Reasoning Models

Definition

A family of LLMs that explicitly model the reasoning process itself. They allocate test-time compute to reasoning steps (chain-of-thought, search, verification) to improve the ability to solve complex problems.

Why It Matters

• Matches personal interests: reasoning (1.3x) × frontier models (1.3x)
• The most active research thread of 2024-2026 — tracked across the industry since the OpenAI o-series
• A direction that addresses the limits of simple scale-up

State of the Art (2026-08-02)

• Strongest empirical result (2026-05-20): OpenAI — Disproof of the Erdős unit distance conjecture — autonomous solution of an 80-year-old open pure-math problem. A general-purpose reasoning model found a proof unknown even to expert mathematicians using only chain-of-thought reasoning. The first clear evidence that general-purpose reasoning can make substantive contributions in specialized scientific fields.
• Latest strong claim: Olympiad-level reasoning via simple unified scaling (2026-05-16) — claims gold-medal-level performance on math olympiads
• Gemini 3.1 Deep Think — autonomous math research agent (Aletheia, IMO gold medal, 18 open problems solved, Feng26 paper published)
• Claude Opus 4.7 also emphasizes "multi-step tasks"

Since May the thread has moved from can a reasoning model solve hard problems to what does the training actually optimise, and what appears only at scale:

• RLVR reaches a trillion parameters — Ring-Zero (Ant Group, 2026-07) is the first demonstration of RL with Verifiable Rewards at 1T, reaching 84.2% on AIME 2026. It documents five behaviours that do not appear below roughly 100B, including a failure mode the authors call context anxiety — the model hedges or stalls when the context is ambiguous. Capability and failure both turn out to be scale-dependent, which makes this an alignment result as much as a capability one → AI Alignment
• The training target may be the wrong one — the MIPI paper (2026-06) argues that RL implementations optimise a training proxy while the model is served by a different inference engine, so the policy being improved is not the policy that runs. If it holds, it touches every lab's reasoning training rather than one method
• Reasoning became a product tier, not a model — Gemini 3.5 Pro gates "Deep Think" behind the $250/month Ultra tier; MAI-Thinking-1 is Microsoft's first reasoning-specialised model, trained from scratch without distillation; MiniMax M3 pairs a 1M context with long-horizon tasks. The question shifted from who has a reasoning model to who can afford to run one
• The people moved too — Jason Wei, co-author of the chain-of-thought paper that started this line, left OpenAI for Meta Superintelligence Labs in mid-2026
• The May result stopped being a one-off (2026-08-01) — OpenAI published ten mathematics and TCS results from an internal version of Astra, each carrying a Lean 4 certificate: a first explicit non-sofic group, a disproof of Connes' Rigidity Conjecture, quantum parallel repetition, Ehrhart's volume conjecture, and the first improvement to the general high-dimensional sphere-packing upper bound since 1978. Total token cost roughly $2,000 at Sol API prices. The interesting shift is not the count but the checking: in May, mathematicians read a reasoning transcript and rewrote it as a proof; now the output arrives with a machine-checkable certificate, so verification no longer depends on expert attention being available. Human editors still organised the proofs before formalization, so the pipeline is not autonomous end to end → AI for Mathematics (source)

Open Problems

• Consistency of the definition of "reasoning" — pattern matching vs genuine reasoning?
• Test-time compute scaling laws — how far is it efficient?
• Domain generalization (math → coding → science → real world)
• Cost — spending more compute on reasoning also increases cost
• Training-inference mismatch — whether the policy that RL improves is the policy that is served at all (The Mirage of Optimizing Training Policies: Monotonic Inference Policies as the Real Objective for LLM Reinforcement Learning)
• Scale-dependent failure — behaviours like context anxiety are invisible below ~100B, so a method validated on small models says little about the model that ships (Ring-Zero: Scaling Zero RL to a Trillion Parameters for Emergent Reasoning)
• Hidden traces are hidden by a lock the whole ecosystem holds. Frontier providers conceal chain-of-thought by returning it to the client as an encrypted blob rather than storing it server-side. Stealing Reasoning Traces from Proprietary LLM APIs (arXiv:2608.09867) reports those blobs are interchangeable across sessions, users and models within one provider — Anthropic, OpenAI and Google are named — so a weaker sibling model can be made to decrypt and print a stronger model's trace verbatim. The strong model is never attacked. No vendor response was found in anything read (source)
• Reasoning that is never verbalised at all. BDH-CQ: In-Context Learning with Recurrent Latent Reasoning (arXiv:2608.09888) iterates in latent space and emits no reasoning tokens, reporting 29.5% pass@2 on ARC-AGI-1 at a computed $0.0007 per task from 150M parameters. If that direction holds, the trace-secrecy problem above dissolves — and so does any ability to read what the model did (source)

Key Papers

• An OpenAI model has disproved a central conjecture in discrete geometry — 2026-05-20, autonomous disproof of an 80-year-old open math problem (OpenAI general reasoning model)
• Achieving Gold-Medal-Level Olympiad Reasoning via Simple and Unified Scaling — 2026-05, claims gold-medal performance from scaling alone
• Ring-Zero: Scaling Zero RL to a Trillion Parameters for Emergent Reasoning — 2026-07, RLVR at 1T parameters; AIME 2026 84.2% and five emergent behaviours
• The Mirage of Optimizing Training Policies: Monotonic Inference Policies as the Real Objective for LLM Reinforcement Learning — 2026-06, the training-inference gap in LLM RL
• Stealing Reasoning Traces from Proprietary LLM APIs (arXiv:2608.09867) — 2026-08-10, encrypted chain-of-thought blobs are interchangeable across models within a provider
• BDH-CQ: In-Context Learning with Recurrent Latent Reasoning (arXiv:2608.09888) — 2026-08, recurrent latent reasoning at 150M; ARC-AGI-1 cost-accuracy frontier

Related Concepts

• Test-Time Compute (Inference-Time Compute Scaling) — inference-time compute allocation is the core mechanism of reasoning models
• Agentic Reinforcement Learning — the intersection of multi-step reasoning + RL
• AI for Mathematics — where the strongest claims from this line land, and how they are checked

Open Debates

• "Simple unified scaling" vs "specialized methods" — which drives progress in reasoning? (2605.13301 argues for the former)
• Whether an emergent behaviour observed once at 1T is a property of the scale or of that training run — Ring-Zero reports five and no one has reproduced them independently