OR-License benchmark tests LLM agents on raw client problems
A new evaluation, OR-License, asks whether LLM agents can solve operations research problems directly from original client materials, without expert reformulation.
Research Digest
Preprints and papers, rewritten in plain English.
SLA-LENS Automates Analysis of Error-Follow-Up in Classroom Talk
A new tool scales up context-sensitive annotation of how instructors respond to learner errors in classroom transcripts.
External Memory Proposed to Ease Reasoning Models' Context Burden
A new arXiv paper highlights how reasoning models depend on in-context memory and suggests external memory as a way to manage growing context demands.
LLMs Are 'Lazy Poets': Study Finds Creative Mode Collapse at Scale
A new arXiv paper argues that aligned large language models produce highly self-similar creative outputs, a problem the authors call creative mode collapse.
Study: Encoder-Free Multimodal LLMs Still Learn Implicit Encoders
A new analysis of Gemma-4 12B and Inkling-Small shows that multimodal models without explicit encoders can still develop implicit encoder-like representations.
NVFP4 Training: Comparative Study of Low-Precision Recipes
A new arXiv study examines which design choices are proven for training LLMs directly in 4-bit floating-point formats.
Pixel-Based Language Models: Promise and Pitfalls for Noisy Text
A new arXiv paper outlines why pixel-based language models may help with noisy and multilingual inputs, but their practical reliability remains an open question.
Interruption Budgets Could Replace Question Counts for LLM Agents
A new arXiv paper argues that measuring user burden by the number of clarifying questions misses the difference between necessary and avoidable interruptions.
FlashMap: A New Speculative Decoding Method for Faster LLM Inference
A new arXiv paper proposes FlashMap, a flow-map-based approach to speculative decoding that targets the bottleneck of autoregressive drafting.
Distilling a Universal Local Feature Extractor for Speech Models
A new arXiv paper proposes distillation to create a universal local speech feature extractor, targeting the computational cost of CNN layers in self-supervised models.
New Bengali Dataset Targets Anti-Male Hate Speech Detection
A Bengali dataset and late-fusion transformer framework aim to close a blind spot in hate-speech moderation by focusing on misandry.
Legal Retrieval: Reranker Training Beats Bigger Candidate Pools
A new arXiv paper argues that once embedding recall saturates, industrial legal retrieval systems should focus on scaling hard-negative reranker training.
Converter Generates TypeScript from .NET Computational Models
A proposed architecture aims to let .NET computational models run locally in web applications by generating TypeScript implementations.
Scaling Concept Vectors from Stories for AI Interpretability
A new arXiv paper proposes scaling concept vectors—directions in a model's activation space tied to human concepts like honesty—to improve AI alignment and interpretability.
SOM-based model simulates bilingual naming with limited training data
A new computational model aims to explain individual-level bilingual naming using laterally-connected self-organizing maps that avoid traditional machine learning's data hunger.
LLMs Fail at Multi-Author Style Detection in New arXiv Study
The study reports that off-the-shelf LLMs and simple fine-tuning produce ineffective features for identifying author changes in documents.
Safety Kernel Aims to Close Cross-Lingual Gaps in LLM Guardrails
A new arXiv paper proposes a mechanistic alignment method to defend against cross-lingual attacks that exploit uneven safety in LLMs.
Cipher Training May Boost Rare-Term Learning in Translation Models
A new arXiv paper asks whether in-context fine-tuning truly teaches rare terms and proposes a cipher-based method to improve it.