Learning desk

MultiAgent EDU StackGather good sources. Teach what matters.
T5TauricResearch/TradingAgentsT5A Man Who Invented Modern AI (Before Everyone Else) – Jürgen Schmidhuber [video]T5GPT-4 finished training four years ago todayT5AI Settles a 25 Year-Old Problem We Left BehindT5What it was like working on LLMs and security at Meta (2022-2026)T5Ask HN: How do you go from writing code to deploying with agents?T5What Happened: OpenAI and HuggingFaceT5Apple says Mac users in China can connect to Alibaba's Qwen AI serviceT5Show HN: Try Benzi – A coding harness/agent beating Claude Code itself on SonnetT5The AI Apocalypse Is HereT3Auto mode is now the default in Claude Code for Pro, Max, and Team plansT5Show HN: Tura – Build agent that uses 80% less token and delivers better resultsT5TauricResearch/TradingAgentsT5A Man Who Invented Modern AI (Before Everyone Else) – Jürgen Schmidhuber [video]T5GPT-4 finished training four years ago todayT5AI Settles a 25 Year-Old Problem We Left BehindT5What it was like working on LLMs and security at Meta (2022-2026)T5Ask HN: How do you go from writing code to deploying with agents?T5What Happened: OpenAI and HuggingFaceT5Apple says Mac users in China can connect to Alibaba's Qwen AI serviceT5Show HN: Try Benzi – A coding harness/agent beating Claude Code itself on SonnetT5The AI Apocalypse Is HereT3Auto mode is now the default in Claude Code for Pro, Max, and Team plansT5Show HN: Tura – Build agent that uses 80% less token and delivers better results
← Dispatches

The Dark Room in the Reward Channel: Dense Prediction Rewards Collapse GRPO-Trained LLM Agents -- and What Actually Works

Primary research

#717

T1new
Topic
unassigned (set during synthesis)
First seen
2026-07-24 07:15:55
Last seen
2026-07-24 07:15:55

Source raw items (1)

  • arXiv2026-07-24 07:15:14
    The Dark Room in the Reward Channel: Dense Prediction Rewards Collapse GRPO-Trained LLM Agents -- and What Actually Works

    Dense per-step supervision is an appealing remedy for sparse-reward, long-horizon LLM agents: reward the agent for predicting its next observation, and memory should follow. We show that under group-normalized RL (GRPO), this recipe does not merely fail -- it destroys the policy. Across Qwen3-1.7B/4B/8B on ALFWorld, a potential-based prediction reward drives every run into a degenerate absorbing state (prediction accuracy -> 1.0, task success -> 0,episode length pinned at the horizon): the "dark room" pathology, built automatically by the optimizer. A single-factor ablation localizes the cause -- removing only GRPO's std normalization turns the same reward from catastrophic (0%) into baseline parity -- and a two-line proposition explains why: in all-fail groups the z-scored advantage is invariant to the shaping coefficient, so bounded rewards become unbounded pressure and annealing cannot help. Our central insight generalizes this: what z-scoring amplifies is a dense signal's within-group variance while all-fail groups dominate, so signals whose variance decays by mastery are structurally amplifier-safe.This variance-profile criterion retrodicts our collapses, carries preregistered predictions for arms that had not yet run, and is consistent with published reward-channel successes (a compatibility check, not an independent test). Finally, a controlled signal-delivery matrix (identical signal, varying only the consumption mechanism) shows the reward channel is at best neutral while the auxiliary-loss channel gains ~20 points -- and a shuffled-gold placebo matches the true-gold arm, so the gap survives without correct labels. Endpoints are single-seed; seed replication and group-size controls are preregistered and in progress.