TopBench (Agentic) - Treatment Effect: leaderboard
TopBench (2026) tests implicit predictive reasoning in table question answering: the answer is not in the table and must be predicted from historical rows. Share of treatment-effect questions whose predicted direction of change matches the reference, from 0 to 1, solved by an agent that runs Python in a sandbox; higher is better.
Metric: Trend accuracy (0 to 1) on TopBench's 105 treatment-effect questions, which ask whether a counterfactual change to a profile raises or lowers its outcome given a historical table, solved by the authors' ReAct agent that writes and runs Python in a Docker sandbox over the table: share of questions whose predicted direction matches the reference; values and decisions extracted from the free-text answer by a DeepSeek-V3.2 judge and checked against the answer text; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 8 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Setting | Score | Overall rank |
|---|---|---|---|---|
| 1 | GPT-5.2 | 0.65 | #190 | |
| 2 | Gemini 3 Flash | 0.65 | #180 | |
| 3 | DeepSeek V3.2 | Thinking | 0.65 | #425 |
| 4 | Claude Sonnet 4.5 | 0.63 | #267 | |
| 5 | Qwen 3 235B A22B 2507 Instruct | 0.53 | #625 | |
| 6 | Qwen 3 235B A22B 2507 (Thinking) | 0.53 | #471 | |
| 7 | TableLLM-13B | 0.17 | ||
| 8 | TableLLM-8B | 0.07 |
No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=topbench-agentic-treatment-effect · How It Works · Data refreshed daily, snapshot 2026-11-22.