[D] ran controlled experiments on meta's COCONUT and found the "latent reasoning" is mostly just good training. the recycled hidden states actually hurt generalization

A viral analysis by independent researcher bmarti44 challenges the core claims of Meta's COCONUT method for AI reasoning. COCONUT (Hao et al., 2024) proposed that language models could perform "latent reasoning" by recycling their internal hidden states across processing steps, achieving ~97% accuracy on the ProsQA benchmark compared to ~77% for standard chain-of-thought prompting. The new research, however, suggests this performance boost comes not from the novel architecture but from the sophisticated multi-stage curriculum training used alongside it.

bmarti44 designed a factorial experiment to isolate the effects, training four GPT-2 124M models. The key finding was that a model (M3) using the same curriculum but with fixed learned embeddings—not recycled states—performed nearly identically (96.6% vs. 97.0%). Statistically, there was no significant difference (McNemar p = 0.845). More critically, on out-of-distribution tasks like 7-hop chains (trained on 3-6 hops), the model with recycled content (M2) was beaten by over 10 percentage points. The recycled states were found to actively hurt generalization and create overconfidence on unfamiliar inputs.

The dissection provides converging evidence from corruption analysis and linear probing, with all data and checkpoints publicly released. While limited by scale (GPT-2) and a single task, the work raises fundamental questions about attributing performance gains in AI research. It underscores the importance of rigorous controlled experiments to separate architectural innovation from training data effects, a crucial consideration for developers building on such methods.