June 6, 2026

The paper proves an Attention Bottleneck Theorem establishing information-theoretic limits on how far decoder-only transformers can track state in purely neural chain-of-thought. A Deterministic Horizon exists at approximately 19-31 steps beyond which accuracy collapses super-exponentially. Across 12 models and 8 task domains (SWE-Bench, WebArena, SQL-Multi), tool-integrated reasoning achieves 86-94% accuracy versus 24-42% for neural CoT. Fine-tuning improves performance by less than 5%, confirming the limits are architectural, not training-related. Accepted at ICML 2026.

LLMs readily fix errors when presented as external input but fail to correct identical errors framed as their own prior output. The paper isolates the cause: chat-template role labels (user message vs. internal thought vs. tool output vs. system memory), not the content itself. Relabeling an internal erroneous claim as an external source increases explicit correction rates by 23-93 percentage points across 7 model families and 3 domains (p < 0.001 in 10/13 test cells). A prompt-structure intervention requiring no retraining achieves significant improvements.

Researchers from the National University of Singapore published the Audio Interaction Model (AIM), a unified streaming audio framework that combines offline task execution (transcription, translation, music generation) with real-time audio instruction following through an end-to-end architecture. AIM achieves simultaneous low-latency streaming and high-quality offline audio processing without separate models for each task mode, receiving 101 upvotes on HuggingFace Daily Papers.