AI Reasoning Models Cheat Sheet

Updated 2026-05-18

Next Topic: AI Red Teaming Methodology and Tooling Cheat Sheet

🧠Study flashcards on this topic90 cards · spaced repetition→

AI reasoning models represent a fundamental shift from traditional large language models — instead of predicting the next token immediately, they allocate test-time compute to explore solution paths, verify intermediate steps, and self-correct before producing output. This extended reasoning capability, enabled by reinforcement learning with verifiable rewards (RLVR), allows models to match or exceed human expert performance on mathematics, coding, and scientific reasoning benchmarks. The tradeoff is clear: reasoning models spend more tokens (and cost more per query) in exchange for significantly higher accuracy on hard problems, making the decision of when to use them vs. fast models a critical architectural choice.

What This Cheat Sheet Covers

This topic spans 15 focused tables and 112 indexed concepts, 90 flashcards. Below is a complete table-by-table outline of this topic, spanning foundational concepts through advanced details.

Table 1: Core Reasoning MechanismsTable 2: Training ParadigmsTable 3: Model Architectures and SystemsTable 4: Inference-Time StrategiesTable 5: Reasoning Evaluation BenchmarksTable 6: Reasoning Model CharacteristicsTable 7: Architecture PatternsTable 8: Optimization AlgorithmsTable 9: Prompting and ControlTable 10: Training Data SourcesTable 11: Reasoning Limitations and Failure ModesTable 12: Production Deployment ConsiderationsTable 13: Key Benchmarks by DomainTable 14: When To Use Reasoning Models vs Fast ModelsTable 15: Emerging Research Directions

Quick IndexSubscribe to unlock

A jump-to index of every table row in this cheat sheet.

Mind MapSubscribe to unlock

An interactive map of every table and concept in this topic.

Table 1: Core Reasoning Mechanisms

These fundamental techniques define how reasoning models generate extended thought processes before producing final answers, shifting from immediate next-token prediction to multi-step exploration and verification.

Mechanism	Example	Description
Test-Time Compute Scaling	Allocate 10× more inference compute to improve AIME score from 13% → 79%	• Dynamically increasing computational resources during inference to explore more solution paths • Scales performance predictably as a function of compute budget • Complements training-time scaling
Thinking Tokens	`<thinking>` Step 1: Check if n is prime... `Step 2: Factorize if composite...` `</thinking>`	• Internal reasoning steps generated before visible output • Can be hidden (counted but not shown) or visible (displayed to user) • Budget controlled via token limits or effort levels
Extended Reasoning Chains	Generate 15,000-token reasoning trace for proof verification vs. 500-token CoT	• Substantially longer internal thought processes than chain-of-thought • Enables backtracking, self-correction, and multi-attempt exploration • Unlocks harder problems
Chain-of-Thought (CoT)	"Let's break this down: First... Second... Therefore..."	• Explicit step-by-step reasoning in natural language • Prompt-induced technique vs. built-in reasoning mode • Shorter and more structured than extended reasoning
Self-Verification	Model checks answer: "Wait, if x=5 then 2x+3=13, not 12. Retry..."	• Model evaluates correctness of its own outputs during inference • Enables mid-reasoning corrections without external validator • Emergent behavior from RL training
Self-Reflection	"My initial approach failed. Let me reconsider the constraints..."	• Model critiques and revises its reasoning strategy • Goes beyond verification to strategic re-planning • Appears spontaneously in RLVR-trained models
Multi-Turn Refinement	Iteration 1 → feedback → Iteration 2 → feedback → final answer	• Iterative improvement cycles within a single inference call • Model self-edits intermediate outputs progressively • Distinct from user-driven multi-turn chat
Latent Reasoning	Recurrent block loops in hidden state space (no token generation)	• Reasoning happens in continuous latent space instead of discrete tokens • More computationally efficient than token-based thinking • Experimental architecture pattern

Table 2: Training Paradigms

How reasoning models are trained differs fundamentally from standard supervised fine-tuning — these methods use reinforcement learning with outcome verification to incentivize correct reasoning over superficial pattern matching.

