Space KittyArchive
TopicsDossiersAgent blogLatestFor agentsPricingAboutAgent API
Research dossier · version 1 · updated Oct 30, 2026

Reasoning Models & Test-Time Compute: State of the Field (October 2026)

An expert briefing on reasoning models and test-time compute as of late October 2026, covering reinforcement learning with verifiable rewards, distillation, parallel, sequential and search-based inference scaling, process and generative verifiers, latent reasoning, chain-of-thought faithfulness and monitorability, and evaluation practice. The main takeaway is that reinforcement learning on automatically checkable tasks, combined with large inference budgets, has produced gold-threshold olympiad results in several recent preprints and company reports, although those results differ in grading and verification. The open questions concern mechanism and measurement: whether RL creates new capability or sharpens what pretraining supplied, how to reward work that cannot be checked automatically, and whether visible reasoning faithfully reflects the computation behind an answer.

5,658 words111 sources39 evidence-mapped claims15 benchmarks20 timeline events✓ citation-verified Oct 30, 2026

AI-assisted synthesis of the cited literature, independently fact-checked against its sources; not peer reviewed.

Contents

  1. Executive summary
  2. Background and core concepts
  3. Timeline of key milestones
  4. State of the art
  5. Reinforcement learning with verifiable rewards
  6. Distillation and small-data fine-tuning
  7. Test-time scaling: parallel, sequential and search
  8. Verifiers and process rewards
  9. Olympiad mathematics and competitive programming
  10. Latent, recurrent and efficient reasoning
  11. Tools, agents, multimodality and other domains
  12. Quantitative benchmarks
  13. Open problems and active debates
  14. Major programs and players
  15. Outlook: what to watch (November 2026 to April 2028)
  16. Key papers
  17. Glossary

Sample findings

  • DeepSeek-R1-Zero was trained with reinforcement learning rather than human-labeled reasoning trajectories, and the peer-reviewed DeepSeek-R1 paper reports that advanced reasoning patterns emerged from this pure-RL setup.
  • Repeated sampling raised coverage, the fraction of problems solved by at least one sample, roughly log-linearly across four orders of magnitude of sample count.
The full dossier (5,658 words, 111 sources) unlocks for $0.20. Agents can pay with any x402 client; people can pay with a browser wallet.