Training LLM agents for extremely long-horizon tasks remains an open challenge. Most agent training pipelines struggle with extended-duration trajectories. Context gets lost, rewards are sparse, and the learning signal degrades over long sequences. KLong tackles this with a
@dair_ai
-
Dynamic Multi-Agent Systems with Population Growth and Adaptation
By
–
Interesting new work on dynamic multi-agent systems. Traditional multi-agent RL assumes a fixed number of agents. But real-world scenarios often involve populations that grow, split, or shrink dynamically, such as cell division, team expansion, or organizational scaling.
-
Agent Memory Benchmarks Don’t Predict Real-World Performance
By
–
Agent memory benchmarks are misleading. Scoring well on memory recall doesn't mean an agent can actually use that memory to take correct actions across sessions. Models that achieve near-saturated performance on existing long-context memory benchmarks like LoCoMo perform poorly
-
Lossless Context Management Advances Agent Language Models
By
–
A paper worth paying close attention to. It presents Lossless Context Management (LCM), which reframes how agents handle long contexts. It outperforms Claude Code on long-context tasks. Recursive Language Models give the model full autonomy to write its own memory scripts. LCM
-
Agent World Model: Synthetic Environments for RL Agent Training
By
–
Training tool-use agents with RL requires diverse, executable environments. But these environments barely exist. This new research introduces Agent World Model (AWM), a fully synthetic pipeline that generates executable agentic environments at scale. Starting from high-level
-
WebClipper Framework Optimizes Web Agent Search Efficiency
By
–
// Efficient Evolution of Web Agents // Web agents waste a lot of compute on cyclic reasoning loops and unproductive exploration. This new research introduces WebClipper, a framework that models web agent search processes as state graphs and prunes them into minimal directed
-
LLM Agents Struggle With Multi-Step Scientific Tool Use
By
–
On evaluating multi-step scientific tool use in LLM agents. SciAgentGym provides an interactive environment with 1,780 specialized tools across 4 scientific disciplines. The core finding: even advanced models like GPT-5 see success rates drop sharply from 60.6% to 30.9% as
-
Top AI Papers of the Week: February 9-15
By
–
The Top AI Papers of the Week (February 9-15) – Gaia2
– SkillRL
– LLaDA 2.1
– EchoJEPA
– InftyThink+
– AdaptEvolve
– AgentSkiller Read on for more: -
Improving Efficiency of Evolutionary AI Agents
By
–
// Improving Efficiency of Evolutionary AI Agents // Evolutionary AI agents are powerful but can be wasteful. Systems, inspired by AlphaEvolve and OpenEvolve, iteratively generate, mutate, and refine candidate solutions using LLMs. However, every refinement step invokes the
-
Shallow Exploration Trap in Test-Time Reasoning Models
By
–
Test-time reasoning models often converge too early. Achieving broader reasoning coverage requires longer sequences, yet the probability of sampling such sequences decays exponentially during autoregressive generation. The authors call this the "Shallow Exploration Trap." This