AI Papers of the Week
Every paper worth reading in AI, hand-picked one week at a time.
Thinking Mid-training: RL of Interleaved Reasoning
Meta FAIR addresses the gap between pretraining (no explicit reasoning) and post-training (reasoning-heavy) with an intermediate SFT+RL mid-training phase. The approach annotates pretraining data with interleaved reasoning traces, then uses supervised fine-tuning followed by RL to teach models when and how to think during continued pretraining. Applied to Llama-3-8B, the full pipeline achieves a 3.2x improvement on reasoning benchmarks compared to direct RL post-training, demonstrating that reasoning benefits from being trained as native behavior early in the pipeline.
MemFactory
MemFactory introduces the first unified, highly modular training and inference framework specifically designed for memory-augmented AI agents. It abstracts the memory lifecycle into atomic, plug-and-play components using a “Lego-like” architecture, natively integrating Group Relative Policy Optimization (GRPO) to fine-tune internal memory management strategies. The framework decomposes memory into mixable components that support recent approaches including Memory-R1, RMM, and MemAgent out of the box, achieving relative gains of up to 14.8% compared to baseline models.
Composer 2
Cursor releases the technical report for Composer 2, a specialized model designed for agentic software engineering that demonstrates strong long-term planning and coding intelligence while maintaining efficiency for interactive use. The report details a process for training domain-specialized models that starts with continued pretraining and scales up with reinforcement learning.
PivotRL
PivotRL is a turn-level reinforcement learning algorithm from NVIDIA designed to tractably post-train large language models for long-horizon agentic tasks. The method operates on existing SFT trajectories, combining the compute efficiency of supervised fine-tuning with the out-of-domain accuracy of end-to-end RL. PivotRL identifies “pivots,” informative intermediate turns where sampled actions exhibit high variance in outcomes, and focuses training signal on these critical decision points. The approach achieves +4.17% higher in-domain accuracy and +10.04% higher out-of-domain accuracy compared to standard SFT, while matching end-to-end RL accuracy with 4x fewer rollout turns. PivotRL is adopted by NVIDIA’s Nemotron-3-Super-120B-A12B as the workhorse for production-scale agentic post-training.
KARL
Databricks presents KARL, a system for training enterprise search agents via reinforcement learning that achieves state-of-the-art performance across a diverse suite of hard-to-verify agentic search tasks. The work also introduces KARLBench, a new evaluation framework spanning six search domains.
Memex(RL)
As tasks get longer and more complex, LLM agents lose track of what they have learned, what they have tried, and what still needs to be done. Memex(RL) introduces an indexed experience memory mechanism that scales agent capability on long-horizon tasks without discarding evidence or blowing up the context window.
Discovering Multi-Agent Learning Algorithms with LLMs
Google DeepMind uses AlphaEvolve, an evolutionary coding agent powered by LLMs, to automatically discover new multi-agent learning algorithms for imperfect-information games. Rather than relying on manual algorithm design, the system navigates vast algorithmic design spaces and discovers non-intuitive mechanisms that outperform state-of-the-art baselines.
AgentConductor
AgentConductor introduces a reinforcement learning-enhanced multi-agent system for code generation that dynamically generates interaction topologies based on task characteristics. Rather than using fixed communication patterns between agents, an LLM-based orchestrator adapts the topology to match problem complexity, achieving state-of-the-art accuracy across five code generation datasets.
CoT Faithfulness via REMUL
Researchers propose REMUL, a training approach for making chain-of-thought reasoning more faithful and monitorable. A speaker model generates reasoning traces that multiple listener models attempt to follow and complete, using RL to reward reasoning that is understandable to other models. Tested across BIG-Bench Extra Hard, MuSR, ZebraLogicBench, and FOLIO, REMUL improves three faithfulness metrics while also boosting overall accuracy, producing shorter and more direct reasoning chains.
GLM-5
GLM-5 is a foundation model from Zhipu AI designed to transition from vibe coding to agentic engineering. The model introduces novel asynchronous agent RL algorithms that separate generation from training for improved efficiency, and uses DSA technology to reduce computational requirements while preserving long-context understanding.
LLaDA 2.1
Ant Group releases LLaDA 2.1, a major upgrade to discrete diffusion language models that breaks the speed-quality trade-off through Token-to-Token (T2T) editing. By weaving token editing into the conventional Mask-to-Token decoding scheme, LLaDA 2.1 introduces two configurable modes: Speedy Mode for aggressive throughput and Quality Mode for benchmark-leading accuracy. The release also includes the first large-scale RL framework for diffusion LLMs.
