The DAIR Daily - AI Papers of the Week

ALMA: Meta-Learning Memory for Agents

\n In a groundbreaking shift for agentic systems, researchers from Jeff Clune’s group have unveiled ALMA (Automated meta-Learning of Memory designs for Agentic systems). The system introduces a \"Meta Agent\" that autonomously discovers memory architectures through open-ended exploration in code space. By searching over database schemas, retrieval mechanisms, and update strategies expressed as executable code, ALMA consistently outperforms all human-designed memory baselines across four sequential decision-making benchmarks. This represents a significant step toward AI systems that learn to be continual learners.\n

\n Read the full report at arXiv.org →\n

\n\n

LLaDA 2.1: Speed Meets Quality

\n Ant Group has released LLaDA 2.1, a major upgrade to discrete diffusion language models. By introducing Token-to-Token (T2T) editing, the model allows for aggressive drafting followed by refinement. The \"Flash\" variant hits an astounding 892 tokens per second, fundamentally changing the speed-quality trade-off in LLM decoding.\n

\n Source: arXiv:2602.08676 →\n

\n\n

SkillRL: Recursive Skill Discovery

\n SkillRL introduces a recursive framework that distills raw experience into reusable behavioral patterns. This \"SkillBank\" approach reduces token footprints while preserving reasoning utility, achieving state-of-the-art performance with an 89.9% success rate on the ALFWorld benchmark.\n

\n Source: arXiv:2602.08234 →\n

\n\n

InftyThink+: Infinite Reasoning

\n Addressing the quadratic cost of long chain-of-thought reasoning, InftyThink+ uses trajectory-level RL to let models autonomously decide when to summarize and resume. The framework yielded a 21% accuracy gain on AIME24, outperforming vanilla long-CoT methods.\n

\n Source: arXiv:2602.06960 →\n

\n\n

Agyn: Software Engineering as an Org

\n Agyn models software development as an organizational process with specialized agents (Manager, Researcher, Engineer, Reviewer). It resolves 72.2% of SWE-bench tasks without specific tuning, suggesting organizational design matters as much as model size.\n

\n Source: arXiv:2602.01465 →\n

\n\n

Research Briefs

EchoJEPA

A foundation model for echocardiography trained on 18M scans, showing 20% improvement in cardiac assessment and exceptional robustness to noise.

\n arXiv:2602.02603 →\n

AdaptEvolve

Confidence-driven model routing cuts evolutionary agent costs by 38% while retaining 97.5% accuracy, making iterative refinement viable for production.

\n arXiv:2602.11931 →\n

Gaia2

Meta FAIR's new dynamic agent benchmark forces agents to handle temporal pressure and multi-agent coordination in changing environments.

\n arXiv:2602.11964 →\n

AgentArk

Distilling multi-agent debate into a single LLM for 4.8% better reasoning, approaching full multi-agent performance at a fraction of the cost.

\n arXiv:2602.03955 →\n

AgentSkiller

A 14B model beating GPT-o3 on tool-use benchmarks via high-quality synthetic trajectories, proving data quality over parameter count.

\n arXiv:2602.09372 →\n

Code