abstract-cot

Date kicked off: May 12, 2026

Reproduction of Thinking Without Words: Efficient Latent Reasoning with Abstract Chain-of-Thought (Ramji, Naseem, Astudillo, IBM Research AI, arXiv:2604.22709) on fortyfive HPC.

Start with PLAN.md — that's the comprehensive plan. This README is just orientation.

TL;DR

Train a small instruction-tuned LM (Granite 4.0 Micro 3B → Qwen3-4B → Qwen3-8B) to replace its verbal chain-of-thought with a short sequence of reserved tokens (<TOKEN_A> … <TOKEN_BL>), getting ~10× token efficiency at comparable accuracy on MATH-500, AlpacaEval, and HotpotQA.

Three training phases:

1. Bottlenecked SFT with a block-structured attention mask (answer attends to abstract tokens, not the verbal CoT).
2. Self-distillation from on-policy abstract sequences.
3. GRPO with constrained decoding and an external generative reward (gpt-oss-20b via MindRouter).

Phases 1+2 iterate T=3 times (policy iteration warm-up); then Phase 3 (RL).

Status

Pre-Phase-0. Folder is seeded with PLAN.md and skeleton directories. No code written, no jobs run.

Quickstart (eventually — once Phase 0 is done)

# On fortyfive login node:
git pull
bash scripts/setup_env.sh           # builds ~/venvs/abscot
bash scripts/prepare_data.sh        # downloads HF models + datasets
sbatch slurm/submit_warmup.slurm    # ~3 days on cmci-gpu-8
sbatch slurm/submit_rl.slurm        # ~3-5 days on cmci-gpu-8
sbatch slurm/submit_eval.slurm

Key references

• Paper: notes/paper-v2.pdf (arXiv:2604.22709v2)
• Datasets: allenai/Dolci-Think-SFT-7B, allenai/Dolci-Think-RL-7B
• Reward model: openai/gpt-oss-20b via MindRouter
• Base models: ibm-granite/granite-4.0-micro, Qwen/Qwen3-4B, Qwen/Qwen3-8B