Formal Raising: Embedding Verifiable Ethics in AI Development
A new arXiv preprint argues that AI should be raised like Turing's 'child machine' with formal, verifiable ethical properties built into its development.
Research Digest
Preprints and papers, rewritten in plain English.
IoT Agents Need a Way to Choose Which Device to Call
A new arXiv paper addresses the problem of selecting the correct IoT device when multiple devices offer the same capability to an LLM agent.
ToolOPD: Credit-Aware Distillation for Tool-Use Agents
A new arXiv paper proposes a credit-aware variant of on-policy distillation to better train small language agents that use tools.
CVaR Selection Pressure for Robust LLM-Discovered Heuristics
A new approach applies CVaR selection pressure to make LLM-discovered heuristics robust against worst-case performance.
Joint Weight Perturbations Aim to Block Harmful Fine-Tuning
A new arXiv paper proposes a collective weight-perturbation defense to keep language models aligned after fine-tuning.
Action-Conditioned World Models for Time Series
A new framework, ActTime, treats future actions as controllable inputs in time-series models, enabling comparison and optimization before execution.
Federated Quantum-Inspired Tuning for Privacy-Preserving Medical LLMs
A new framework combines federated learning with quantum-inspired optimization to let medical institutions fine-tune LLMs without sharing raw data.
Bayesian model maps heavy-metal risk to children in Ghana's Densu Basin
A probabilistic analysis of 94 groundwater samples finds lead, arsenic, and cadmium drive most risk, with about a third of sites likely exceeding safe hazard levels.
Human Expert Skills Could Make LLM Fact-Checking Agents Cheaper
A new approach proposes transferring human expert strategies to LLM agents to reduce the computational cost of automated fact-checking.
Spherical Fixed-Point Reasoners Move Adaptive Loop Halting to a Hypersphere
A new paper proposes constraining Fixed-Point Reasoners to a hypersphere, changing the geometry of adaptive loop halting in looped transformers.
Auditing Tool Traces in Agentic LLMs for Scientific Predictions
A new arXiv paper proposes methods to localize which external sources an agentic LLM used and test whether its predictions truly depend on them.
Benefit-Targeted Routing for LLM Evidence Screening in Supplier Discovery
A study of 6,164 manufacturing source-task pairs shows how routing and bounded revision can control when LLM judgments are used in evidence screening.
ARISE: adaptive residual-informed stability ensemble for feature selection
A new feature-selection method ranked first in every tested dataset–metric combination for small-sample molecular classification.
Benchmarking When Proactive AI Should Speak Up in Group Chats
A new arXiv preprint proposes a benchmark for evaluating the timing of proactive agents' contributions in multi-party chat.
When AI Agents Trust Their Own Records Over Reality
A new paper describes 'claim laundering,' a failure mode in which agent decision models rely on stale internal records instead of current environmental signals.
Tackling Repetitive Actions in Multi-Turn RL for LLM Agents
A new arXiv paper proposes productive state-space exploration to address inefficient action selection and repetitive behavior in multi-turn reinforcement learning for LLM agents.
OrderGPT: Learning the Best Order for Generating Discrete Data
A new model treats generation order as a learnable choice rather than a fixed left-to-right or any-order assumption.
Auditing How Seller Content Steers AI Shopping Recommendations
A new study examines how seller-controlled content can influence what large language models recommend when shopping on a user's behalf.