Agentic AI

Researchers argue that the infrastructure surrounding an LLM, known as the harness, dictates agent performance and security but lacks a standard definition. They decompose this layer into eight comparable functional components and propose a framework to record the specific harness-model pairing used in evaluations. This documentation is critical because harness variations can cause multi-fold score discrepancies for identical models.

• Harness configuration significantly impacts reported agent scores, causing multi-fold variances for the same model.
• Proposed framework decomposes the agent harness into eight distinct, comparable functional components.
• Recording the harness-model pairing is necessary for reproducibility and accurate benchmarking.
• Existing model and agent cards fail to capture the operational infrastructure sequence, leaving security and eval gaps.

LanG 2.0 is an agentic security system that uses small, energy-efficient local models instead of frontier LLMs for threat detection and correlation. It relies on structural safeguards like entity-anchored correlation, human approval workflows, and tamper-evident ledgers rather than raw model accuracy to ensure safety. The system significantly improved campaign correlation precision and rule acceptance rates compared to its predecessor while maintaining affordability.

• Small local models can replace frontier LLMs in SOC workflows when paired with structural safeguards
• Campaign correlation precision improved from 18.75% to 51.81% over the previous version
• Human approval gates and tamper-evident ledgers are key to safe autonomous rule generation
• Energy efficiency and cost reduction are primary drivers for using local models over cloud APIs

Researchers introduce Mage-VL-Cowork, a 4B-parameter multimodal agent designed to unify visual GUI interaction with code and terminal execution. The model is trained through a five-stage curriculum leveraging a novel data synthesis framework called Mage-CUA-Forge. This framework generates diverse desktop and web tasks by evolving teacher harnesses rather than relying on static distillation, significantly reducing training data costs.

• Unifies visual GUI control with code/terminal execution in a single 4B policy
• Mage-CUA-Forge synthesizes 293k tasks across 6,915 skills using evolved harnesses
• Reduces per-task data generation expenditure by 82.6% via reusable workflows
• Progressive curriculum spans midtraining, SFT, and RL for hybrid capabilities
• Moves beyond static teacher distillation by instantiating specialized CUA harnesses

Researchers introduce CLR-Bench to determine if self-evolving agents genuinely learn hidden rules or merely memorize corrections as growing exception lists. The benchmark tests agents on cumulative hidden rules tied to real-world web data, using opaque labels to prevent simple lookup. Probes confirm base models possess the underlying knowledge, isolating the agent's ability to infer and apply the latent grouping logic.

• Distinguishes true generalization from rote memorization of exceptions in agent state
• Uses cumulative hidden rules on opaque labels to block direct web retrieval
• Verifies base model knowledge via probes to isolate agent learning capability
• Highlights risk of unbounded state growth in agents relying on exception lists

Current test-time scaling methods for LLM agents struggle in shared, persistent enterprise environments because they assume parallel execution is possible. The authors introduce Agentic World Modeling (AWM), which avoids direct execution and instead predicts next observations to simulate agent trajectories. This approach addresses the limitations of existing generative world models that compress state into single prompts, offering a way to handle partial observability and irreversible side effects.

• Solves parallel execution bottleneck in shared, persistent enterprise environments
• Replaces direct execution with next-observation prediction to avoid side effects
• Overcomes state compression limits of prior generative world models
• Enables reliable simulation of transitions under partial observability
• Allows test-time scaling without costly or infeasible environment replication

TraceSpec addresses the limitations of aggregate speculative decoding metrics in LLM agents by connecting trajectory evaluation with adaptive draft training. The framework combines dynamic execution with fixed-prefix replay and introduces a new corpus of 10,000 training and 2,500 test states for browsing and software engineering tasks. Analysis reveals that while overall acceptance rates may improve with context length, performance often degrades in specific late-stage browsing actions, highlighting the need for phase-aware optimization.

• Aggregate acceptance metrics mask performance drops in late-stage agent actions
• TraceSpec-Traj provides 12,500 states for training and testing agent-specific drafts
• Workload type and execution phase significantly influence speculative decoding efficacy
• Adaptive draft training requires trajectory-aware supervision rather than global metrics

Agentic Soft-Token Compression (ASTC) introduces a training-free method to manage context in long-running LLM agents by using internal soft tokens instead of external hard-token rewriting. Unlike prior approaches that rely on separate compressors which may misalign with the model's representation, ASTC leverages the LLM's native recurrent dynamics to identify and compress already absorbed information. This approach aims to reduce computational costs and preserve decision-relevant evidence without lossy external selection.

• Eliminates need for external compressors or context-management agents
• Uses model's native recurrent dynamics for alignment with internal representations
• Training-free method reduces computational overhead in long-horizon tasks
• Preserves decision-relevant evidence better than lossy hard-token rewriting