A new preprint (arXiv:2608.00017) shows frozen-weight LLM agents endorsing 31% to 54% of their own wrong answers. The authors call the failure mode 'Memory Reward Inflation' — the Echo Gap.
Key facts
- 31%-54% of wrong answers self-endorsed as correct
- arXiv:2608.00017, 'Memory Reward Inflation'
- Echo Gap persists across stronger LLMs
- Frozen weights, no parameter updates
- Preprint by @rohanpaul_ai, arXiv 2608.00017
The promise of self-improving agents is that they get better over time. A new preprint challenges that assumption: a frozen-weight agent can get worse purely through the memories it stores and trusts. The mechanism is simple — the agent logs past episodes, scores them with an LLM judge, and reuses high-scoring ones as precedents, all without touching model weights. According to the arXiv preprint, the flaw is in the grader itself: across the paper's factual banks, the tested models endorsed 31% to 54% of their own wrong answers as correct. That score then enters persistent memory, and the mistake influences future decisions instead of vanishing after one bad answer.
Why stronger models don't fix it
The paper's key negative result: upgrading the grading LLM doesn't repair the damage. Stronger or different LLMs still produce grading errors that correlate with the original self-grading bias. The authors call this the Echo Gap — a systematic blind spot that compounds as the agent accumulates self-approved precedents. This is not the classic catastrophic forgetting or distribution shift; it's a quiet accumulation of self-reinforced errors in a system that never updates its parameters.
The finding has practical implications for memory-augmented agent frameworks, where episodic storage and retrieval are core design choices. If the reward model is the weak link, then memory persistence turns a one-off mistake into a permanent bias. The paper does not disclose the exact architectures or model sizes tested, but the consistency of the 31%-54% range across factual banks suggests the effect is not model-specific.
The structural problem
The Echo Gap is a reminder that self-improvement loops are only as sound as their reward signal. In RLHF, a bad reward model gets corrected by human feedback; in self-improving agents, there is no such correction loop. The agent's own grader is the sole arbiter, and when it is wrong, the error is enshrined rather than erased. This is a failure mode that frozen-weight agents share with their fine-tuned counterparts — the difference is that frozen weights give a false sense of stability. The preprint does not offer a fix, but the implication is clear: memory hygiene and external validation matter as much as the model itself.
Key Takeaways
- New arXiv paper shows frozen-weight agents endorse 31%-54% of wrong answers, compounding errors via memory.
- Echo Gap resists stronger LLMs.
What to watch
Watch for follow-up work proposing memory validation filters or external reward signals. If the 31%-54% self-endorsement range replicates across more model families, expect agentic frameworks to add memory audits. Also track whether the authors release code or benchmarks for measuring Echo Gap in production agents.
