes. Gemma and Qwen play complementary roles. No additional training.