The Video Signal

Topic

Reinforcement Learning

All digests tagged Reinforcement Learning

· 41:28

Hugging Face Journal Club: Kimi K3

The discussion summarizes the Kimi K3 tech report, detailing a highly complex and scaled training pipeline for large language models. Key innovations include specialized agentic structures (general, deep research, coding), multi-tier on-policy distillation across nine domain experts, and advanced infrastructure techniques like partial rollout schemes and dynamic resource scheduling. The talk emphasizes that the model's performance is achieved through sophisticated engineering efforts rather than a single breakthrough concept.

Key takeaways

  1. Multi-Tier Expert Specialization 2:00

    The training pipeline involves generating nine specialized domain experts (3 reasoning levels: low, high, max) for three general domains (general tasks, deep research, coding). These are combined using multi-tier on-policy distillation into a single student model.

  2. Partial Rollout Scheme 4:08

    A novel RL technique where the system samples $k$ rollouts from $n$ prompts. It uses an additional parameter $\lambda$ to measure completed rollouts within a budget, updating weights on that subset and pushing incomplete rollouts into subsequent iterations.

  3. Extensible Chat Template 6:42

    The Kimi K3 model utilizes an 'extensible token markup language' for its chat template. This design aims to be easy to learn during SFT while remaining flexible enough to accommodate future modalities without re-engineering the core template.

  4. Advanced Infrastructure Scaling 30:35

    The system employs sophisticated resource management, including a dynamic rollout auto throttling scheduler based on KV cache pressure and a method for collocating RL training and inference by automatically balancing GPU resources between the two tasks.

Watch on YouTube Full article

· 1:17:20

Training Agents 3: Reinforcement Learning

This session introduces Group Relative Policy Optimization (GRPO), a reinforcement learning (RL) method that advances agent training beyond mere imitation (SFT/Distillation). GRPO trains models by sampling multiple completions per prompt and using the group's relative scores—calculated via a reward function—as the primary training signal. This approach is highly effective for complex tasks, allowing agents to learn from their own varied trajectories in an iterative loop.

Key takeaways

  1. GRPO Mechanics 20:30

    GRPO calculates advantages relative to the group average (Reward - Group Average / Group Standard Deviation). This method eliminates the need for a separate value model, reward model, or critic, simplifying the RL loop. The process involves generating multiple rollouts, scoring them with a verifiable Python function (the reward function), and updating the policy based on these relative advantages.

  2. Training Pipeline Progression 5:05

    The training pipeline typically progresses from Supervised Fine-Tuning (SFT) for dense, off-policy signals, to Distillation for richer online rollouts, and finally to RL/GRPO for sparse, on-policy learning. SFT is often used first to bootstrap the model's understanding of the task structure.

  3. Reward Function Design 26:40

    Defining a verifiable reward function is critical; it acts as a 'contract' defining success. Rewards can be composed of multiple components, such as a format check (e.g., ensuring JSON structure) and an accuracy check (e.g., passing unit tests). For agentic tasks, using test suites or compiling code provides robust signals.

  4. Interpreting Training Curves 35:00

    Monitoring training curves via tools like Track.io is essential for debugging. Key metrics include the reward (should rise), entropy (should remain stable/flat, not dive or spike), and completion length. Failure modes—such as 'reward hacking' (high reward but low test accuracy) or 'collapse' (low entropy)—require deep data inspection.

Watch on YouTube Full article