Research notes, essays, and technical overviews on AI, RL, causal inference, and LLMs by Feng Yu.https://meiluosi.github.io/Astroen2026年10月28日 17:16:33https://meiluosi.github.io/en/posts/strandbeest-wind-evidence-and-negative-results/https://meiluosi.github.io/en/posts/strandbeest-wind-evidence-and-negative-results/Two independent physics solvers (MuJoCo and K2), E0–E3 evidence grading, a 30-cell scale study on start wind speed, and our hardest negative results: why pin friction shifts start wind by 2.59x, why kinematic screening anti-correlates with dynamics among top candidates, and why robust optimization flips across solvers.Wed, 28 Oct 2026 00:00:00 GMThttps://meiluosi.github.io/en/posts/llm-training-eval-06-conclusions/https://meiluosi.github.io/en/posts/llm-training-eval-06-conclusions/Connect immutable artifacts, comparable evaluations, failure records, and the next research decision.Fri, 23 Oct 2026 03:05:00 GMThttps://meiluosi.github.io/en/posts/llm-training-eval-05-delivery/https://meiluosi.github.io/en/posts/llm-training-eval-05-delivery/Separate input contracts, weight conversion, generation, parsing, and task-level regression.Fri, 23 Oct 2026 03:04:00 GMThttps://meiluosi.github.io/en/posts/llm-training-eval-04-reproduction/https://meiluosi.github.io/en/posts/llm-training-eval-04-reproduction/Distinguish numerical variation from changed training semantics and test full-state checkpoint restoration.Fri, 23 Oct 2026 03:03:00 GMThttps://meiluosi.github.io/en/posts/llm-training-eval-03-costs/https://meiluosi.github.io/en/posts/llm-training-eval-03-costs/Separate token denominators, approximate FLOPs, measured runtime, memory components, and end-to-end research costs.Fri, 23 Oct 2026 03:02:00 GMThttps://meiluosi.github.io/en/posts/llm-training-eval-02-evidence/https://meiluosi.github.io/en/posts/llm-training-eval-02-evidence/Separate correctness from preference, calibrate judges, and disclose ordering, ties, contamination, and uncertainty.Fri, 23 Oct 2026 03:01:00 GMThttps://meiluosi.github.io/en/posts/llm-training-eval-01-experiments/https://meiluosi.github.io/en/posts/llm-training-eval-01-experiments/Define interventions, budgets, paired repeats, failure rules, and the limits of training conclusions.Fri, 23 Oct 2026 03:00:00 GMThttps://meiluosi.github.io/en/posts/llm-training-tool-06-evaluation/https://meiluosi.github.io/en/posts/llm-training-tool-06-evaluation/Separate candidate coverage from repeated reliability and report costs, constraints, and paired regressions.Fri, 23 Oct 2026 02:05:00 GMThttps://meiluosi.github.io/en/posts/llm-training-tool-05-failures/https://meiluosi.github.io/en/posts/llm-training-tool-05-failures/Separate retries from training, validate repaired actions, and avoid reinforcing errors or hiding selection bias.Fri, 23 Oct 2026 02:04:00 GMThttps://meiluosi.github.io/en/posts/llm-training-tool-04-environments/https://meiluosi.github.io/en/posts/llm-training-tool-04-environments/Detect cross-task cache contamination, separate random sources, and distinguish replay from fresh execution.Fri, 23 Oct 2026 02:03:00 GMThttps://meiluosi.github.io/en/posts/llm-training-tool-03-rewards/https://meiluosi.github.io/en/posts/llm-training-tool-03-rewards/Trace returns through tool interactions and examine how costs, shaping, and retries change the objective.Fri, 23 Oct 2026 02:02:00 GMThttps://meiluosi.github.io/en/posts/llm-training-tool-02-supervision/https://meiluosi.github.io/en/posts/llm-training-tool-02-supervision/Inspect decision prefixes, target masks, repeated supervision, and closed-loop evaluation before training tool-use models.Fri, 23 Oct 2026 02:01:00 GMThttps://meiluosi.github.io/en/posts/llm-training-tool-01-trajectories/https://meiluosi.github.io/en/posts/llm-training-tool-01-trajectories/Separate requested actions, executed results, and final answers in a controlled inventory task.Fri, 23 Oct 2026 02:00:00 GMThttps://meiluosi.github.io/en/posts/llm-training-rl-06-evaluation/https://meiluosi.github.io/en/posts/llm-training-rl-06-evaluation/Separate parameter updates from inference budgets, candidate coverage from selection, and proxy rewards from independent task success.Fri, 23 Oct 2026 01:05:00 GMThttps://meiluosi.github.io/en/posts/llm-training-rl-05-grpo/https://meiluosi.github.io/en/posts/llm-training-rl-05-grpo/Constructed group calculations explain reward normalization, homogeneous groups, sampling cost, and sequence-versus-token aggregation.Fri, 23 Oct 2026 01:04:00 GMThttps://meiluosi.github.io/en/posts/llm-training-rl-04-ppo/https://meiluosi.github.io/en/posts/llm-training-rl-04-ppo/Positive and negative advantage examples explain PPO-Clip, its limits, reference policies, and terminal handling in GAE.Fri, 23 Oct 2026 01:03:00 GMThttps://meiluosi.github.io/en/posts/llm-training-rl-03-policy-gradient/https://meiluosi.github.io/en/posts/llm-training-rl-03-policy-gradient/An enumerated softmax policy explains reward-weighted gradients, baseline conditions, variance, and implementation boundaries.Fri, 23 Oct 2026 01:02:00 GMThttps://meiluosi.github.io/en/posts/llm-training-rl-02-verifiers/https://meiluosi.github.io/en/posts/llm-training-rl-02-verifiers/An executable integer checker illustrates verifier coverage, malformed outputs, incomplete tests, and the difference between composite reward and task success.Fri, 23 Oct 2026 01:01:00 GMThttps://meiluosi.github.io/en/posts/llm-training-rl-01-formulation/https://meiluosi.github.io/en/posts/llm-training-rl-01-formulation/A small integer task separates token actions, terminal rewards, sampling budgets, and parameter updates.Fri, 23 Oct 2026 01:00:00 GMThttps://meiluosi.github.io/en/posts/fly-connectome-01-framework/https://meiluosi.github.io/en/posts/fly-connectome-01-framework/How a real fruit fly mushroom body circuit becomes a fixed learning skeleton, and how five ablation arms isolate wiring from rule.Wed, 21 Oct 2026 00:00:00 GMT