Skip to content

Repository files navigation

reasoning-cliff

CI License: MIT Python

AI "reasoning" models collapse beyond a complexity threshold — and paradoxically think LESS as problems get harder. Apple ML Research (Shojaee et al., 2506.06941) called it "the illusion of thinking." Three subsequent papers showed the original experiments had confounds: token limits, impossible puzzle instances, and evaluation frameworks that couldn't tell "ran out of space" from "genuinely failed to reason."

reasoning-cliff is a reproducible harness that runs the same experiments with all three confounds controlled: multiple token budgets, solvability-verified puzzle instances, and a four-way failure classifier. The cliff is real — it just sits further right than the original paper found, and it's smaller. Here's the map.

What it provides

  • Deterministic puzzle generators and verifiers (Hanoi, River Crossing)
  • Multi-budget token control (DEFAULT, DOUBLE, UNCAPPED)
  • Four-way failure typing (CLEAN, TOKEN_STARVED, INTENTIONAL_TRUNCATION, GENUINE_COLLAPSE)
  • DuckDB experiment persistence
  • Publication-style matplotlib charts for collapse and effort curves
  • CLI for estimate/run/plot/demo workflows

Install

python3 -m venv .venv
source .venv/bin/activate
pip install -e .[dev]

Quick start

reasoning-cliff estimate --config config/experiment_config.json --models claude-sonnet-4-6,gpt-4o
reasoning-cliff run --config config/experiment_config.json --models claude-sonnet-4-6,gpt-4o --puzzles hanoi,river_crossing --complexity-range 2:8 --token-budgets DEFAULT,DOUBLE,UNCAPPED --trials 3 --output-db report/results.duckdb
reasoning-cliff plot --db report/results.duckdb --output-dir outputs/figures

Methodology note

This repository explicitly addresses the Shojaee/Lawsen/Rethinking dispute by:

  • Running controlled token-budget sweeps for each (model, puzzle, complexity) cell.
  • Excluding mathematically unsolvable River Crossing instances from evaluation.
  • Separating token-starved failures from genuine collapse failures.

See docs/CONTROVERSY_NOTE.md.

Open source and governance

Not affiliated with Apple, Apple ML Research, or any replication authors.

Releases

Packages

Contributors

Languages