RippleCP: A Metric for Checkpoint Value in Coding Agents
A new framework quantifies which recovery checkpoints in coding agents are worth materializing, addressing a gap in checkpoint system design.
Research Digest
Preprints and papers, rewritten in plain English.
Frozen medical AI models miss lessons from real-world cases, paper argues
A new preprint proposes a model-agnostic way for multimodal medical AI to keep learning after deployment.
Per-Call-Site Evaluation of Small Models in Agentic Home Automation
A new study evaluates small language models at each call site in a deployed home-automation agent, suggesting that frontier models may be reserved for the most complex steps.
Constraint-Margin DPO aims to fix LLM planning preferences
A new arXiv preprint proposes modifying direct preference optimization so constraint violations are weighted by their severity rather than treated equally.
Can LLMs Offer a Common View of Justice?
A new arXiv paper asks whether large language models can help reconcile human judgments of fairness that are inevitably shaped by social position and culture.
Probability-Only Model Jev Aims to Turn Calibrated Probabilities into Decisions
An arXiv study introduces Jev, a probability-only 'System One' model that outputs calibrated probabilities in milliseconds, and examines how to use it for decision-making.
FreeEvolve Learns to Replace Hand-Designed Agent Search Loops
A new method called FreeEvolve aims to automate the optimization loop itself, not just the agent designs it produces.
Breaking the Space Barrier for Structured LM Output
A new arXiv paper tackles the tokenization problem that arises when automata enforce strict formats on language model output.
AGAR: A Reinforcement Learning Substrate for Evolving Programs with LLMs
A new arXiv paper describes AGAR, a reinforcement learning substrate that lets language models rewrite candidate programs under a search loop for algorithm discovery.
Trajectory Abstractions Proposed as Behavioral Variables for Language Agents
A new framework aims to give researchers reusable behavioral variables for studying language agents across tasks and models.
AI Evaluation Needs Context from Its Ecosystem, Paper Argues
A new arXiv paper says benchmark validity depends on the interplay of model providers, users, funders, and regulators.
SearchWorld: World-Model Imagination for UAV Object Search
A new arXiv preprint proposes using spatial value-grounded imagination with world models to help autonomous drones search urban environments.
CanonQ: Extreme Low-Bit Quantization for LLMs
A new quantization scheme tackles the challenge of jointly compressing weights, activations, and KV caches to extremely low bit widths.
Winner's Curse in LLM Self-Improvement: Noisy Selection and Lock-in
A new paper warns that self-improving LLMs can overvalue changes due to measurement noise on small evaluation sets, leading to lock-in on suboptimal updates.
Audit Finds Blind Spots in AppWorld and WorkArena Task Verifiers
A new study uses mutation testing to expose weaknesses in execution-based verifiers for agent benchmarks.
LLM-Enabled UAV Dispatch: Survey Maps Shift to System-Level Coordination
A new arXiv survey and taxonomy argues that UAV dispatch is evolving from isolated path planning toward LLM-driven system-level coordination.
DUDA-Bench: Benchmarking LLM Agents for Urban Diagnosis
A new benchmark, DUDA-Bench, tests how well LLM agents can diagnose urban problems from multimodal data to support evidence-based planning.
MIMESIS: Learning User Simulators to Train Interactive Agents
A new method learns realistic user simulators from data, offering a scalable alternative to expensive human feedback for training interactive language agents.