This training is designed to cover end to end the reasoning-LLM paradigm that began with the launch of OpenAI o1 in September 2024 and has been firmly placed at the center of the world AI ecosystem throughout 2025-2026 with DeepSeek R1, Gemini 2.5 Deep Think, Claude Extended Thinking, Qwen3 Reasoning, GLM-4.6-thinking, and GPT-5 reasoning mode. What separates reasoning models from classic LLMs is not merely longer output; it is the explicit training of chain-of-thought behavior as a post-training objective, the structural separation between the thinking trace and the final answer, test-time compute becoming the first-order determinant of model performance, and the alignment paradigm shifting via process reward. In Turkey, a training that addresses this discipline end to end — from System-1 vs System-2 cognition theory, R1-style RL emergence, reasoning-distillation recipes, MCTS-based search, hybrid/mixed-mode reasoning, to production inference engineering — is virtually nonexistent; existing content either stays at the level of o1/R1 paper summaries or remains shallow at the prompt-CoT level. This program is designed to fill that gap as Turkey's most comprehensive production-grade reasoning-models reference training.
The strategic backbone of the program is the first module, which clearly frames the structural difference between classic chat LLMs (System 1 — fast/intuitive) and reasoning LLMs (System 2 — slow/deliberative). Kahneman's dual-process cognitive model is projected onto the LLM plane; the OpenAI o1 → o3 → o4 evolution, the open-source shock created by DeepSeek R1 in January 2025, and the subsequent moves of Gemini 2.5 Deep Think + Claude Extended Thinking + Qwen3 are mapped historically. Which task is optimal for reasoning vs classic LLMs — a concrete decision matrix is provided: reasoning models are clearly superior for math-olympiad problem solving, algorithm design, complex code debugging, multi-step planning, scientific reasoning, and formal verification; classic LLMs are more efficient for short Q&A, fast RAG, summarization, and simple translation. A disciplined approach to selecting optimal choice in the cost-latency-quality triangle is established.
The second module, addressing reasoning models' internal anatomy, covers end to end how chain-of-thought behavior is induced during post-training, the structural distinction between thinking blocks and output tokens, and provider-specific protocol differences. The Anthropic Claude Extended Thinking API's thinking-blocks + budget_tokens (1K-32K) structure; OpenAI Reasoning Items + reasoning_effort (minimal / low / medium / high) parameters; Google Gemini 2.5 Pro Thoughts + thinking_config; the
The third module, addressing the most critical component of reasoning-model training — the reward signal — from two perspectives, covers the outcome reward (ORM) and process reward (PRM) distinction in detail. In the ORM approach, reward is given via rule-based signals such as a math problem's final answer (SymPy exact match + numeric tolerance), code's test pass rate (pytest), or format compliance (regex verifier); the cornerstone of DeepSeek-Math and R1's success is this approach. PRM, on the other hand, scores each intermediate reasoning step — AllenAI Tülu 3 PRM (2025), Math-Shepherd automatic process-supervision generation, the OpenAI PRM800K dataset, and Yuan 2024 Implicit PRM (deriving PRM from outcome reward) approaches are comparatively covered. Which reasoning task is optimal for ORM and which for PRM — this decision is clarified with concrete benchmark data; PRM's reward-hacking risks and the mitigation recipe are provided.
The theoretical peak of the program is the fourth module, dedicated to the test-time compute scaling laws introduced by Charlie Snell and DeepMind's team in 2024. The marginal-gain balance between pre-train compute and test-time compute, optimal compute allocation by task difficulty, parallel scaling methods (best-of-N, self-consistency majority voting, weighted voting with PRM), sequential scaling (iterative refinement via Self-Refine and Reflexion), and search-based scaling (MCTS, REBASE, beam-guided) are addressed comprehensively. A concrete FLOPS-economy comparison is made between 1 GPU-hour of reasoning compute and 1 GPU-hour of pre-training compute; the discipline of dynamically allocating the compute budget by task difficulty is established.
The fifth module dissects the DeepSeek R1 and R1-Zero pipelines end to end. R1-Zero — the first model to prove that reasoning emergence is possible with pure RL without cold-start SFT — is a paradigmatic finding; the aha-moment phenomenon, reasoning-length emergence, the language-mixing problem, and the mitigation recipe are covered in detail. R1's multi-stage pipeline (SFT cold-start → reasoning RL → SFT mix → general RL) is analyzed step by step. Rule-based reward design — SymPy exact match + numeric tolerance for math, pytest pass-rate for code, regex verifier for format compliance — is performed practically. Open-source reproduction projects (HuggingFace Open-R1, ByteDance DAPO, SimpleRL-Zoo, TinyZero, Open-Reasoner-Zero) are comparatively examined; participants can attempt reasoning emergence on their own 7B-32B base models with GRPO + rule-based reward.
