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