Agentic AI

Multi-agent systems often use KV cache transfers to avoid text decoding overhead, but full cache sharing scales poorly with context size and agent count. This research introduces receiver-conditioned communication, where agents only transmit the specific cache data required for the receiving agent's local task. The approach significantly reduces memory and context costs, achieving a 94% reduction compared to traditional full cache sharing methods.

• Avoids linear memory scaling by filtering KV cache data per receiver
• Achieves 94% reduction in cache overhead for multi-agent coordination
• Maintains accuracy by sending only task-relevant evidence
• Reduces latency by skipping full text generation and decoding

A case study demonstrates that off-the-shelf coding agents like Codex and Claude Code can autonomously optimize scientific performance engineering tasks. Starting with a PyTorch-dependent CUDA implementation for fixed-radius nearest-neighbor search, the agents followed an executable goal requiring exact-correctness tests and profiling. The agents successfully removed the PyTorch dependency and conducted hypothesis-driven optimizations to produce a standalone C++ implementation.

• General-purpose agents can execute rigorous scientific performance engineering without code transformation prescriptions.
• Agents autonomously removed PyTorch dependencies from CUDA implementations based on correctness and profiling goals.
• Hypothesis-driven optimization led to a standalone C++ result, proving autonomous scientific refactoring capability.
• Executable goals with exact-correctness tests enable reliable agent-led performance tuning.

EngramRAG addresses associative blindness and static topology stagnation in LLM agents by implementing a dual-state memory system. It pairs a low-latency Waking State with an asynchronous Dreaming State that consolidates memories using Complementary Learning Systems principles. The core innovation is Usage-Modulated Personalized PageRank (U-PPR), which applies Hebbian plasticity to update graph transition probabilities based on entity usage frequency.

• Solves multi-hop relational dependency failures common in conventional agent memory.
• U-PPR dynamically boosts centrality of persistent entities via Hebbian plasticity.
• Asynchronous consolidation prevents temporal decay of core persona invariants.
• Replaces static graphs with usage-adaptive topology to reduce memory stagnation.

Omni-IO Skills introduces a standardized agent harness that enables existing foundation models to operate natively across text, image, audio, video, 3D, and code without costly model retraining. The framework addresses coordination challenges in complex workflows by using hierarchical Skills, a unified multimodal execution interface, and dependency-aware orchestration. It also includes a persistent Asset Registry to manage intermediate assets and cross-turn revisions within multi-asset execution graphs.

• Avoids expensive foundation model updates by decoupling modality support from the core model.
• Standardizes coordination of procedures, dependencies, and intermediate assets across turns.
• Uses hierarchical Skills and a persistent Asset Registry for robust multimodal orchestration.
• Represents complex multi-asset workflows as Declare Execution Graphs for clarity.
• Enables plug-and-play integration, making existing agents omni-native immediately.

Long-horizon coding agents typically wait for expensive, complete tool-call sequences to receive verifiable rewards, which amplifies early errors and increases inference costs. The authors introduce Contextual Early Reward (CER), a method that predicts terminal rewards by analyzing behavioral evidence in the early trajectory prefix. CER synthesizes adaptive rubrics from historical task experiences to evaluate the current stage, enabling earlier decision-making. On SWE-bench Verified, CER improved reward model accuracy by up to 4.2 percentage points while using only 15.3% of the tokens required by the best baseline.

• Predicts terminal rewards early in the trajectory, avoiding costly full-sequence inference.
• Uses adaptive rubrics derived from historical task experiences to assess current progress.
• Achieves 4.2pp RM@8 improvement on Nemotron 3 Ultra over strongest baselines.
• Reduces token usage to 15.3% of baseline while matching or exceeding performance.
• Mitigates sparse terminal rewards and unstable training common in long-horizon tasks.