RSA (Recursive Self-Aggregation) = Sequential refinement 🔄 + Parallel exploration ⚡ → Unified into a hybrid evolutionary loop for deeper reasoning. 📄 Paper + website: rsa-llm.github.io/ 🧵 Details in the thread

RSA is simple 🚀 1️⃣ Generate a population of N reasoning chains in parallel 2️⃣ Subsample into N subsets of K chains 3️⃣ Prompt the model to aggregate each subset → new improved population of CoTs 4️⃣ Repeat for T loops That’s the whole algorithm: Recursive Self-Aggregation

RSA scales sequentially & in parallel: more steps T or larger aggregation K jointly with population N → better performance! 🔥 Gains: AIME-25 47 → 73% HMMT-25 29 → 50% Reasoning Gym Games 55 → 70% LiveCodeBench 49.5 → 56.3% No verifiers ❌ No prompt opt ❌ No RL yet!

RSA boosts performance across all models and all reasoning tasks. Tested on Qwen, Nemo, GPT-OSS — thinking & non-thinking, MoE & dense, full-attention or hybrid SSM. Benchmarks: AIME, HMMT LiveCodeBench, Reasoning Gym 📈 Gains are consistent and significant throughout.

RL makes RSA even stronger 📈 Naive RL can hurt aggregation, but aggregation-aware RL ✅ 👉 Generate K responses 👉 Create aggregation prompts 👉 Train the model to aggregate Boosts performance & generalizes to new tasks — RSA is worth it!

RSA beats all simple budget-matched test-time scaling baselines! Shoutout to concurrent work by @wzhao_nlp & team for AggLM (RL-trains single-step aggregators). We independently discovered this, but also find that sequential aggregation with larger population is key for scaling.