Native Context Editing: A Grammar for LLM Context Edits
A new framework proposes that context editing is a native language ability and teaches LLM agents a grammar for it.
Research Digest
Preprints and papers, rewritten in plain English.
HyperTaxa Uses Multiscale Genome Representations for Microbial Taxonomy
A new arXiv paper proposes HyperTaxa, a method that uses multiscale genome representations to improve microbial taxonomic inference when reference coverage is incomplete.
EvalRx: Agentic Framework for Model Diagnosis and Repair
An open-source agentic framework aims to move model evaluation beyond identifying failures to diagnosing causes and verifying fixes.
Residual Task State: A New Lens on Agent Failure Recovery
A new arXiv paper argues that recovering from failures in long-horizon agents requires accounting for the persistent state that links a defect to its symptom.
Robust Evidence for Rule-Based Conclusions Under Source Failures
A new paper formalizes how to certify Horn-rule conclusions when sources may fail in correlated ways.
EvoSkill v2: Three-Agent System Evolves Agent Skills via Instruction Editing
EvoSkill v2 improves AI agents by automatically editing their task instructions, leaving the model and harness untouched.
Program-Guided Reasoning for Frozen Judgment Models
A new framework asks whether reusable programs can chain local judgments into complex multistep reasoning.
EmbodiedWorld-200K targets general-purpose 3D embodied agents
A new dataset and interactive world models aim to move embodied AI beyond single-task, single-environment benchmarks.
Uncertainty Estimation Guides Expert Re-Annotation for Diatom ID
A new arXiv study suggests that uncertainty estimates can help target expert review in automated diatom classification for freshwater biomonitoring.
Probabilistic Forecasting for Global PM2.5 Pollution
A new arXiv preprint outlines a global probabilistic approach to forecasting fine particulate matter, a key air pollutant.
Attention-based cross-talk architecture proposed for motor fault diagnosis
A new arXiv preprint applies an attention-based cross-talk architecture to compound fault diagnosis in motors, a key task for reliable condition monitoring in physical AI systems.
TimesJev: Faster Time Series Reasoning Without Long Chain-of-Thought
A new model applies System 1-style fast thinking to time series tasks, avoiding the token-heavy reasoning chains typical of LLMs.
Why LLMs stumble on long tasks: missing procedural memory
A new arXiv paper argues that large language models fail at long-horizon tasks because they lack explicit, modifiable procedural memory.
Testing Reflective Optimizers on Local Reasoning Models
A new study examines whether GEPA and SkillOpt gains replicate on locally served reasoning models.
Bandit algorithm stays near-optimal under noise and adversarial corruption
A new generalized linear bandit algorithm provably achieves near-optimal regret with only one update per round, even when observation noise varies and rewards are adversarially corrupted.
Data Thinning Gives GP-UCB Near-Optimal Regret Bounds
A new paper shows that data thinning lets UCB-style Gaussian process bandits achieve near-optimal regret bounds, closing a gap long dominated by non-adaptive methods.
IndustrialCAD-1400 Puts Text-to-CAD Models to the Test
A new benchmark evaluates whether text-to-CAD systems can generate detailed industrial assemblies with functional parts and interfaces.
Accessibility Profiles Reveal What Benchmark Scores Hide
A new arXiv paper argues that single-outcome benchmark scores miss the variability of equivalent phrasings, proposing accessibility profiles to measure breadth and reliability.