Research papers, read to you in five minutes.
Research papers, read to you in five minutes.

FDE-Gym tests whether agents deliver systems that actually move a customer's business KPI, not just satisfy the written request. The best model scores 0.56 and fully resolves 14 of 95 tasks; the real failure isn't vague problems but hidden rules, where leading models drop from 0.73 to 0.39 once constraints go unstated.

Agents · Added Oct 16

More Methods, Better Agents? A Reality Check on Self-Distillation in Agentic Reinforcement Learning

Benchmarking ten self-distillation methods for RL agent training against plain GRPO across 132 runs and 12K H800-hours: only two beat baseline on average, win-counts hide full collapses (one method scored 0% on WebShop), and matched-compute comparisons erase even the winners' edges.

LLM Training · Added Oct 16

Beyond Domain-Routed Supervision: Cross-Teacher Guarding for Multi-Teacher On-Policy Distillation

Multi-teacher distillation into one small student quietly destroys medical accuracy when combined with math or code, because routed updates still collide on shared weights. Road MOPD scores each trajectory with every teacher, reweights tokens on agreement and projects away conflicting gradients, recovering 5.6–8.8 points without hurting the other domains.

Agents · Added Oct 16

Player-1: Learning to Play from Synthetic Experience

Screen-reading agents choke on games: no cursor, held inputs, world moves on its own. Player-1 skips per-game demos — SFT on 6k trajectories, then PPO across 16 synthetic mini-games with programmatic rewards — lifting Minecraft success 10%→38% and transferring to unseen Doom and Watch Dogs 2.

Agents · Added Oct 16

Doc2Harness: Benchmarking LLMs on Long-Horizon Harness Engineering of Coding Agents

Coding benchmarks score models on fixing issues inside mature repos, but shipped agents are model-plus-harness, and nobody tests whether a model can build the harness itself. Doc2Harness does: spec plus skeleton, scripted mock-model playback, positive controls against vacuous passes. Top model fails 47% of 328 cases.

Never fall behind the literature again.

Free account. Follow topics, build your queue.

Sign up free

Agents · Added Oct 17

FDE-Gym: The Last Mile of AI Delivery

FDE-Gym tests whether agents deliver systems that actually move a customer's business KPI, not just satisfy the written request. The best model scores 0.56 and fully resolves 14 of 95 tasks; the real failure isn't vague problems but hidden rules, where leading models drop from 0.73 to 0.39 once constraints go unstated.

Agents · Added Oct 16

More Methods, Better Agents? A Reality Check on Self-Distillation in Agentic Reinforcement Learning

Benchmarking ten self-distillation methods for RL agent training against plain GRPO across 132 runs and 12K H800-hours: only two beat baseline on average, win-counts hide full collapses (one method scored 0% on WebShop), and matched-compute comparisons erase even the winners' edges.

LLM Training · Added Oct 16

Beyond Domain-Routed Supervision: Cross-Teacher Guarding for Multi-Teacher On-Policy Distillation

Multi-teacher distillation into one small student quietly destroys medical accuracy when combined with math or code, because routed updates still collide on shared weights. Road MOPD scores each trajectory with every teacher, reweights tokens on agreement and projects away conflicting gradients, recovering 5.6–8.8 points without hurting the other domains.

Agents · Added Oct 16

Player-1: Learning to Play from Synthetic Experience

Screen-reading agents choke on games: no cursor, held inputs, world moves on its own. Player-1 skips per-game demos — SFT on 6k trajectories, then PPO across 16 synthetic mini-games with programmatic rewards — lifting Minecraft success 10%→38% and transferring to unseen Doom and Watch Dogs 2.

Agents · Added Oct 16

Doc2Harness: Benchmarking LLMs on Long-Horizon Harness Engineering of Coding Agents

Coding benchmarks score models on fixing issues inside mature repos, but shipped agents are model-plus-harness, and nobody tests whether a model can build the harness itself. Doc2Harness does: spec plus skeleton, scripted mock-model playback, positive controls against vacuous passes. Top model fails 47% of 328 cases.

Never fall behind the literature again.

Free account. Follow topics, build your queue.

Browse topics

Research papers, read to you in five minutes.Research papers, read to you in five minutes.

Research papers, read to you in five minutes.
Research papers, read to you in five minutes.