TL;DR — Audio LLMs trained only on outcome rewards produce hallucinatory, inconsistent, and unscalable reasoning chains. CESAR instead rewards the reasoning process itself via online RL (GRPO), resolving test-time inverse scaling and reaching SOTA on MMAU — outperforming Gemini 2.5 Pro and GPT-4o Audio.

The problem: test-time inverse scaling

Adding chain-of-thought reasoning to Audio LLMs often degrades performance — a phenomenon we term test-time inverse scaling. Longer reasoning chains yield progressively worse results, because outcome-only rewards leave the reasoning process completely unsupervised.

Without process rewards

The reasoning chain is invisible to the reward. Models produce hallucinatory, inconsistent reasoning that accumulates errors over longer chains — reasoning becomes a liability.

With CESAR

Process rewards score consistency, structured analytical patterns, causal reasoning, and calibrated depth — turning reasoning into a genuine, scalable capability.

Method

CESAR uses Group Relative Policy Optimization (GRPO) with a multi-faceted reward suite that goes beyond simple correctness — every component targets a specific failure mode of unsupervised reasoning.

1

Correctness

Standard outcome reward — is the final answer right?

2

Consistency

Reasoning ↔ question and reasoning ↔ answer alignment.

3

Structure

Analytical patterns, logical connectives, domain vocabulary.

4

Depth

Overthinking penalty — calibrate reasoning length to the task.

2026-07-15T22:18:21.112288 image/svg+xml Matplotlib v3.10.8, https://matplotlib.org/
Figure 1. Outcome-only RL (top) rewards only final-answer correctness and format. CESAR (bottom) adds process rewards — consistency, structure, and an overthinking penalty — on the reasoning chain itself.

Results

SOTA
MMAU Test-mini
> 2.5 Pro
beats Gemini 2.5 Pro
> GPT-4o
beats GPT-4o Audio
+0.038
positive scaling slope
ModelMMAU (total)Test-time scaling
Qwen2.5-Omni-7B (base)65.2%Inverse (slope −0.51)
Ke-Omni-R (outcome-only RL)74.6%Flat (slope −0.007)
CESAR w/o overthinking penalty76.5%Positive
CESAR (Ours)77.1% BestPositive (slope +0.038)
2026-07-15T21:57:59.529021 image/svg+xml Matplotlib v3.10.8, https://matplotlib.org/ Sound Music Speech Total 50 55 60 65 70 75 80 85 90 Accuracy % (MMAU) Qwen2.5-Omni-7B Ke-Omni-R CESAR (w/o Overthink.) CESAR (Ours)
Figure 2. Per-category MMAU accuracy. CESAR leads the base model (Qwen2.5-Omni-7B) and the outcome-only baseline (Ke-Omni-R) on Sound, Music, Speech and overall; its no-penalty variant does too, except on Music where it matches Ke-Omni-R — process rewards improve perception, not just reasoning.

Test-time scaling

With process rewards, more reasoning stops hurting and starts helping. The base model collapses as reasoning length grows; CESAR stays flat-high and even rises.

2026-07-15T21:57:59.349975 image/svg+xml Matplotlib v3.10.8, https://matplotlib.org/ 0 20 40 60 80 100 Avg reasoning tokens 20 40 60 80 Accuracy % Sound 0 20 40 60 80 100 Avg reasoning tokens 20 40 60 80 Accuracy % Music 0 20 40 60 80 100 Avg reasoning tokens 20 40 60 80 Accuracy % Speech 0 20 40 60 80 100 Avg reasoning tokens 20 40 60 80 Accuracy % Total Qwen2.5-Omni-7B Ke-Omni-R CESAR (w/o Overthink.) CESAR (Ours)
Figure 3. Accuracy vs. average reasoning tokens on MMAU. Qwen2.5-Omni-7B suffers inverse scaling — accuracy collapses — while CESAR, its no-penalty variant, and Ke-Omni-R stay flat and high.
2026-07-15T21:57:59.472378 image/svg+xml Matplotlib v3.10.8, https://matplotlib.org/ Qwen2.5- Omni-7B Ke-Omni-R CESAR (Ours) −0.6 −0.5 −0.4 −0.3 −0.2 −0.1 0.0 0.1 Scaling slope Δaccuracy / token -0.510 -0.007 +0.038 reasoning helps reasoning hurts (inverse scaling)
Figure 4. Fitted scaling slopes (Δaccuracy per reasoning token). Qwen2.5-Omni-7B is strongly negative (−0.51); Ke-Omni-R is flat (−0.007); only CESAR is clearly positive (+0.038) — reasoning genuinely helps.

Publication journey

Oct 2025
arXiv preprint released
arXiv:2510.20867.
Oct 2025
Submitted to ICLR 2026
The Fourteenth International Conference on Learning Representations (Submission #8335).
Jan 2026
Accepted at ICLR 2026 (Poster)
Apr 2026
Presented · Rio de Janeiro, Brazil
ICLR 2026.

BibTeX

@inproceedings{fan2026incentivizing,
  title     = {Incentivizing Consistent, Effective and Scalable Reasoning
               Capability in Audio {LLM}s via Reasoning Process Rewards},
  author    = {Jiajun Fan and Roger Ren and Jingyuan Li and Rahul Pandey and
               Prashanth Gurunath Shivakumar and Ivan Bulyko and Ankur Gandhe
               and Ge Liu and Yile Gu},
  booktitle = {The Fourteenth International Conference on Learning Representations},
  year      = {2026},
  url       = {https://openreview.net/forum?id=DUr48hxO2h}
}