Published:
Routing replay (R3) stabilizes MoE RL training, but the routing data is 97% of the generation payload. This post traces the bottleneck — the single-threaded manager pipeline, not bandwidth — and the failed ‘obvious’ fix that revealed a fundamental constraint of mixing NCCL with inference.
Published:
Role-playing was the first multi-agent pattern — assign personas, let agents debate or collaborate. But it was largely a product of 2023-2024 model capabilities. As models improve, the real value of multi-agent systems turns out to be structural.
Published:
How LLM agents learn to manage their own context — from harness-driven compaction to memory tools and sub-agents — and why this may be the key bottleneck for long-horizon reasoning.
Published:
A unified treatment of the five sources of distribution mismatch in LLM reinforcement learning and their corrections.
Published:
Pass@k is ubiquitous in evaluating reasoning models, but the metric is more subtle than it appears. Computing it correctly requires the unbiased estimator, and the nonlinearity of pass@k means it effectively upweights hard problems compared to pass@1.
Published:
A training-free approach to step-level credit assignment: estimate V(prefix) via log-probability, compute marginal utility across episodes — plus the implementation pitfalls that silently destroy the signal.
Published:
On-policy distillation integrates teacher guidance into RL training, but the implementation is full of silent failures. This post documents the architecture, pitfalls, and design choices from building OPD in VeRL.
Published:
An empirical investigation into what drives output length growth during RL training, revealing that dataset difficulty composition is the primary driver behind the ‘overthinking’ phenomenon.