Internalizer: Scaling Context-to-Parameter Mapping to Very Large LLMs
A new hypernetwork approach aims to let very large language models carry context directly in their weights via LoRA adapters.
Research Digest
Preprints and papers, rewritten in plain English.
MemTrial: Learning When to Trust Memory in LLM Portfolio Agents
A new arXiv paper proposes MemTrial, a framework that helps LLM portfolio agents decide when to trust their memory of past experiences.
Dynamic Interviewer Architecture Adapts Questions to User Expertise
A new system for automated interviews uses local large language models to adapt questioning dynamically while keeping evidence traceable.
Estimating expectations under autoregressive language models
A new arXiv paper tackles the computational cost of computing expected values of test functionals under autoregressive language models.
Learning Search Control to Cut Planning Memory Use
A new arXiv paper proposes learning domain-specific indexical policies to avoid storing exponentially many states during heuristic search.
Narrative Wrapping Bypasses LLM Safety Refusals, New Benchmark Shows
A new arXiv study measures how role-play and narrative contexts make safety-aligned language models answer harmful requests they would otherwise refuse.
Generative UIs Could Ease the Burden of Text-Based Agent Interaction
A new arXiv paper proposes data-aware, ephemeral generative UIs as an alternative to text-heavy human-agent interaction.
Social Pain and Brain Dynamics in Adolescent Self-Injury
A new study examines how social distress disrupts emotion-action brain states in adolescents with non-suicidal self-injury.
Why Output Alignment Matters for Long-Term Forecasting
A new arXiv paper argues that aligning model outputs with the target sequence is a crucial but overlooked factor in long-term time series forecasting.
Governing Persistent Memory in LLM Agents: Admission and Presentation
A new arXiv preprint argues that persistent memory in LLM agents needs explicit governance over what is recalled and how it is presented.
Harness Compilation Offline Procedure for Small Vision-Language Models
A new offline procedure, Harness Compilation, adapts how a frozen small vision-language model shares work with external evidence sources.
Lower Reconstruction Loss Can Hurt Low-Bit LLM Quantization
A new arXiv preprint shows that minimizing reconstruction loss in weight-only post-training quantization does not necessarily improve low-bit LLM quality.
New Benchmark Tests AI on Unresolved Scientific Problems
OpenProblemBench evaluates AI systems on 82 unresolved problems in foundational theoretical sciences, aiming to measure progress beyond established knowledge.
New Method Recursively Decomposes Long-Horizon Tasks for AI Agents
A new arXiv preprint proposes TaReD, a tool-aware recursive decomposition method to help AI agents handle complex, long-horizon tasks without relying on a single reasoning chain.
Distillation as a Tool to Expose Misaligned AI
A new arXiv paper suggests distillation could incriminate strategically deceptive models, though the abstract is truncated.
UniData Pipeline Aims to Cut Cost of Multimodal Instruction Data
A new arXiv paper proposes a universal pipeline for generating multimodal instruction datasets, targeting the labor bottleneck in MLLM training.
New benchmark evaluates LLM assistance for peer review
A framework and benchmark from arXiv aims to move large language model assistance for peer review beyond simple imitation of human reviewers.
Possibilistic Reasoning for Open-Ended Scientific Discovery
A new arXiv paper proposes using possibilistic reasoning to let LLM-driven discovery adapt hypotheses while preserving statistical validity.