Agentic AI

This paper introduces Multi-Head Latent Control, a method that allows large language models to make agentic decisions—such as tool invocation, deferral, or abstention—by analyzing internal latent signals rather than relying on external prompts or orchestration. The approach addresses the fragility and maintenance costs of current input-side routing techniques by embedding control logic directly into the model's generation process. This creates a unified interface for decision-making that adapts dynamically to the model's internal reasoning state.

• Shifts agent control from external prompt routing to internal latent analysis for better stability.
• Enables dynamic decisions like tool use or deferral based on real-time model states.
• Reduces maintenance overhead as model backbones evolve by decoupling control from input signals.
• Provides a unified interface for complex agentic behaviors without heavy fine-tuning.

Amazon RDS for Db2 is now available in the Asia Pacific (Thailand, Malaysia, Taipei), Mexico (Central), and Canada West (Calgary) regions. This service allows customers to deploy Db2 databases quickly with auto-configured parameters for performance. It supports Multi-AZ synchronous replication for high availability and offers both hourly pay-as-you-go and Bring Your Own License options.

• Db2 workloads can now be deployed in 5 new regions, reducing latency for APAC and North American customers.
• Multi-AZ configurations provide high availability via synchronous replication to a standby instance.
• Flexible licensing is supported: AWS Marketplace hourly rates or BYOL models are available in these regions.

A new benchmark inspired by cognitive psychology set-shifting tests how LLM agents handle hidden changes in tool reliability during active sessions. The evaluation framework injects redundant tool libraries where the optimal choice silently swaps, revealing that agents quickly revert to narrow, recurring routines rather than adapting fluidly. This highlights a rigidity in current agent architectures when facing dynamic, hidden environmental shifts.

• Agents default to rigid routines after hidden reliability shifts, limiting adaptability
• New benchmark isolates set-shifting capability using branched evaluation schedules
• Redundant tool libraries expose how models prioritize familiar vs. optimal paths
• Current agents struggle to dynamically re-optimize tool selection mid-session

The STOCKTAKE paper introduces a 26-week supply-chain replenishment benchmark designed to measure the gap between an LLM agent's perception and its subsequent actions. Unlike existing evaluations that conflate misreading the world with failing to act, this benchmark uses a factored partially observable Markov decision process with hidden factors. It enables a fair reference policy via an exact Bayes filter, allowing precise attribution of failure modes in multi-week decision tasks.

• STOCKTAKE isolates 'knowing-doing' gaps in LLM agents for supply-chain tasks.
• Uses a 26-week POMDP with six hidden factors to prevent privileged info leakage.
• Exact Bayes filter reference policy enables accurate failure attribution.
• Moves beyond final cost metrics to diagnose specific agent dysfunction.

This research introduces tool efficiency as a direct quantitative metric for evaluating how often LLM agents make useful tool calls. It defines marginal tool utility to assess individual tool calls, determining if a tool can be removed without hurting accuracy. The study uses an LLM-as-a-Judge approach to calculate these metrics, moving beyond accuracy as a proxy for efficiency.

• Tool efficiency measures the rate of useful tool calls in agent trajectories.
• Marginal tool utility identifies tools safe to remove without accuracy loss.
• LLM-as-a-Judge determines the utility sign for each specific tool call.
• This approach shifts focus from accuracy proxies to direct efficiency metrics.