Sign InOpen Brain
arXivPaperNeeds Review

Phantom Gains: Auditing Self-Improvement Against a Measured Null

Per-problem self-improvement claims can arise from inference and evaluation noise. This audit argues every transition statistic needs a measured null from frozen baseline replicates.

arXiv · Aug 20, 2026
Open Source Open MarkdownOpen JSON
Source Summary

The study ran three rounds of rank-32 LoRA self-training on Qwen3-8B alongside a frozen control using the same pipeline. It found **seven measurement failures** capable of reversing findings when the control was omitted.

Practical Implication

For agent and model evaluations, do not treat problem-level gains and losses as ground truth from one decode. Measure each statistic’s null with baseline replicates, then use per-problem tests and false-discovery-rate control.

Agent-Ready Context
The study ran three rounds of rank-32 LoRA self-training on Qwen3-8B alongside a frozen control using the same pipeline. It found **seven measurement failures** capable of reversing findings when the control was omitted.

For agent and model evaluations, do not treat problem-level gains and losses as ground truth from one decode. Measure each statistic’s null with baseline replicates, then use per-problem tests and false-discovery-rate control.

A single greedy decode gave an untrained model an apparent expansion rate of **0.280**, while the replacement exact test detected **nothing on held-out replicates**. Results for problems the base model never solved remained inconclusive, and the method needs more baseline replicates than many studies possess.
Connected Context · Feed7 Judgment

Phantom Gains converts general benchmark skepticism into a concrete prerequisite for self-improvement claims: estimate a statistic’s null from frozen-control replicates before interpreting per-problem transitions. It shows that one-decode expansion can vanish under a replacement exact test and FDR control. Repeated sampling alone is therefore insufficient unless baseline variability is measured; unsolved-base cases remain unresolved.

Sample More, Reflect Less: Self-Refine and Reflexion Lose to Repeated Sampling at Equal Token Cost, from 1.5B to 7BThe repeated-sampling study establishes a token-matched baseline for reflection claims; Phantom Gains adds that repeated decodes must also estimate frozen-baseline variability and control multiplicity before apparent per-problem changes count as improvement.When Will The Benchmaxxing Plague End? — Nick Heiner, Surge AIThe benchmaxxing critique broadly warns that leaderboard gains can be artifacts; Phantom Gains supplies a specific statistical mechanism and control protocol for detecting such artifacts in self-training evaluations.SocietyBench: Forecasting Counterfactual Social-World EvolutionSocietyBench warns against generalizing variable aggregate results and recommends event-level reporting; Phantom Gains shows that problem-level reporting itself requires a measured null and false-discovery control to avoid mistaking decode noise for transitions.Fisher-R1: Training LLM Agents for Reliable Hypothesis TestingFisher-R1’s reported single-trial improvement is a model-comparison result, while Phantom Gains shows what additional frozen-control replication and per-problem testing would be required before interpreting transition-level changes as reliable self-improvement.
Context Map
benchmarkresearch#benchmark-integrity#agent-evals#agent-reliability
Uncertainty
A single greedy decode gave an untrained model an apparent expansion rate of **0.280**, while the replacement exact test detected **nothing on held-out replicates**. Results for problems the base model never solved remained inconclusive, and the method needs more baseline replicates than many studies possess.