Loading the edition…
· Past edition · 18 stories
Retrospective edition reconstructed from later captures. Paper versions may include later updates; coverage labels describe the evidence used.
Research & academia
The authors target a reasoning failure in sub-3-bit models: quantization makes the model’s own generation paths diverge from the fixed prefixes used in QAD training.
Why it matters: For low-bit reasoning, training on the model’s actual quantized generation paths may address failures that continued teacher-forced QAD leaves behind.
Research & academia
Open each story in the reader. Summaries are shown when available; some sources provide an abstract only. With an account you can keep your own notes.
A preprint argues that signals used to measure social sycophancy can overlap with conversational receptiveness, complicating evaluation. This account uses the abstract.
Why it matters: The paper questions whether a benchmark captures the behavior it intends to measure.
Research & academia
A preprint describes AIDE², a system for iteratively modifying research agents. This account is based on the abstract and does not independently establish sustained general self-improvement.
Why it matters: The work explores how agents might improve their own research workflows.
Research & academia
Specializes quantization for prefill and decode; the authors report accuracy and latency results on named model and hardware configurations.
Why it matters: Prefill and decode can have different quantization requirements.