Agent post-training becomes predictable engineering

Organizations

Updates

1. Jul 17

   SEED turned agents' own trajectories into hindsight skills distilled jointly with outcome-based RL, while a systematic study named Student-Teacher Mismatch and Length Exploitation as on-policy distillation's failure modes and fixed them with advantage clipping and log-scale compression across seven benchmarks.

   SEED self-evolves on-policy distillation · Two failure modes of on-policy distillation

2. May 8

   ScaleLogic found RL training compute scales as a power of reasoning depth (R² > 0.99), with easy-to-hard curricula adding up to 10.66 benchmark points, and Fudan's Balanced Aggregation fixed two GRPO aggregation biases with a parameter-free drop-in.

   ScaleLogic finds a power law for RL compute · Fudan's Balanced Aggregation fixes GRPO bias

3. Apr 19New

   C2 learned rubric reward models from binary preferences alone, letting an 8B model match a 32B rubric-annotated baseline and gain 6.5 points on RM-Bench.

   C2 learns rubric reward models from preferences

Predictions

1. Oct 2026

   Annotation-free reward modeling becomes the default for labs with limited labeling budgets, speeding fine-tuning of specialized models in fields where expert annotators are scarce. Apr 19

2. Nov 2026

   RL compute scaling laws appear as standard methodology sections in frontier-lab technical reports, replacing empirically tuned, undisclosed RL training schedules. May 8

3. Jan 2027

   OPD signal-quality diagnostics become standard components of agent post-training, and OPD paired with outcome-based RL becomes a default in agentic training stacks. Jul 17