DeepSeek-V3.2 is an open large language model that balances high computational efficiency with superior reasoning and agent capabilities. Key innovations include DeepSeek Sparse Attention (DSA) for reduced complexity in long contexts, a scalable reinforcement learning framework achieving GPT-5-level performance, and a large-scale agentic task synthesis pipeline for improved generalization in tool-use scenarios.
Combines internalizing general skills with task-specific skill utilization via a difficulty-aware router to improve in-distribution and out-of-distribution performance for agentic RL. Uses privileged distillation for hard tasks and diagnostic probing for easy tasks; evaluated on ALFWorld and WebShop.
Uses search-agent reading traces and tiered distractors to train LLMs for long-context, multi-hop reasoning, and introduces a rubric reward that supervises entity-level steps (applied only to correct finals). Improves evidence-grounded reasoning and resists reward hacking across 4B–30B models.
Analyzes how single-domain RL fine-tuning on LLMs induces cross-domain interference and shows this damage concentrates in a low-dimensional shared conflict subspace; proposes a local perturbation theory and short domain "refresh" procedures that selectively recover earlier domains with minimal collateral loss.
A 20B retrieval subagent trained with reinforcement learning inside a stateful search harness that externalizes recoverable search state (candidate pool, curated evidence, verification records). The harness lets the policy focus on semantic search decisions, improving curated recall and transfer robustness.
The paper by DeepMind introduced Deep Q-Networks (DQN), the first deep learning model to learn control policies directly from raw pixel input using reinforcement learning. By combining Q-learning with convolutional neural networks and experience replay, DQN achieved superhuman performance on several Atari 2600 games without handcrafted features or game-specific tweaks. Its impact was profound: it proved deep learning could master complex tasks with sparse, delayed rewards, catalyzing the modern wave of deep reinforcement learning research and paving the way for later breakthroughs like AlphaGo.
The paper introduced AlphaGo, the first program to defeat a human professional Go player without handicap. It combined deep neural networks — trained with supervised learning and reinforcement learning — with Monte Carlo tree search (MCTS), enabling efficient move selection and board evaluation in Go’s massive search space. AlphaGo’s victory against European champion Fan Hui marked a historic AI milestone, showcasing that combining learning-based policies with search can surpass prior handcrafted methods, reshaping both game AI and broader AI research directions.