The sixth module addresses the engineering of transferring the capabilities of R1 and other large reasoning models to compact 1.5B-32B models via SFT-based distillation. Dataset-construction recipes of the OpenThoughts-114K, S1.1K (Stanford 2025 — reaching o1-mini level with 1,000 samples), LIMO-817 (95% AIME accuracy with just 817 samples), Bespoke-Stratos-32B, Sky-T1, and DeepScaleR-1.5B projects; strategies for collecting reasoning traces from teacher models (R1 / o3 / Gemini Deep Think); and the quality-filtering discipline (balancing difficulty + correctness + diversity) are covered in detail. A custom reasoning-distillation pipeline is built for the Turkish math, code, and legal domains; a domain-specific reasoning model is produced on top of Qwen3-0.5B or Llama 3.2 1B. The distillation + GRPO hybrid approach is presented as a production recipe.
The seventh module covers end to end the heuristic prompt + inference patterns that elicit reasoning behavior even on classic LLMs without a dedicated reasoning model. Chain-of-Thought (Wei 2022), Self-Consistency (Wang 2023), Tree-of-Thoughts (Yao 2023), Self-Refine (Madaan 2023), Reflexion (Shinn 2023), Reasoning via Planning (Hao 2023), and Plan-and-Solve (Wang 2023) algorithms are each covered with mathematical formulation + Python implementation + benchmark results. A native reasoning model (R1, o3) vs prompt-based CoT (GPT-4o, Claude Sonnet) comparison is made with concrete benchmarks — when native reasoning vs when prompt-based patterns are sufficient is clarified evidence-based.
The eighth module details the adaptation of DeepMind AlphaZero's MCTS approach to LLM reasoning. The MCTS loop of Selection (UCB) → Expansion → Simulation → Backpropagation is projected onto the LLM token tree; step-value estimation via PRM-guided MCTS, the ReST-MCTS (Zhang 2024) self-training pipeline, the REBASE algorithm (beam-search-style cost-efficient MCTS), and the AlphaMath (MCTS + LLaMA math reasoning) projects are presented with implementation details. Parallel rollouts with vLLM, tuning the c_puct exploration parameter, and MCTS budget control (time, FLOPS, max-depth) are conveyed with production examples.
The ninth module addresses the hybrid/mixed-mode structures and thinking-budget engineering offered by most modern reasoning models (Claude Opus 4.7, Qwen3, GLM-4.6, GPT-5). The Anthropic budget_tokens (1K-32K) + interleaved thinking; the OpenAI reasoning_effort parameter; the Qwen3 /think /no_think directive; the GLM-4.6 thinking_mode; the Gemini 2.5 thinking_config — all are covered comparatively in practice. Dynamic thinking-budget routing via a classifier that predicts query difficulty, stage-based budget with early exit, and confidence-aware routing (classic LLM → reasoning fallback pattern) design are performed. In production experience, this discipline reduces reasoning cost by 40-70%.
The tenth module addresses the engineering discipline of serving reasoning models' long thinking traces (4K-32K tokens) with low latency and high throughput in production. It is conveyed through vLLM continuous batching + PagedAttention, SGLang RadixAttention + reasoning-aware caching (shared reasoning prefix optimization), TensorRT-LLM speculative decoding (EAGLE-3, MEDUSA), the NVIDIA Dynamo inference platform, and reasoning-inference performance comparisons among AMD MI325X and TPU v6/v7. Reusing system prompt + few-shot reasoning via prefix cache, KV-cache pagination, and the 2-4x latency-reduction recipe with draft model + verify are shown practically.
The eleventh module covers end to end the modern benchmark set evaluating reasoning models. AIME 2024-2026 (high-school olympiad — symbolic checker), MATH-500 (Hendrycks), GPQA Diamond (graduate-level science), FrontierMath (Tao 2024 — hardest math), ARC-AGI 1 and 2 (Chollet abstract reasoning), LiveCodeBench (timestamp-aware code), SWE-bench Verified (solving real GitHub PRs), Codeforces ELO, IOI 2024, HumanEval+ — each is analyzed in detail. Saturation analysis (benchmarks still open after o3 + R1 in 2026), data-contamination detection (canary strings + n-gram analysis), pass@k / cons@N / maj@N metrics, and the discipline of producing custom domain benchmarks (legal, healthcare, finance) are established.
In the capstone module, each participant builds an end-to-end reasoning system for their own domain: scenario selection (math tutor, code-debug agent, legal reasoner, clinical triage system, financial-modeling assistant, or the participant's own use case); model selection (Claude Opus 4.7 + Extended Thinking, OpenAI o3/o4, DeepSeek R1, Qwen3, distilled S1/LIMO); reasoning pattern (native reasoning, CoT, ToT, MCTS, hybrid); production-inference stack (vLLM, SGLang, or TensorRT-LLM); evaluation framework (custom domain benchmark + AIME/MATH-500 baseline); a 90-day operational roadmap (cost monitoring, thinking-budget tuning, reasoning-router optimization). By the end of the training, participants reach a level of technical competence to dissect the internals of OpenAI o3/o4, DeepSeek R1, Gemini 2.5 Deep Think, Claude Extended Thinking, and Qwen3; apply the test-time compute scaling laws to optimize their compute budget; build an R1-style RL reasoning-emergence pipeline; produce their own reasoning model with distillation recipes (S1, LIMO, Bespoke-Stratos); implement MCTS, ToT, Self-Refine, and Reflexion patterns; serve reasoning models with vLLM / SGLang / TensorRT-LLM; and evaluate with the modern benchmark set. The training consists of 3 days, 12 modules, and over 100 hands-on lessons.