Terminal-RSI: Recursive Self-Improvement for Terminal Agents
A new arXiv paper proposes an approach that turns verifier outcomes into policy updates and partial progress into completed tasks.
Research Digest
Preprints and papers, rewritten in plain English.
Learning from Unverified Peers via Shared Residual Distillation
A new post-training method uses a model's own unverified solutions as learning signals through shared residual distillation.
Prism-1: Position- and Rival-Independent Scoring for One-Pass Decision Models
A new scoring method aims to make System One decision models less sensitive to option order and competing choices.
New method RADAR repairs and audits LLM confidence scores
A new approach called RADAR targets the limits of scalar recalibration in LLM confidence scoring.
CART: A Routing Approach for Multimodal Continual Instruction Tuning
A new method aims to help multimodal LLMs learn new tasks sequentially without losing prior capabilities.
Federated Skill Evolution for LLM Agents via Version Control
A new framework lets large language model agents evolve reusable skills from distributed execution experience without centralizing data.
Multi-Sourced AI Tutors Aim to Support Non-Native English Speakers
A new arXiv paper proposes a multi-sourced approach to AI tutoring that accounts for the needs of non-native English speakers in computing courses.
Shortcut Cues Can Skew AI Safety Guard Evaluations
A new arXiv paper warns that evaluation data for AI safety guards may contain shortcut cues that make labels predictable without genuinely testing agent safety.
GNN-Guided Refinement for DSA-MP Layout Decomposition
A new approach uses graph neural networks to guide local refinement in layout decomposition for directed self-assembly combined with multiple patterning lithography.
HTN-Orchestrate: Hierarchical Planning for Multi-Agent Tasks
A new framework uses hierarchical task networks to bridge local subtask execution and global planning in multi-agent orchestration.
Random Projection Aims to Curb Privacy Leaks in Retrieval-Augmented Generation
A new arXiv preprint proposes random projection to prevent retrieval-augmented generation from leaking sensitive information.
Transformers Need Custom Design for Continuous Diffusion Language Models
A new arXiv paper argues that continuous diffusion language models require transformer architectures tailored to their specific needs rather than borrowed from causal LLMs.
EliteReplay aims to keep LLM solutions diverse during RL post-training
A new method called EliteReplay applies quality-diversity search to prevent reinforcement learning from narrowing the range of solutions an LLM can produce.
Why Scaffolding Still Helps Coding Agents
A new arXiv study asks why developer-supplied scaffolding still boosts coding agents even as language models get stronger.
LEASH Framework Uses Uncertainty to Escalate Computer-Use Agents
A new method, LEASH, cuts costs by escalating from a weak to a strong model only when the weak model is uncertain.
MLExpert: Toward Autonomous Machine Learning Engineering
A new arXiv paper proposes a generalist agentic model that can autonomously develop and improve AI systems, advancing the path toward recursive self-improvement.
FDE-Gym Tests AI Agents on Ambiguous Customer Requests
A new benchmark simulates the messy, clarification-heavy work of forward-deployed engineers, moving beyond well-specified task tickets.
Four Gemini 4 Argon Agents Defeat Ender Dragon in Minecraft
A new arXiv preprint reports that a team of four Gemini 4 Argon agents, each controlling a Minecraft player via Mineflayer, cooperated to defeat the Ender Dragon.