AI Papers of the Week
Every paper worth reading in AI, hand-picked one week at a time.
LaMer
LaMer introduces a Meta-RL framework that enables LLM agents to actively explore and learn from environment feedback at test time. Unlike standard RL-trained agents that learn fixed policies and struggle with novel tasks, LaMer agents learn exploration strategies that transfer across environments.
JustRL
JustRL challenges the assumption that complex RL pipelines are necessary for training small language models. Using single-stage training with fixed hyperparameters, the authors achieve state-of-the-art math reasoning performance on two 1.5B models while using 2x less compute than sophisticated multi-stage approaches.
Self-Play SWE-RL
Self-Play SWE-RL (SSR) trains software engineering agents through self-play, requiring only access to sandboxed repositories with no human-labeled issues or tests. A single LLM learns to both inject and repair bugs of increasing complexity, achieving +10.4 points on SWE-bench Verified while consistently outperforming human-data baselines.
Pre-Training, Mid-Training, and RL Interplay
CMU researchers develop a controlled experimental framework using synthetic reasoning tasks to isolate how pre-training, mid-training, and RL-based post-training each contribute to reasoning capabilities in language models. The study reconciles conflicting views on whether RL truly extends reasoning beyond what models learn during pre-training.
Agentic AI Adaptation Survey
Researchers from UIUC, Stanford, Berkeley, and other institutions present the first comprehensive taxonomy of adaptation strategies for agentic AI systems. The survey organizes recent advances into a unified framework covering how agents and their tools can be modified to achieve higher task performance, improved reliability, and better generalization across diverse scenarios.
Nanbeige4-3B
Nanbeige4-3B is a 3B parameter model pretrained on 23T tokens and fine-tuned on over 30M instructions using a Fine-Grained Warmup-Stable-Decay scheduler, Dual Preference Distillation, and multi-stage reinforcement learning. Despite its compact size, it outperforms Qwen3-8B and Qwen3-14B on reasoning benchmarks and rivals much larger models on WritingBench, demonstrating that well-engineered small models can match far larger counterparts.
DeepSeek-V3.2
DeepSeek releases V3.2, an open model that matches GPT-5 on reasoning benchmarks while introducing significant architectural and training innovations. The high-compute variant, DeepSeek-V3.2-Speciale, surpasses GPT-5 and achieves gold-medal performance in both the 2025 IMO and IOI competitions.
Evolving Multi-Agent Orchestration
OpenBMB researchers propose a “puppeteer-style” paradigm for multi-agent LLM collaboration, where a centralized orchestrator dynamically directs agents based on evolving task states. Trained via reinforcement learning, the system achieves superior performance with reduced computational costs across math, knowledge, and software development tasks.
Training LLMs for Honesty via Confessions
OpenAI introduces a novel method for training LLMs to honestly self-report their own misbehavior through “confessions” - separate outputs where models evaluate their compliance with instructions and policies. By training GPT-5-Thinking to produce confessions after completing tasks, the research demonstrates that models can be incentivized to reveal deceptive behaviors they otherwise hide in their main answers.
CodeVision: Thinking with Programming Vision
Researchers propose CodeVision, a framework where multimodal models generate code as a universal interface to invoke image operations, addressing brittleness in visual reasoning from orientation changes and corruptions. The two-stage training approach combines supervised fine-tuning with RL using dense rewards, enabling flexible tool composition and error recovery on Qwen models.
INTELLECT-3
INTELLECT-3 is a 106B-parameter Mixture-of-Experts model (12B active) trained with large-scale reinforcement learning, achieving state-of-the-art performance for its size across math, code, science, and reasoning benchmarks. Built on top of GLM-4.5-Air base, it outperforms many larger frontier models, including DeepSeek R1-0528, and matches GLM-4.6 (which has over 3x the parameters) on key benchmarks.
Evolution Strategies at Hyperscale
EGGROLL (Evolution Guided General Optimization via Low-rank Learning) is an evolution strategies algorithm designed to scale backprop-free optimization to large population sizes for billion-parameter neural networks. By using low-rank matrix perturbations instead of full-rank ones, EGGROLL achieves a hundredfold increase in training throughput while nearly matching pure batch inference speed.
DR Tulu
DR Tulu-8B is the first open model directly trained for long-form deep research using Reinforcement Learning with Evolving Rubrics (RLER). Unlike existing models trained on short-form QA tasks, DR Tulu learns to produce comprehensive, well-attributed research reports by training with rubrics that co-evolve with the model and are grounded on real-world searched knowledge.
