New York Site views —Page views —
← All series

Debates in language-model research

Compare competing explanations and examine what the evidence can establish.

Reading order

  1. Reinforcement learning for reasoning: new abilities or more frequent correct samples?
  2. Is a written chain of thought the reason a model answered?
  3. Is longer reasoning better? Evidence on overthinking and early stopping
  4. Process and outcome rewards: what improves when intermediate steps are supervised?
  5. Why do SFT and reinforcement learning differ? Memorization, exploration and generalization
  6. Can synthetic data keep improving models? Quality, coverage and distribution collapse
  7. Does reasoning distillation transfer ability or habits of expression?
  8. Does higher reward mean greater ability? From reward hacking to independent validation
  9. What does preference optimization learn? Helpfulness, style and sycophancy
  10. Can a model teach itself? When self-improvement depends on external feedback
  11. When retrieved evidence conflicts with model memory, which should we trust?
  12. Does a long context provide long-term memory? Reading, retention and updating
  13. Retrieval or parametric memory: where is knowledge easier to correct?
  14. Continual learning and external memory: does reduced forgetting establish learning?
  15. Do graphs help through relational modeling or through extra information and compute?
  16. Does multi-agent collaboration help through specialization or extra compute?
  17. Can self-reflection correct errors? Rewriting an answer versus receiving feedback
  18. Is an agent planning or searching? What a successful trajectory establishes
  19. Does tool learning transfer beyond the call pattern?
  20. Does long-term memory help an agent or preserve its mistakes?
  21. What does a world model learn? Video prediction and causal understanding
  22. Is the model reading the image or guessing from a language prior?
  23. Is multimodal chain of thought faithful? Checking visual evidence in explanations
  24. From watching video to taking action: what is missing from observational learning?
  25. From simulation to reality: evaluation distributions and the cost of failure
  26. Does passing Lean verification establish understanding of a proof?
  27. Two leaderboards, two winners: where did the rankings diverge?
  28. How far can a scaling law extrapolate? Hold out the largest training run
  29. What did a research agent discover? From a reviewed paper to a verifiable point set
  30. Three testable questions from 29 close readings
  31. After AI helps with homework, can you still do it on your own?
  32. Can you recognize AI writing at a glance?
  33. Does the footnote support the sentence?
  34. When a whole class asks AI for ideas, why do good stories start to look alike?
  35. Whose side does AI take when each friend tells the story?
  36. Can my mother still hear my voice in an AI-edited birthday card?
  37. Same question, different wording: why did the AI send me to another door?
  38. What happened to the rain plan? Finding conditions lost in an AI summary
  39. Did the translation keep the promise?