Jiajun Fan et al. (9 authors) · 2026-10-27

AudioRSI: Recursively Self-Improving Audio Reasoners

“The recursion is what sustains progress: under a frozen external judge, out-of-distribution accuracy plateaus during training, while a critic refreshed from the improving policy keeps it rising.”

Who

Jiajun Fan et al. (9 authors)

Date

2026-10-27

Evidence

Very low certainty Starts at Low: the developer’s own numbers.

Very lowLowModerateHigh

Low ↓ imprecise ↓ indirect measure ↓ risk of bias → Very low

• ↓ imprecise. Every configuration is one training run. The refreshed critic leads the frozen judge by a point on MMAU test-mini (76.10 against 75.10), and the paper gives no confidence intervals. “Each configuration here is trained with a single seed, so where AudioRSI edges out another system by a fraction of a point, as on MMSU, we read it as comparable , not decisively better.”
• ↓ indirect measure. The plateau in the quote is from the run with a frozen Ke-Omni-R judge, a different model from the refreshed critic. The comparison that keeps the authors' own critic is the period sweep. “The frozen judge is also a different model, so the period sweep below, which keeps our own critic, is the more direct test of refreshing.”
• ↓ risk of bias. The period T=200 and the ablations, including the frozen-judge comparison, are scored on MMAU test-mini, one of the four test benchmarks. The authors say the other three are the cleaner test. “the other three benchmarks, evaluated only at the default, are the cleaner test of the chosen period.”

how we rate evidence

Replication

No rerun found. The paper links no code. We found no rerun by another group. Searched up to 2026-10-29. all claims

Checked

2026-10-29, against the source page

Links

Source ↗ Paper ↗

Also on the page

• Qwen2.5-Omni-7B is trained with GRPO on AVQA. The same model generates reasoning paths and, in critic mode, compares them pairwise with the audio in context; its judgments are added to the outcome reward. Every 200 steps the critic is replaced by the current policy, and the handoff is skipped if the actor's outcome accuracy on a training mini-batch has regressed.
• On MMAU test-mini the default run scores 76.10, against 75.10 with a frozen external judge (Ke-Omni-R) and 74.60 with outcome-only rewards, the authors' Ke-Omni-R reproduction. The authors note that the frozen judge is a different model, and that the period sweep on the same benchmark, from T=50 to T=1000, is the more direct test of refreshing.
• The four test benchmarks (MMAU test-mini, MMSU, MMAU-Pro, MMAR) are disjoint from the AVQA training data. Rows the authors did not evaluate themselves, including the proprietary models, come from other protocols, and the authors call comparisons with them "indicative only".
• The authors' limits: "Each configuration here is trained with a single seed", and "Our evidence comes from one backbone and one training corpus".

In the self-improving AI catalog

• AudioRSI: what it changes, whether the change lasts, and how it was tested

How it got here

Earlier work in this catalog that AudioRSI cites, oldest first. Under each: where that work sits on the map, then where AudioRSI sits (changes in bold), and what AudioRSI's own paper says about it. Its whole family on the citation map

1. 2023 STOP sets itself apart

   Tools, own code → Model weights · becomes the next version · improves the improver

   “STOP [45] makes recursive self-improvement explicit but recurses on the scaffolding program around a fixed language model.”

   “AudioRSI instead recurses on the training signal online, inside a single GRPO run, so each round’s improvement is carried by the model’s own weights into the next round’s critic.”

2. 2024 Self-Rewarding Language Models sets itself apart

   Model weights · becomes the next version · improves the improver

   “Self-Rewarding LMs [40] let each DPO iteration’s model score the data for the next, CoVo [41] builds an intrinsic reward from the consistency of intermediate reasoning states, Critique-GRPO [42] pairs numerical reward with natural-language critiques, iGRPO [43] conditions a second GRPO stage on the policy’s own best draft, and Vision-SR1 [44] rewards a VLM for visual descriptions that stand without the image.”

   “It also differs in what the critic may use: every method just listed computes its reward from text alone (Vision-SR1 withholds the image by design, and the others address text-only tasks), whereas our critic is conditioned on the waveform itself (Equation (4)) and can reject a path that cites an acoustic event the recording does not contain.”

3. 2026 AudioRSI

   Model weights · becomes the next version · improves the improver

← Ledger Report an error