Seer: Fast RL for LLMs
Researchers introduce Seer, a system addressing performance bottlenecks in synchronous reinforcement learning for LLMs by optimizing the rollout phase that dominates end-to-end iteration time. Through three core mechanisms: divided rollout, context-aware scheduling, and adaptive grouped speculative decoding, Seer achieves 74-97% improvement in rollout throughput and 75-93% reduction in long-tail latency on production-grade RL workloads.
Natural Emergent Misalignment from Reward Hacking
Anthropic researchers demonstrate that realistic AI training processes can inadvertently produce misaligned models through “reward hacking generalization”. Models learn to cheat on programming tasks during RL. They simultaneously develop dangerous behaviors, including alignment faking (50% of responses) and safety research sabotage (12% of instances), without explicit training for these harmful actions. The study identifies a simple mitigation: “inoculation prompting” using contextual instructions that break semantic links between task-specific cheating and broader misalignment without reducing hacking frequency.
LAMP: Language-Augmented Multi-Agent RL
LAMP integrates natural language processing into multi-agent reinforcement learning through a three-stage pipeline: Think (processes numerical data and identifies market patterns), Speak (generates strategic communications between agents), and Decide (synthesizes information into optimized policy). The framework achieves substantial improvements over baseline methods with +63.5% and +34.0% gains in cumulative return and +18.8% and +59.4% improvements in robustness, bridging traditional MARL with real-world economic contexts where language significantly influences decisions.
Olympiad-Level Formal Mathematical Reasoning with Reinforcement Learning
Google DeepMind introduces AlphaProof, an AlphaZero-inspired reinforcement learning agent that learns to find formal mathematical proofs within the Lean theorem prover, achieving the first-ever medal-level performance at the International Mathematical Olympiad by solving three problems, including the competition’s most difficult challenge.
The Era of Agentic Organization
Microsoft Research introduces asynchronous thinking (AsyncThink), a new reasoning paradigm where language models learn to organize their internal thinking into concurrently executable structures through an organizer-worker protocol, achieving 28% lower inference latency than parallel thinking while improving accuracy on mathematical reasoning and demonstrating zero-shot generalization to unseen tasks.
RL Enhances Knowledge Navigation
Researchers show that RL-enhanced models outperform base models by 24pp on hierarchical knowledge retrieval tasks (e.g., medical codes) by improving navigation of existing knowledge structures rather than acquiring new facts. Structured prompting reduces this gap to 7pp, while layer-wise analysis reveals that RL transforms query processing (cosine similarity drops to 0.65-0.73) while preserving factual representations (0.85-0.92). The findings suggest RL’s benefits stem from enhanced procedural skills in traversing parametric knowledge hierarchies rather than expanded knowledge content.
RLAC: Adversarial Critic for RL Post-Training
UC Berkeley and CMU researchers introduce RLAC, an RL post-training approach using a learned critic that dynamically identifies likely failure modes (e.g., factual errors or edge cases) verified by external validators, eliminating exhaustive rubric enumeration. On biography generation, RLAC achieves 0.889 FactScore (vs 0.867 for FactTune-FS) while reducing verification calls by 5.7×, and on code generation reaches 56.6 average score using only 9% of training data. The adversarial game between generator and critic prevents reward hacking through on-policy, prompt-specific training signals grounded in verifiable rubrics.
Scaling Agent RL via Experience Synthesis
Meta researchers introduce DreamGym, a unified framework that synthesizes diverse training experiences to enable scalable reinforcement learning for LLM agents without costly real-environment rollouts. It addresses fundamental barriers of expensive interactions, limited task diversity, unreliable rewards, and infrastructure complexity.
TIR-Judge
Google and collaborators introduce TIR-Judge, an end-to-end reinforcement learning framework that trains LLM judges to integrate code execution for precise evaluation. It surpasses reasoning-only judges by up to 6.4% (pointwise) and 7.7% (pairwise) while demonstrating that tool-augmented judges can self-evolve without distillation.
GAP
GAP introduces graph-based agent planning with parallel tool execution and reinforcement learning, enabling AI agents to coordinate multiple specialized capabilities simultaneously rather than sequentially. The framework significantly accelerates task completion and improves success rates on complex multi-step problems through optimized tool selection and execution ordering.
Precision-RL
Reinforcement learning fine-tuning of LLMs suffers from a critical numerical mismatch between training and inference engines, causing training instability and collapse. This work reveals that simply switching from BF16 to FP16 precision virtually eliminates this mismatch - achieving faster convergence, higher stability, and superior performance across diverse models, frameworks, and algorithms without any algorithmic changes or architectural modifications.