AI Papers of the Week
Every paper worth reading in AI, hand-picked one week at a time.
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.
STRUCTUREDAGENT
STRUCTUREDAGENT introduces a hierarchical planning framework for long-horizon web tasks using dynamic AND/OR trees. The framework separates planning responsibilities: the system constructs and maintains the planning tree while the LLM is invoked only for local operations like node expansion or repair. A structured memory module tracks candidate solutions to improve constraint satisfaction. Results on WebVoyager, WebArena, and custom shopping benchmarks show improved performance over standard LLM-based web agents, with the added benefit of interpretable hierarchical plans that enable easier debugging and human intervention.
AgentIR
Deep research agents generate explicit reasoning before every search call, but existing retrievers completely ignore these rich signals about search intent and problem context. AgentIR introduces reasoning-aware retrieval that jointly embeds the agent’s reasoning trace alongside its query, along with DR-Synth, a data synthesis method for generating training data from standard QA datasets. On BrowseComp-Plus, AgentIR-4B achieves 68% accuracy with Tongyi-DeepResearch compared to 50% with conventional embedding models twice its size and 37% with BM25.
Think Harder or Know More
This paper investigates transformer models featuring both adaptive per-layer looping, where each block learns to iterate its hidden state via a learned halting mechanism, and gated memory banks that provide additional learned storage. The key finding is that looping primarily benefits mathematical reasoning while memory banks help recover performance on commonsense tasks. Combining both mechanisms yields a model that outperforms an iso-FLOP baseline with three times the number of layers on math benchmarks. Analysis of model internals reveals layer specialization: early layers loop minimally and access memory sparingly, while later layers do both more heavily.
Theory of Mind in Multi-Agent LLMs
This work introduces a multi-agent architecture combining Theory of Mind (ToM), Belief-Desire-Intention (BDI) models, and symbolic solvers for logical verification, evaluating it on resource allocation problems across multiple LLMs. The central finding is counterintuitive: simply adding cognitive mechanisms does not automatically improve coordination.
ParamMem
Self-reflection enables language agents to iteratively refine solutions, but models tend to generate repetitive reflections that add noise instead of useful signal. ParamMem introduces a parametric memory module that encodes cross-sample reflection patterns into model parameters, enabling diverse reflection generation through temperature-controlled sampling.
Reaching Agreement Among LLM Agents
This paper introduces Aegean, a consensus protocol that frames multi-agent refinement as a distributed consensus problem. Rather than static heuristic workflows with fixed loop limits, Aegean enables early termination when sufficient agents converge, achieving 1.2-20x latency reduction across four mathematical reasoning benchmarks while maintaining answer quality within 2.5%. The consensus-aware serving engine performs incremental quorum detection across concurrent agent executions, cutting wasted compute on stragglers.
Diagnosing Agent Memory
This paper introduces a diagnostic framework that separates retrieval failures from utilization failures in LLM agent memory systems. Through a 3x3 factorial study crossing three write strategies with three retrieval methods, the authors find that retrieval is the dominant bottleneck, accounting for 11-46% of errors, while utilization failures remain stable at 4-8% regardless of configuration. Hybrid reranking cuts retrieval failures roughly in half, delivering larger gains than any write strategy optimization.
Deep-Thinking Tokens
Google researchers challenge the assumption that longer outputs indicate better reasoning. They introduce deep-thinking tokens, a metric that identifies tokens where internal model predictions shift significantly across layers before stabilizing. Unlike raw token count, which negatively correlates with accuracy (r = -0.59), the deep-thinking ratio shows a robust positive correlation (r = 0.683).
Codified Context
Single-file AGENTS.md manifests don’t scale beyond modest codebases. A 1,000-line prototype can be fully described in a single prompt, but a 100,000-line system cannot. This paper presents a three-component codified context infrastructure developed during construction of a 108,000-line C# distributed system, evaluated across 283 development sessions.
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.
