The Morning
From the arXiv
CliffCompaction: Cost-Efficient Compaction for Long-Horizon Coding Agents
liffCompaction is an autocompaction technique for long-horizon coding agents that significantly reduces costs (up to 50%) by only truncating or dropping information, never rephrasing it. This method maintains or improves performance on benchmarks like Terminal-Bench and KernelBench, enabling more efficient test-time scaling and achieving state-of-the-art results. Its core contribution lies in cost-effective context management without sacrificing fidelity.
PACT: From Credit Assignment to Critic Alignment
This paper addresses the lack of a clear mathematical definition for token-level credit in reinforcement learning for LLMs. It proposes three regularity conditions (Completeness, Prefix Consistency, and Neutrality) that uniquely define token-level credit, prov…
Recursive self-improvement of AI research agents
This paper introduces recursive self-improvement for AI research agents, where an agent iteratively modifies and improves its own code. The core method, AIDE^2, proposes, benchmarks, and retains code changes that enhance performance on AI R&D tasks. Its main c…
REFLEX with Jev for Efficient Selective Control in LLM Agents
This paper introduces REFLEX, an LLM agent architecture that uses a fast, typed decision layer (Jev) for efficient, selective control. REFLEX calls a powerful LLM only when confidence is low or generation is necessary, significantly reducing computation while …
Flash-dLLM: IO-Aware KV Caching and Parallel Decoding for Fast, Memory-Efficient Diffusion LLMs
Flash-dLLM tackles the inference inefficiency of Diffusion LLMs by addressing I/O bottlenecks in KV caching and parallel decoding. Its core method is an I/O-aware fused KV-cache kernel that minimizes redundant memory transfers. This optimization enables effici…
A2M: Trace-Optimized Agent Hijacking in the MCP Ecosystem
This paper introduces A2M, a two-stage framework for hijacking agents that use the Model Context Protocol (MCP). A2M first optimizes tool metadata to attract agent invocation and t…
Beyond Repeated Sampling: Learning Search Policies for LLM Reasoning
This paper proposes a new method for LLM reasoning that moves beyond simple repeated sampling. Instead of generating many similar answers, it learns to generate diverse "concepts" …
From Alignment to Access Control: A Framework for GenAI Policy Enforcement
This paper proposes a framework for systematically analyzing and dissecting existing approaches to policy enforcement in Generative AI (GenAI) applications. Its core contribution i…
Greedy Decoding Is Not Precision-Invariant: Cross-Precision Output Divergence in LLM Inference
This paper demonstrates that greedy decoding in LLMs is not precision-invariant, meaning the same model and prompt can produce different outputs when using BF16 versus FP16 precisi…
Grow the Harness, Not the Context: From Strategy-Free Scaffolds to Reusable Specialist Agents
This paper proposes "Growing Harness," a method to train LLM agents by learning reusable code for recurring control decisions, rather than repeatedly asking the LLM to make them. B…
The Town Square
The Pentagon has stated that an overreliance on artificial intelligence was a contributing factor in a missile strike that hit an Iranian school.
Workshops
Agent Substrate is the core system for building and deploying AI agents, providing a foundational framework with essential functionalities for agent development and management.
This repository provides an "Office Harness for AI Agents," enabling them to interact with and manipulate various document formats like spreadsheets, docs, slides, and PDFs within a single runtime environment.