SkillRL
SkillRL introduces a recursive skill-augmented RL framework that bridges the gap between raw experience and policy improvement through automatic skill discovery. Instead of storing noisy raw trajectories, SkillRL distills experience into reusable high-level behavioral patterns and evolves them alongside the agent policy during training.
InftyThink+
InftyThink+ is an end-to-end RL framework for infinite-horizon reasoning that optimizes the entire iterative reasoning trajectory. Standard long chain-of-thought suffers from quadratic cost, context length limits, and lost-in-the-middle degradation. InftyThink+ addresses all three by letting models autonomously decide when to summarize, what to preserve, and how to resume, trained through trajectory-level reinforcement learning.
TinyLoRA
This paper from Meta FAIR asks how small a LoRA adapter can get and still teach a model to reason. The answer: remarkably small. The authors propose TinyLoRA, a method that scales low-rank adapters down to as few as one trainable parameter by projecting through fixed random tensors and sharing weights across all modules. The key insight is that RL makes fundamentally more information-dense updates than SFT, enabling effective learning with orders of magnitude fewer parameters.
Reinforcement Learning via Self-Distillation
This paper introduces Self-Distillation Policy Optimization (SDPO), an on-policy RL algorithm that converts rich textual feedback from verifiable environments into dense credit assignment without requiring an external teacher model. SDPO uses the current model conditioned on feedback as a “self-teacher” to retrospectively identify mistakes in its own rollouts, substantially outperforming GRPO across scientific reasoning, tool use, and competitive programming.
Communication Methods in Multi-Agent RL
A systematic survey of 29 papers reviewing how agents coordinate in multi-agent reinforcement learning, covering fully connected message passing, implicit communication, attention-based selective methods, graph-based relational approaches, and role-based hierarchical frameworks. The analysis reveals that attention- and graph-based methods dominate recent research, while implicit communication is seeing renewed interest for its scalability in decentralized settings where explicit channels are infeasible.
TTT-Discover: Learning to Discover at Test Time
TTT-Discover introduces test-time training for scientific discovery, performing reinforcement learning at test time so the LLM can continue to train with experience specific to the test problem. Unlike prior work like AlphaEvolve that prompts a frozen LLM, this approach enables the model itself to improve while attempting to solve hard problems.
Reasoning Models Generate Societies of Thought
This paper reveals that enhanced reasoning in models like DeepSeek-R1 and QwQ-32B emerges not from extended computation alone, but from simulating multi-agent-like interactions - a “society of thought” - enabling diversification and debate among internal cognitive perspectives with distinct personality traits and domain expertise.
Self-Evolving Search Agents Without Training Data
Dr. Zero introduces a framework for developing multi-turn search agents that improve themselves autonomously without labeled training data. A proposer generates diverse questions to train a solver initialized from the same base model, creating a self-evolution loop with automated curriculum difficulty scaling.
Unified Long-Term and Short-Term Memory for LLM Agents
AgeMem introduces a unified framework that integrates both long-term and short-term memory operations into an LLM agent’s decision-making policy. The system enables agents to autonomously determine what and when to store, retrieve, update, summarize, or discard information by exposing memory operations as tool-based actions.
MemRL
MemRL enables LLM agents to improve continuously without retraining by separating a frozen model’s reasoning from an evolving memory system. A Two-Phase Retrieval mechanism filters candidates by semantic relevance, then ranks them using learned Q-values that improve through trial-and-error, outperforming existing methods on HLE, BigCodeBench, ALFWorld, and Lifelong Agent Bench.
Nemotron-Cascade
Nemotron-Cascade introduces cascaded domain-wise reinforcement learning (Cascade RL) to build general-purpose reasoning models capable of operating in both instruct and deep thinking modes. Rather than blending heterogeneous prompts from different domains, Cascade RL orchestrates sequential, domain-wise RL stages that reduce engineering complexity while delivering state-of-the-art performance.
GDPO
GDPO addresses a critical flaw in training language models with multiple competing objectives. The authors discover that when applying Group Relative Policy Optimization (GRPO) to multi-reward settings, normalizing distinct rollout reward combinations causes them to collapse into identical advantage values, degrading training signal quality and stability.
Training AI Co-Scientists Using Rubric Rewards
This paper from Meta Superintelligence Labs presents a scalable method to train language models to generate better research plans without expensive human supervision or real-world execution. The approach automatically extracts research goals and goal-specific grading rubrics from scientific papers, then uses reinforcement learning with self-grading to improve plan generation.