Intro

LLMs being trained with RLVR (Reinforcement Learning from Verifiable Rewards) start off with a 'chain-of-thought' (CoT) in whatever language the LLM was originally trained on. But after a long period of training, the CoT sometimes starts to look very weird; to resemble no human language; or even to grow completely unintelligible.

Why might this happen?

I've seen a lot of speculation about why. But a lot of this speculation narrows too quickly, to just one or two hypotheses. My intent is also to speculate, but more broadly.

Specifically, I want to outline six nonexclusive possible causes for the weird tokens: new better language, spandrels, context refresh, deliberate obfuscation, natural drift, and conflicting shards.

And I also wish to extremely roughly outline ideas for experiments and evidence that could help us distinguish these causes.

I'm sure I'm not enumerating the full space of [...]

---

Outline:

(00:11) Intro

(01:34) 1. New Better Language

(04:06) 2. Spandrels

(06:42) 3. Context Refresh

(10:48) 4. Deliberate Obfuscation

(12:36) 5. Natural Drift

(13:42) 6. Conflicting Shards

(15:24) Conclusion

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

---

Images from the article:

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.