Agentic AI

LayerRAG-Bench evaluates agentic RAG systems across nine models and 240 tasks, exposing reliability gaps beyond simple grounding. While schema normalization successfully resolves schema-drift errors, it fails to recover from stale evidence, missing tool outputs, denied permissions, or incorrect session context. The study highlights that evaluating groundedness alone leads to significant false positives when evidence is outdated or misaligned with the session state.

• Schema normalization fixes drift but ignores stale evidence or wrong-session context.
• Groundedness-only metrics produce false positives under stale or mismatched evidence.
• Benchmark covers 9 fault scenarios including auth denials and missing tool outputs.
• Evaluated across OpenAI, Anthropic, and Gemini models in 8 enterprise domains.

SkillSmith addresses the gap between textual knowledge composition and parametric skill consolidation by treating model weights as a reasoning modality. The approach allows LLMs to natively integrate weight-space libraries with past textual experiences for complex problem solving. This unifies two previously orthogonal mechanisms into a single framework for targeted performance improvements.

• Unifies text-based reflection and weight-space merging into one agentic framework.
• Treats model weights as a native modality for LLM reasoning and composition.
• Enables targeted performance gains by combining procedural and parametric skills.
• Moves beyond treating textual and parametric knowledge as separate pursuits.

ChronoMem introduces a version-control layer for LLM agent memory, addressing the brittleness of current forward-only systems. It commits whole-memory snapshots at each write and maintains structured version histories to enable semantic rollback. This approach allows agents to recover from corrections, concept drift, and memory corruption. The system is integrated into Google's open-source Agent Development Kit.

• Solves memory corruption and concept drift by enabling rollback to prior states
• Commits full memory snapshots at each write for granular version control
• Integrated into Google's open-source Agent Development Kit for immediate use
• Moves beyond forward-only evolution to support inspection and reverting

This research paper introduces a causal audit to verify whether latent multi-agent systems actually transmit and use task-relevant information. By applying controlled message replacements at the sender-receiver boundary, the study isolates whether end-task performance depends on the presence, content, or identity of the latent message. The findings highlight that high representational capacity does not guarantee the receiver utilizes the sender's encoded information.

• End-task performance alone cannot confirm if latent messages carry task-relevant signals.
• Controlled message replacements isolate sender contribution from receiver behavior.
• Five measurements quantify encoded info, receiver sensitivity, and task value.
• Represents a methodological shift from correlation to causation in MAS evaluation.

Researchers identify retrieval-equivalence collapse in RL search agents, where distinct query strings yield overlapping evidence sets. The proposed Harness-G framework structures retrieval to resolve this aliasing at the policy-environment interface. This approach stabilizes training by addressing formulation issues often overlooked by denser credit signals.

• Retrieval aliasing causes distinct queries to produce identical evidence, wasting compute.
• Standard RL optimizations miss structural flaws at the policy-environment boundary.
• Harness-G uses graph structures to enforce unique, non-redundant retrieval paths.
• Stabilizing retrieval decisions improves multi-turn interaction efficiency and convergence.