▲The Inference-Cost Efficiency Race

Score history

Daily conviction score, 0 to 100. Higher means the thesis is more strongly corroborated.

Now 100 · +2 since Oct 11 · ranged 98 to 100

Showing the last few days. Unlock full score history.

Why the conviction moved

• Oct 12
  Strengthened +2

  An arXiv paper corrects the dependencies missing inside parallel speculative drafts, which should raise how many drafted tokens get accepted and so lower serving latency and cost per query. A companion paper trains causal models on chains of edits so output can be revised after it is written. Together they show decoding order itself is becoming a target for efficiency research.

• Oct 10
  Strengthened +2

  A new arXiv paper lets the acting model decide which earlier tool results to compress, reversibly. That targets tool-output bloat in context, which drives per-query tokens and KV-cache cost in agent workloads.

Showing the last 2 days. Unlock the full record.

Source trail

• Supporting · October 12, 2026

  Two arXiv papers target the ordering limits of autoregressive decoding, one for speculative speed and one for revisable output

  An arXiv paper corrects the dependencies missing inside parallel speculative drafts, which should raise how many drafted tokens get accepted and so lower serving latency and cost per query. A companion paper trains causal models on chains of edits so output can be revised after it is written. Together they show decoding order itself is becoming a target for efficiency research.

  arXiv cs.CL (Speculative Decoding)

• Supporting · October 10, 2026

  New arXiv papers target two agent bottlenecks: tool-output bloat in context and lack of enterprise training data

  A new arXiv paper lets the acting model decide which earlier tool results to compress, reversibly. That targets tool-output bloat in context, which drives per-query tokens and KV-cache cost in agent workloads.

  arXiv cs.AI