Evaluating AGENTS.md
This research evaluates whether AGENTS.md files, the repository-level context files that developers write to help AI coding agents understand their codebases, actually improve agent performance. Testing four coding agents (Claude Code with Sonnet-4.5, Codex with GPT-5.2 and GPT-5.1 mini, and Qwen Code with Qwen3-30b-coder), the findings are counterintuitive.
PAHF
Meta introduces PAHF (Personalized Agents from Human Feedback), a continual agent personalization framework that addresses a critical gap: most AI agents cannot adapt to individual user preferences that evolve over time. PAHF couples explicit per-user memory with both proactive and reactive feedback mechanisms.
Doc-to-LoRA
Sakana AI introduces Doc-to-LoRA (D2L), a lightweight hypernetwork that meta-learns to compress long documents into LoRA adapters in a single forward pass. Instead of processing long contexts through expensive quadratic attention, D2L converts the document into parameter-space representations that the target LLM can use without re-consuming the original text.
Emergent Socialization in AI Agent Society
A study on Moltbook, a social network with no humans where all participants are LLM-driven agents, challenges the assumption that scale and interaction density alone produce meaningful social dynamics. The researchers find that while global semantic content stabilizes quickly, individual agents maintain diversity without converging, displaying strong individual inertia and minimal adaptive response to interaction partners.
Lossless Context Management (LCM)
Lossless Context Management (LCM) is a deterministic architecture for LLM memory that outperforms Claude Code on long-context tasks. Benchmarked on the OOLONG eval using Opus 4.6, the LCM-augmented coding agent Volt achieves higher scores than Claude Code at every context length between 32K and 1M tokens. LCM extends the recursive paradigm pioneered by Recursive Language Models (RLMs) with two engine-managed mechanisms.
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.
MemoryArena
MemoryArena introduces a benchmark for evaluating how agents utilize memory across multiple interconnected sessions. The key finding is that scoring well on memory recall does not mean an agent can actually use that memory to take correct actions across sessions. Models with near-saturated performance on existing benchmarks like LoCoMo perform poorly in agentic multi-session settings.
MAPLE
MAPLE proposes separating memory, learning, and personalization into specialized sub-agents rather than treating them as a unified capability. The framework achieves a 14.6% improvement in personalization scores over stateless baselines and increases trait incorporation from 45% to 75%, validated through the MAPLE-Personas benchmark.
SkillsBench
SkillsBench evaluates whether LLM agents can generate their own procedural knowledge across 86 tasks spanning 11 domains, with curated Skills and deterministic verifiers. Testing 7 agent-model configurations over 7,308 trajectories, the benchmark reveals a critical gap: agents benefit enormously from consuming procedural knowledge but cannot reliably author it themselves.
LongCLI-Bench
LongCLI-Bench benchmarks how well AI agents handle complex, extended tasks through command-line interfaces. Across 20 demanding tasks spanning initial development, feature expansion, error resolution, and code optimization, leading agents succeed less than 20% of the time. The study finds that most failures occur early in task execution, and human-agent collaboration through plan injection and interactive guidance yields significantly greater improvements than automated self-correction alone.
CogRouter
CogRouter enables adaptive reasoning depth for LLM agents by dynamically selecting from four hierarchical cognitive levels at each step, from instinctive responses to strategic planning. Using confidence-aware advantage reweighting during training, Qwen2.5-7B with CogRouter achieves 82.3% success rate on agentic benchmarks, substantially outperforming larger models while consuming fewer tokens by skipping heavy reasoning on routine steps.
Team of Thoughts
Team of Thoughts presents a multi-agent framework for efficient test-time scaling through orchestrated tool calling. The system uses an orchestrator tool design where agents with different capabilities are coordinated by a calibrated orchestrator. With self-assessment for tool agents and orchestrator calibration for identifying superior coordination models, Team of Thoughts achieves 96.67% on AIME24 and 72.53% on LiveCodeBench, substantially exceeding homogeneous baselines.