Agentic AI

The GATS framework addresses the high cost and stochastic behavior of LLM-based planning tools like LATS and ReAct by eliminating LLM calls during inference. It combines systematic UCB1-based tree search with a three-layer world model that handles exact symbolic matching, execution log statistics, and LLM predictions only for unknown actions. This approach aims to deliver superior planning performance on synthetic tasks with branching paths and dead-ends while reducing computational overhead.

• GATS removes LLM inference during the planning phase, significantly cutting computational costs.
• A layered world model handles known actions via symbolic matching and logs, reserving LLMs for unknowns.
• UCB1-based tree search provides systematic exploration, reducing the stochastic behavior of prior methods.
• Performance gains are demonstrated on synthetic tasks involving complex branching and dead-ends.
• This architecture offers a deterministic alternative to purely LLM-driven agent planning workflows.

KV-PRM addresses the quadratic scoring cost of text-based Process Reward Models in long multi-agent rollouts by eliminating redundant text re-encoding. Instead of processing full trajectories from scratch, the model directly reads the KV cache generated during the LLM's initial generation phase. This approach allows for efficient scoring of agent steps using a single verify token, significantly reducing computational bottlenecks in test-time scaling.

• Eliminates quadratic scoring cost by reusing existing KV caches instead of re-encoding text.
• Enables efficient Process Reward Modeling for long-context multi-agent rollouts.
• Reduces test-time scaling overhead, making PRMs more practical for complex agent tasks.
• Validates KV cache reuse as a viable optimization for inference-time reward signals.

This paper addresses the context loss problem in agentic LLMs that generate code via multi-turn tool use, where sessions typically start from zero. The proposed architecture, shared selective persistent memory, retains reusable context such as task specifications, data schemas, and tool configurations while discarding irrelevant session-specific reasoning traces. This approach avoids the token inefficiency and quality degradation associated with naively persisting entire conversation histories.

• Agentic LLMs lose productive context like schemas and constraints between sessions, hurting efficiency.
• Selective persistence keeps reusable artifacts (specs, schemas) while dropping verbose reasoning traces.
• Shared memory workspaces allow multiple sessions to leverage retained context without token bloat.
• Reduces irrelevant context noise, improving generation quality compared to full history retention.

This paper introduces GRACE, a method to manage persistent system instructions for LLM agents by storing them as a typed semantic graph instead of flat text. It validates updates within local neighborhoods to prevent conflicts as the instruction set grows over long operational horizons. The approach reconstructs accepted graph changes as incremental edits to the final system prompt, ensuring reliable context evolution under distribution shifts.

• Replaces fragile flat-text instruction maintenance with a structured typed semantic graph.
• Validates updates locally to reduce verification complexity as context scales.
• Enables reliable long-horizon agent operation by handling accumulated instruction interactions.
• Reconstructs graph updates as incremental edits to maintain a consistent system prompt.

OpenProver is an open-source system for automated theorem proving that uses a Planner-Worker-Verifier architecture to generate and check proofs in Lean 4. A Planner agent manages a scratchpad and repository, delegating tasks to parallel Workers while a Verifier ensures correctness. The system supports an interactive terminal interface, allowing human operators to monitor and guide the proof search process directly.

• Uses a Planner-Worker-Verifier architecture for structured, parallel proof generation.
• Integrates Lean 4 for automatic formal verification of all generated proofs.
• Fully open-source with reproducible evaluation capabilities.
• Offers interactive terminal mode for human-guided proof search and steering.