Paradigm	Example	Description
RLVR (Reinforcement Learning with Verifiable Rewards)	Train on math: `+1` if final answer correct, `0` if wrong (verified by execution)	• RL training where rewards are objectively verifiable via external checks (test execution, formal proof, symbolic math) • Incentivizes genuine reasoning over reward hacking

The rest of this section is part of the full cheat sheet

Table 3: Model Architectures and Systems

Specific reasoning model implementations, each with distinct architectural choices and deployment characteristics, representing the production ecosystem as of 2026.

Model	Example	Description
DeepSeek-R1	671B MoE model trained with pure RL • 79.8% on AIME 2024, 97.3% on MATH-500	• First fully open-weights reasoning model comparable to o1 • Exhibits self-verification and reflection without explicit training • Distilled variants at 1.5B-70B sizes

The rest of this section is part of the full cheat sheet

Table 4: Inference-Time Strategies

Techniques applied during model inference to improve reasoning quality by allocating more compute, exploring multiple paths, or selecting optimal outputs from samples.

Strategy	Example	Description
Best-of-N Sampling	Generate 32 solutions, use ORM to rank, return top-1	• Sample multiple independent responses and select best via reward model • Linear scaling in compute (32× cost for 32 samples) • Embarrassingly parallel

The rest of this section is part of the full cheat sheet

Table 1: Core Reasoning Mechanisms

Mechanism	Example	Description
Test-Time Compute Scaling	Allocate 10× more inference compute to improve AIME score from 13% → 79%	• Dynamically increasing computational resources during inference to explore more solution paths • Scales performance predictably as a function of compute budget • Complements training-time scaling
Thinking Tokens	`<thinking>` Step 1: Check if n is prime... `Step 2: Factorize if composite...` `</thinking>`	• Internal reasoning steps generated before visible output • Can be hidden (counted but not shown) or visible (displayed to user) • Budget controlled via token limits or effort levels
Extended Reasoning Chains	Generate 15,000-token reasoning trace for proof verification vs. 500-token CoT	• Substantially longer internal thought processes than chain-of-thought • Enables backtracking, self-correction, and multi-attempt exploration • Unlocks harder problems
Chain-of-Thought (CoT)	"Let's break this down: First... Second... Therefore..."	• Explicit step-by-step reasoning in natural language • Prompt-induced technique vs. built-in reasoning mode • Shorter and more structured than extended reasoning
Self-Verification	Model checks answer: "Wait, if x=5 then 2x+3=13, not 12. Retry..."	• Model evaluates correctness of its own outputs during inference • Enables mid-reasoning corrections without external validator • Emergent behavior from RL training
Self-Reflection	"My initial approach failed. Let me reconsider the constraints..."	• Model critiques and revises its reasoning strategy • Goes beyond verification to strategic re-planning • Appears spontaneously in RLVR-trained models
Multi-Turn Refinement	Iteration 1 → feedback → Iteration 2 → feedback → final answer	• Iterative improvement cycles within a single inference call • Model self-edits intermediate outputs progressively • Distinct from user-driven multi-turn chat
Latent Reasoning	Recurrent block loops in hidden state space (no token generation)	• Reasoning happens in continuous latent space instead of discrete tokens • More computationally efficient than token-based thinking • Experimental architecture pattern

Table 2: Training Paradigms

Paradigm	Example	Description
RLVR (Reinforcement Learning with Verifiable Rewards)	Train on math: `+1` if final answer correct, `0` if wrong (verified by execution)	• RL training where rewards are objectively verifiable via external checks (test execution, formal proof, symbolic math) • Incentivizes genuine reasoning over reward hacking

Table 3: Model Architectures and Systems

Specific reasoning model implementations, each with distinct architectural choices and deployment characteristics, representing the production ecosystem as of 2026.

Model	Example	Description
DeepSeek-R1	671B MoE model trained with pure RL • 79.8% on AIME 2024, 97.3% on MATH-500	• First fully open-weights reasoning model comparable to o1 • Exhibits self-verification and reflection without explicit training • Distilled variants at 1.5B-70B sizes

Table 4: Inference-Time Strategies

Techniques applied during model inference to improve reasoning quality by allocating more compute, exploring multiple paths, or selecting optimal outputs from samples.

Strategy	Example	Description
Best-of-N Sampling	Generate 32 solutions, use ORM to rank, return top-1	• Sample multiple independent responses and select best via reward model • Linear scaling in compute (32× cost for 32 samples) • Embarrassingly parallel