aiqbench analysis
This research addresses the performance gap between speech-based and text-based large language models by introducing VIRBA, a reinforcement learning framework designed to align model behavior across diverse acoustic conditions like varying accents and noise. By optimizing for cross-acoustic invariance and robust reasoning policies, the study demonstrates significant improvements in spoken QA and audio reasoning tasks without relying solely on teacher imitation.
Enhancing Speech Large Language Models through Reinforced Behavior Alignment
arXiv:2509.03526v3 Announce Type: replace Abstract: Speech large language models (SpeechLMs) can process spoken requests, yet still lag behind text-prompted large language models (LLMs) on instruction following and reasoning. We argue that this gap is not only an ASR or representation problem: the same semantic instruction can induce different response policies when spoken with different speakers, accents, noise conditions, prosody, or disfluencies. We propose Verifiable Invariant Reinforced Behavior Alignment (VIRBA), a reinforcement-learning framework for aligning SpeechLM behavior across acoustic realizations of the same semantic intent. VIRBA builds multi-view spoken instruction groups, scores sampled responses with semantic preference, rule-verifiable correctness, cross-acoustic invariance, and adaptive reasoning rewards, and optimizes the model with Cross-Acoustic Group Relative Policy Optimization (CA-GRPO). The resulting objective moves SpeechLM alignment beyond teacher imitation toward robust reasoning policies that remain stable across realistic spoken realizations. Experiments with recent SpeechLM baselines, disfluency robustness, spoken QA, audio reasoning, and speech-to-text translation show the largest gains on reasoning-heavy and acoustically perturbed spoken prompts.
Abstract:Speech large language models (SpeechLMs) can process spoken requests, yet still lag behind text-prompted large language models (LLMs) on instruction following and reasoning. We argue that this gap is not only an ASR or representation problem: the same semantic instruction can induce different response policies when spoken with different speakers, accents, noise conditions, prosody, or disfluencies. We propose Verifiable Invariant Reinforced Behavior Alignment (VIRBA), a reinforcement-learning framework for aligning SpeechLM behavior across acoustic realizations of the same semantic intent. VIRBA builds multi-view spoken instruction groups, scores sampled responses with semantic preference, rule-verifiable correctness, cross-acoustic invariance, and adaptive reasoning rewards, and optimizes the model with Cross-Acoustic Group Relative Policy Optimization (CA-GRPO). The resulting objective moves SpeechLM alignment beyond teacher imitation toward robust reasoning policies that remain stable across realistic spoken realizations. Experiments with recent SpeechLM baselines, disfluency robustness, spoken QA, audio reasoning, and speech-to-text translation show the largest gains on reasoning-heavy and acoustically perturbed spoken prompts.
Computation and Language (cs.CL); Audio and Speech Processing (eess.AS)
arXiv:2509.03526 [cs.CL]
(or arXiv:2509.03526v3 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2509.03526 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 253 words
Related coverage
Physics-enhanced reinforcement learning for real-time optimal control of dynamical systems
arXiv · Jul 20, 2026
Enhanced Evolutionary Multi-Objective Deep Reinforcement Learning for Reliable and Efficient Wireless Rechargeable Sensor Networks
arXiv · Jun 17, 2026
MCite-RL: Towards Reliable Multimodal RAG via Citation-enhanced Agentic Reinforcement Learning
arXiv · Aug 25, 2026
CMRL: Collision-Aware and Memory-Enhanced Reinforcement Learning for UAV Navigation in Multi-Scale Obstacle Environments
arXiv · Sep 9, 2026