Reading the Mind of an Audio LLM: Its Brain Map and Middle-Layer Workspace
Published:
A speech model answers a spoken question in one word. We read its middle layers while it listened — the answer was already there, in words it never says. Read more
Published:
A speech model answers a spoken question in one word. We read its middle layers while it listened — the answer was already there, in words it never says. Read more
Published:
RL post-training runs on one fixed coefficient that must both protect the model and get out of its way. Subtract the advantage — one line. Read more
Published:
Voice agents could be measured, never improved. We closed the loop — two open models talking in native audio — and task success more than doubled. Read more
Published:
Voice agents could be measured, never improved. We closed the loop — two open models talking in native audio — and task success more than doubled. Read more
Published:
A speech model answers a spoken question in one word. We read its middle layers while it listened — the answer was already there, in words it never says. Read more
Published:
Telling an Audio LLM to think cost it 3.40 points. Rewarding the reasoning process instead earns 3.40 — the same margin, mirrored. Read more
Published:
RL post-training runs on one fixed coefficient that must both protect the model and get out of its way. Subtract the advantage — one line. Read more
Published:
RL post-training runs on one fixed coefficient that must both protect the model and get out of its way. Subtract the advantage — one line. Read more
Published:
A speech model answers a spoken question in one word. We read its middle layers while it listened — the answer was already there, in words it never says. Read more
Published:
A speech model answers a spoken question in one word. We read its middle layers while it listened — the answer was already there, in words it never says. Read more
Published:
Telling an Audio LLM to think cost it 3.40 points. Rewarding the reasoning process instead earns 3.40 — the same margin, mirrored. Read more
Published:
Voice agents could be measured, never improved. We closed the loop — two open models talking in native audio — and task success more than doubled. Read more
Published:
Telling an Audio LLM to think cost it 3.40 points. Rewarding the reasoning process instead earns 3.40 — the same margin, mirrored. Read more
Published:
RL post-training runs on one fixed coefficient that must both protect the model and get out of its way. Subtract the advantage — one line. Read more
Published:
Telling an Audio LLM to think cost it 3.40 points. Rewarding the reasoning process instead earns 3.40 — the same margin, mirrored. Read more
Published:
Voice agents could be measured, never improved. We closed the loop — two open models talking in native audio — and task success more than doubled. Read more