AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

Do Audio LLMs Listen or Read? Analyzing and Mitigating Paralinguistic Failures with VoxParadox

Jiacheng Pang, Ashutosh Chaubey, Mohammad Soleymani

ICML 2026 · Large language models › Language, safety and interpretability · approach: Speech and audio language models · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add PCLM (instead of SFT only) to Audio Flamingo 3 (ablation)AF3 SFT · VoxParadox (avg of 10 tasks)GT accuracy (%): 34.80 -> 60.00Improved
add PCLM to Audio Flamingo 3 over the original modelAF3 · VoxParadox (avg of 10 tasks)GT accuracy (%): 17.40 -> 60.00Improved
add PCLM to Qwen2-Audio over the original modelQwen2-Audio · VoxParadox (avg of 10 tasks)GT accuracy (%): 14.85 -> 68.65Improved
add PCLM (instead of SFT only) to Qwen2-Audio (ablation)Qwen2-Audio SFT · VoxParadox (avg of 10 tasks)GT accuracy (%): 45.70 -> 68.65Improved
add DPO on top of PCLM for AF3 (ablation)AF3 + PCLM · VoxParadox (avg of 10 tasks)GT accuracy (%): 60.00 -> 65.20Improved
add DPO on top of PCLM for Qwen2-Audio (ablation)Qwen2-Audio + PCLM · VoxParadox (avg of 10 tasks)GT accuracy (%): 68.65 -> 72.30Improved
add DPO to SFT without PCLM for AF3 (ablation)AF3 SFT · VoxParadox (avg of 10 tasks)GT accuracy (%): 34.80 -> 40.30Improved
add DPO to SFT without PCLM for Qwen2-Audio (ablation)Qwen2-Audio SFT · VoxParadox (avg of 10 tasks)GT accuracy (%): 45.70 -> 44.50Hurt
add DPO and PCLM to AF3 to reduce transcript-followingAF3 · VoxParadox (avg of 10 tasks)adversarial-label agreement (%): 68.50 -> 22.60Improved
add DPO and PCLM to Qwen2-Audio to reduce transcript-followingQwen2-Audio · VoxParadox (avg of 10 tasks)adversarial-label agreement (%): 70.25 -> 15.95Improved
reverse the audio clips to remove intelligible speechAF3 · VoxParadox (avg of 10 tasks)GT accuracy (%): 17.40 -> 38.80Improved
concatenate intermediate encoder layers 5 and 15 tokens after final-layer tokens (ablation)AF3 · VoxParadox (avg of 10 tasks)GT accuracy (%): 17.40 -> 19.75Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy