Alex Wang pictures a reasoning model’s answer as one possible path through words from a question to a result. Training can make paths that reach an accepted answer more likely. It does not guarantee that each additional step in one answer gets closer to the truth.
- Trying several independent answers increases the chance that at least one succeeds, if a reliable checker can recognize it. Forcing a single answer to run longer has no equivalent guarantee.
- When training rewards only the final answer, a correct answer does not establish that the intermediate explanation is faithful.
- A checker that accepts the wrong answer gives the model a shortcut to learn. More searching can make that failure easier to find.
Wang derives these claims under simplified, binary-reward assumptions and cites prior tests of forced longer reasoning. This is a conceptual account of model training, not a new measurement of deployed systems. For coding agents, it is a reminder to check the work and the tests against the actual request.
The 70-comment thread on Hacker News adds a practitioner’s counterpoint about when extra reasoning is worth its cost.
What the thread adds
- gertlabs — reports that in their own training work, “you don’t need to enable reasoning on every single rollout — you get diminishing returns past a certain point.” They describe turning it off for some easier training problems. Wang did not test this approach.
- robrenaud — answers a concern that long text alone cannot ground a model: “The model isn’t grounded, but it can be wrapped in a system that is.”
From a 70-comment thread, in HN’s own order.