Releases·Model Release
VIDRAFT’s Darwin-398B-JGOS Scores 90.9% on GPQA Diamond
VIDRAFT's Darwin-398B-JGOS Scores 90.9% on GPQA Diamond
0 engagement1 sourceFri, Sep 25, 2026, 11:00 PM
AI Overview
- VIDRAFT's Darwin-398B-JGOS scores 90.9% on GPQA Diamond, per a September 25, 2026 announcement.
Related
Developments·Paper
Study: Prompting and trained effort settings shorten LLM reasoning traces with little accuracy change
A new arXiv paper evaluates methods to shorten LLM reasoning before answering, testing a numeric/concision prompt on Qwen3-14B and trained effort settings on gpt-oss-20b and -120b. Across 198 GPQA Diamond and 500 MMLU-Pro questions, the Qwen prompt shortened reasoning traces by 12-17% with small, mixed accuracy changes at matched token limits. The findings suggest shorter reasoning can reduce cost and latency without major accuracy loss.
A new arXiv paper evaluates methods to shorten LLM reasoning before answering, testing a numeric/concision prompt on Qwen3-14B and trained effort settings on gpt-oss-20b and -120b. Across 198 GPQA Diamond and 500 MMLU-Pro questions, the Qwen prompt shortened reasoning traces by 12-17% with small, mixed accuracy changes at matched token limits. The findings suggest shorter reasoning can reduce cost and latency without major accuracy loss.
0 engagement1 sourcearxivTue, Aug 4, 2026, 09:54 AM
Developments·Paper
Majority-vote self-consistency backfires on GPQA Diamond for small LLMs
A new arXiv study shows that self-consistency via majority vote, a common inference-time technique, reduces per-problem accuracy on a majority of GPQA Diamond questions for two small instruction-tuned models. The effect was pre-registered on a confirmatory split, suggesting the degradation is real for small models on hard science problems.
A new arXiv study shows that self-consistency via majority vote, a common inference-time technique, reduces per-problem accuracy on a majority of GPQA Diamond questions for two small instruction-tuned models. The effect was pre-registered on a confirmatory split, suggesting the degradation is real for small models on hard science problems.
0 engagement1 sourcearxivTue, Aug 11, 2026, 08:08 PM
Releases·Model Release
Qwen3.8-27B Benchmarks
Qwen3.8-27B benchmarks have been released, showing performance metrics for the new model. The results provide insights into its capabilities relative to other models.
Qwen3.8-27B benchmarks have been released, showing performance metrics for the new model. The results provide insights into its capabilities relative to other models.
257 engagement1 sourceredditFri, Aug 14, 2026, 03:02 PM
Developments·Benchmarkfeatured
Qwen 3.8 Max tops Artificial Analysis agentic index, ahead of Opus 5
Qwen 3.8 Max has been ranked as the best overall model by Artificial Analysis's agentic index, surpassing Opus 5. The ranking reflects strong performance in agentic tasks, positioning Qwen as a leading model for AI agents.
Qwen 3.8 Max has been ranked as the best overall model by Artificial Analysis's agentic index, surpassing Opus 5. The ranking reflects strong performance in agentic tasks, positioning Qwen as a leading model for AI agents.
1.5k engagement1 sourceredditThu, Aug 6, 2026, 06:50 PM
Developments·Benchmark
CAIS releases HLE-Diamond, an updated version of the Humanity's Last Exam benchmark
CAIS releases HLE-Diamond, an updated version of Humanity's Last Exam
CAIS releases HLE-Diamond, an updated version of Humanity's Last Exam
453 engagement2 sourcesredditWed, Sep 23, 2026, 05:10 PM