Trail-Weighted Evidence Pruning for Faster Bayesian Prediction
A new scoring method ranks evidence trails by edge strength to prune less relevant information in Bayesian networks, potentially making outcome prediction more efficient.
Research Digest
Preprints and papers, rewritten in plain English.
Why Inference-Time Scaffolds Boost AI: A New Theoretical Lens
A new arXiv paper uses computational separations to explain why external harnesses that repeatedly call a model can improve its capabilities.
Seeing the task: robots that watch before teaching autistic students
A new method lets a camera-equipped humanoid robot recognise a task visually and plan its own steps, potentially making robot-assisted autism education more flexible.
Study: Multi-Task Prompts Reveal 'Implicit Budgeting' Fragility in LLMs
New research argues that language models fail in distinctive ways when a single prompt bundles several tasks, a problem tied to how models implicitly budget their outputs.
Multimodal LLMs Show Structured Hallucination Risk in Cheque Extraction
A new preprint finds that multimodal large language models can produce structured errors when extracting courtesy amounts from cheques under visual ambiguity, and proposes a compressed-image countermeasure.
ICLHerd: Herding-Based Data Selection with In-Context Learning Proxy
A new arXiv paper proposes using in-context learning as a proxy for herd-based training-data selection in LLMs.
MMedAgent-Plus: A Multimodal Agent for Coordinating Medical Workflows
A new tool-use agent aims to coordinate the interdependent analyses that make up real-world medical workflows.
Hallucination Tests Hide Model-Specific Failures
Aggregate scores on hallucination stress tests can look identical across models even when each model trips on different prompts, prompting a new curation approach.
Edge Model Exposure Poses Cross-Device Security Risk
Extracting a neural network from one edge device could compromise other devices running similar models, a new paper warns.
Auditing AI preference labels via recorded derivations
A new arXiv paper proposes recording intermediate decisions to make preference labels in AI training more inspectable.
Does Knowing an Attack Was Caught Change AI Sabotage?
A new arXiv paper examines whether AI agents that learn an earlier attack was caught become better at evading control on later tasks.
Certifying Replay Witnesses for COBOL-to-Java Migrations
A new approach aims to make synthetic test data for migrated COBOL applications repeatable and verifiable.
Context-Driven Class Incremental Learning for Multimodal Models
A new arXiv paper highlights a gap in extending class-incremental learning to multimodal foundational models in context-driven settings.
VPFlow: Value-Predictive Planning for Offline Goal Reaching
A new preprint proposes VPFlow, a value-predictive planning approach for offline goal-conditioned RL with imperfect world models.
3D-MoE Co-Design Tackles Memory Bottleneck in MoE Decode
A new hardware-software co-design uses 3D-stacked near-memory processing to address the memory-bound nature of Mixture-of-Experts inference.
Context Compaction Evaluated on Later Questions
New evaluation asks whether context compaction that helps the current question also serves future ones.
Inverse Scaling Found in Vision-Language Models
A new arXiv study challenges the assumption that bigger vision-language models always improve, showing systematic cases where performance declines.
FPGA Architecture Combines CKKS Encoding and Encryption on the Client
A proposed FPGA design reduces hardware area by sharing resources between the floating-point encoding and modular encryption steps of CKKS homomorphic encryption.