AI Paper Atlas

Research map / Large language models / Agents, code and math

Scaling the Scaling Logic: Agentic Meta-Synthesis of Logic Reasoning

Bowen LIU, Zhi Wu, Runquan Xie, Zhanhui Kang, Jia Li

ICML 2026 · Large language models › Agents, code and math · approach: Inductive and abstract reasoning evaluation · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
train on agent-evolved task families instead of seed families (ablation)GRPO on Qwen3-8B-Base · SynLogicpass@1: 13.1 -> 18.7 (step 240)Improved
train on evolved families instead of seed families (ablation)GRPO on Qwen3-8B-Base · AIME25pass@1: 11.0 -> 14.0 (step 240)Improved
evolve Enigmata families with the pipeline at matched instance count and steps (ablation)Enigmata data, Qwen3-8B-Base · BBHpass@1: 61.8 -> 67.3Improved
evolved vs seed data at equal sample count (236 vs 236) (ablation)SSLogic-Evolve · logic and math averagesaverage delta: +0.8 logic, +1.2 mathImproved
add SSLogic-Evolve data to math (AIME) training (ablation)AIME-trained Qwen3-8B-Base · AIME25pass@1: 15.2 -> 22.4 (step 240)Improved
add ARC-AGI data to the math + SSLogic-Evolve mix (ablation)AIME + Evolve training · math benchmarkspass@1 average delta: -3.9 to -5.0 at steps 160-200 (-3.7 at 240)Hurt
train Qwen3-8B (Thinking) on Evolve instead of Seed (ablation)Qwen3-8B (Thinking) · SynLogicpass@1: 49.4 -> 52.1Improved
train Qwen3-8B (Thinking) on Evolve instead of Seed (ablation)Qwen3-8B (Thinking) · math benchmarks (AIME24, Brumo25, HMMT25)pass@1: AIME24 74.3 -> 74.1; Brumo25 71.4 -> 70.9No clear effect
evolve data, token-matched comparison (about 1.09B tokens) (ablation)SSLogic-Evolve vs Seed · Macro Accuracyrelative macro accuracy: +7.3% relativeImproved
evolved data effect on KORBench categories (ablation)Evolve vs Seed · KORBenchpass@1: Logic 32.4 -> 45.6; Puzzle 7.2 -> 8.4Mixed
mixing math with Evolve data raises reflection token rateRL training dynamics · reflection tokens per 100 response tokensfrequency: 17.57 -> 27.80Improved
evolved vs seed data on Qwen3-8B-Base ARC-AGI (ablation)Evolve vs Seed · ARC-AGIpass@1: 3.1 -> 3.1 (step 240)No clear effect

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy