MultiSessionCollab - MATH-Hard: leaderboard
MultiSessionCollab (Mehri et al., 2026) task success on MATH-Hard problems: an agent helps simulated users with personas and interaction preferences over 20 sessions each, and a user's hidden draft answer improves only when the help respects those preferences. Percentage accuracy of the final draft answer over 100 simulated users, higher is better; agents run without cross-session memory, with Llama-3.3-70B-Instruct as the user simulator.
Metric: Task success (%; accuracy of the simulated user's final draft answer, MATH-Hard problems, 100 simulated users x 20 sessions, agent without memory). Source: arxiv.org. Saturation forecast: Estimated already saturated. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-OSS-20B | 41.45 | #1150 |
| 2 | Qwen 2.5 7B Instruct | 26.5 | #2583 |
| 3 | Llama 3.3 70B Instruct | 25.81 | #1353 |
| 4 | Llama 3.1 8B Instruct | 11.5 | #4277 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=multisessioncollab-math-hard · How It Works · Data refreshed daily, snapshot 2026-11-12.