Kimi-style agentic reward modeling adds rubric scoring and budgeted verbosity control
stochasticchasm · x · 2026-07-28
The discussion focuses on an Agentic Generative Reward Model (GRM) for non-verifiable general tasks.
- It keeps tournament-style group rewards with binary comparisons, as in Kimi K2.5.
- For better judgment, the agentic judge must follow a fixed protocol: read the output, generate a rubric, score each candidate against that rubric, and record the scores.
- To reduce reward hacking from overly verbose answers, the method adds budget-based verbosity control: an initial verbosity budget is estimated from a cold-start model, then scaled by a multiplier, and outputs that exceed the budget automatically lose the comparison.
The reply adds that the approach is especially interesting for reasoning-effort allocation, notes extensive SFT, and praises the choice to estimate the initial token budget from the post-SFT model and scale it with a multiplier.
Related event: Kimi K3 Reproduces RLVR Research and Introduces Reasoning Budget Control(4 posts)→
More from coding & agent
- Dev builds MCP layer so Claude Code can work on servers without ever holding SSH keys — NETMESAFE · 2026-10-16
- 'Refusing to code with AI in 2026 is like refusing Google in 2005' — prasenx · 2026-10-16
- Claude Code adds idle compaction to save 1h prompt cache on idle sessions — EricBuess · 2026-10-16
- Handling agents that crash mid-task: the hard part is the unknown outcome across the network — Responsible-Dare5139 · 2026-10-16
- OpenAI DevDay Community Hackathon heads to Berlin with developer, founder and GTM tracks — paw_lean · 2026-10-16
- Data scientist career path is shifting: ML skills no longer enough, AI engineering is the new bar — mdancho84 · 2026-10-16