ClueFlow: Learning Long-Horizon Search through Attributed Research-State Transitions
Trains search agents on attributed research-state transitions; the thread's long-horizon search item.
AI Storyline
5 items · 4 sources · 5 days
Operational story trace
Learning Through
Follow in this browser to see new updates on your Live feed.
Current stateDeveloping
Latest change
Nov 3: MDSKILL, an arXiv paper, internalizes skill memory into agent weights through multi-view on-policy self-distillation, complementing RL for self-evolving agents.
Earlier contextThe story so far
Reinforcement Learning is a loose grouping of agentic-RL papers, not one developing story. Since Oct 19 it has gathered research-state search training, memory-skill learning, and a study of when fine-grained credit assignment helps or hurts.
editor-curated · source-linked
Arc
Oct 19Nov 3 · now
SEARCH · Oct 19ClueFlow trains long-horizon search on attributed research-state transitions1 source · show source ▾
MEMORY · Oct 23Lifelong agents learn memory and skills bidirectionally within a bounded context1 source · show source ▾
Lifelong Context Adaptation through Bidirectional Memory-Skill Learning
Lifelong agents learn from growing history within a bounded context by coupling memory and skills.
THE TURN · Oct 26Fine-grained credit assignment can help or hurt multimodal agentsTested on video editing; finer credit is not automatically better.1 source · show source ▾
When Fine-Grained Credit Assignment Helps and Hurts: Multimodal Agentic Reinforcement Learning through Video Editing
Shows fine-grained credit assignment can help or hurt, tested on video editing.
EXPLAIN · Oct 30CausalXRL explains RL decisions through causal graph reasoning1 source · show source ▾
CausalXRL: Explainable Reinforcement Learning through Causal Graph Reasoning
Adds interpretability: explains RL agent behavior with causal graphs.
NOW · Nov 3MDSKILL internalizes skill memory through multi-view self-distillation1 source · show source ▾
MDSKILL: Learning to Internalize Skill Memory through Multi-View Distillation for Efficient Dynamic Agent Self-Evolution
Newest: distills skill memory into the policy via on-policy self-distillation alongside RL.
Day 1 Monday, Oct 19, 2026
ClueFlow: Learning Long-Horizon Search through Attributed Research-State Transitions
Trains search agents on attributed research-state transitions; the thread's long-horizon search item.
Day 2 Friday, Oct 23, 2026
Lifelong Context Adaptation through Bidirectional Memory-Skill Learning
Lifelong agents learn from growing history within a bounded context by coupling memory and skills.
Day 3 Monday, Oct 26, 2026
When Fine-Grained Credit Assignment Helps and Hurts: Multimodal Agentic Reinforcement Learning through Video Editing
Shows fine-grained credit assignment can help or hurt, tested on video editing.
Day 4 Friday, Oct 30, 2026
CausalXRL: Explainable Reinforcement Learning through Causal Graph Reasoning
Adds interpretability: explains RL agent behavior with causal graphs.
Day 5 Tuesday, Nov 3, 2026
MDSKILL: Learning to Internalize Skill Memory through Multi-View Distillation for Efficient Dynamic Agent Self-Evolution
Newest: distills skill memory into the policy via on-policy self-distillation alongside RL.
What to watch — open questions
- Does MDSKILL's internalized skill memory beat prompt-held skill memory on agent benchmarks beyond the paper's own?
- Does ClueFlow's attributed research-state training carry to search agents outside the paper's benchmarks?
- Do the credit-assignment findings from video editing hold on text-only tool-use agents?
- Does bidirectional memory-skill learning stay stable over long interaction histories?
How this thread was built
editor wrote the arc · 5 beatswatcher 1 status change
Storylines are threaded mechanically from the feed: stories that share a distinctive anchor across multiple days and sources. Each item links to its original source. The evidence trace, current state, and open questions are written by the editor routine and refreshed whenever a new beat lands.