In AI training, only 1-in-5 tokens are real decision points — targeting those few beats any elaborate learning strategy.

This explores why picking the right tokens to act on tends to matter more than the machinery you wrap around adaptation. The most striking version of this: in RLVR training, only about 20% of tokens are high-entropy "forking points" where reasoning actually branches, and training exclusively on those matches or beats updating on everything Do high-entropy tokens drive reasoning model improvements?. The learning signal isn't spread evenly across the sequence — it's concentrated in a minority, and finding that minority is the whole game. Adaptation strategy operates on top of this; if you're adapting the wrong 80%, no amount of clever weight-routing or retrieval scheduling recovers the loss.

The same asymmetry shows up when you look inside reasoning chains. Models internally rank tokens by functional importance — symbolic computation tokens get preserved while grammar and filler get pruned first — and students trained on these pruned chains outperform those trained on frontier-model compressions Which tokens in reasoning chains actually matter most?. Specific tokens like "Wait" and "Therefore" turn out to be mutual-information peaks: suppress them and reasoning collapses, but suppress an equal number of random tokens and nothing happens Do reflection tokens carry more information about correct answers?. Selection is doing the heavy lifting because the information is genuinely sparse — a few load-bearing decision points surrounded by connective tissue.

What you might not expect is how much this reframes problems that *look* like adaptation problems. The exploration-exploitation trade-off, long treated as a fundamental dial to tune, turns out to be an artifact of measuring at the token level — hidden-state analysis shows exploration and exploitation are nearly uncorrelated, so you can enhance both at once Is the exploration-exploitation trade-off actually fundamental?. And a case can be made that the unit of optimization should move up a level entirely: reinforcing *attention distributions* — where the model decides what to look at — outperforms token-level RL on multimodal reasoning, because attention is where the actual decision happens Can optimizing attention patterns improve multimodal RL better than optimizing tokens?.

Adaptation strategies aren't worthless — they're just downstream. Fast-slow training shows forgetting is a *misallocation* problem: route lessons into prompts vs. weights correctly and you get the same performance faster with less catastrophic forgetting Can splitting adaptation into two channels reduce forgetting?. Notice the framing — the win comes from correctly identifying *which channel* each piece of information belongs in, which is a selection problem wearing an adaptation costume. Similarly, a well-calibrated token-probability uncertainty signal beats elaborate multi-call adaptive retrieval at a fraction of the cost, because the model's own self-knowledge about *when* it's uncertain is a better selector than external heuristics Can simple uncertainty estimates beat complex adaptive retrieval?.

The deeper lesson the corpus leaves you with: more reasoning, more tokens, more adaptation machinery all hit a ceiling — extended reasoning even causes models to abandon answers they'd already gotten right Why does more reasoning sometimes make models worse?. What consistently pays off is precision about *where* the signal lives. Selection beats strategy because the signal is sparse and unevenly distributed, and no strategy can amplify signal you failed to locate.