02
Apodex-1.0: Verification as an Intrinsic Team Behavior
How a unified runtime hosts research, coding, and proof workloads while a team structure pushes beyond the reliability ceiling of a single agent.
9 FIELD NOTES
Notes from the moving edge of LLM systems.
Paper-driven field notes on agents, evaluation, and self-improving harnesses. Read for the mechanism, not the leaderboard.
LATEST DOSSIER2026.07
RSI: When the Harness Starts Rewriting Itself
From context search and workflow evolution to joint optimization of weights and harnesses: what does recursive self-improvement actually optimize?
Read note- Context
- Workflow
- Harness
- Weights
Filter by research thread
THE INDEX
↗
03
Rubric-Based Evaluation: Turning Quality into Verifiable Criteria
Why scalar scores are insufficient, and how structured criteria connect evaluation, reward modeling, and inference-time control.
04
QUEST: One Rubric Tree for General Deep-Research Agents
A unified recipe for task synthesis, context compression, and asynchronous RL across models from 2B to 35B.
05
Co-ReAct: Letting Rubrics Collaborate at Every Step
Rubrics move from post-hoc grading to generating criteria, checking actions, and triggering retries at ReAct branch points.
06
MiroThinker H1: Verification after Interactive Scaling
Local checks audit each step while global checks audit the evidence chain, turning more interactions into a more reliable research process.
07
Marco DeepResearch: A Three-Layer Verification Stack for an 8B Agent
Verification at data synthesis, trajectory construction, and test-time scaling blocks error propagation throughout the system.
08
DR Tulu: Evolving Rubrics alongside the Policy
Search-augmented dynamic rules and discriminative filtering keep reward signals useful for open-ended long-form research.
09
MiroThinker: A Third Scaling Axis beyond Model and Context
Interactive depth becomes an independent scaling dimension, exposing how long ReAct trajectories translate tool use into capability.
Content is adapted with attribution from the reference below under CC BY-SA 4.0; page design is original to LLM Compass. zhoujx4/llm-atlas