Which Long Documents Matter for Context Extension? A Controlled Test
A new controlled study compares long-document data-selection pipelines for context extension under matched training budgets.
Research Digest
Preprints and papers, rewritten in plain English.
Over-Defensive Coding in AI Agents Traced to Silent Forks
New research shows that unnecessary defensive actions in coding agents evade both outcome rewards and credit assignment, starting at silent decision points.
Contextual gates improve compositional activation steering in language models
A new arXiv paper proposes contextual gates to make it easier to steer multiple language-model behaviors at once.
LastWords Revisited: NLP's Unsolved 'Boring' Problems
A new arXiv paper revisits a 2022 Computational Linguistics column that argued some NLP problems remain unsolved despite AI advances.
Item Response Theory Offers New Lens for LLM Text Detector Benchmarks
A new arXiv paper argues that aggregate metrics like accuracy and F1 miss important nuances in evaluating LLM-generated text detectors, proposing Item Response Theory as a measurement framework.
Shapley Attribution Reveals How Rubric Items Drive LLM Judge Scores
A new arXiv preprint proposes using Shapley attribution to unpack which checklist items most influence LLM-based evaluation and reward.
Embedding-Based Taxonomy and Benchmark for Measuring Text Diversity
A new framework aims to standardize how researchers measure diversity in text corpora, with implications for training data and model evaluation.
Single Attention Head Can Trigger Language Confusion in LLMs
Ablating one attention head can make a large language model switch to an unexpected language, suggesting language control is concentrated in a few units.
When Does Better Acceptance Pay in Speculative Decoding?
A new arXiv study proposes a comparison protocol for deciding whether the extra latency of speculative decoding is worth its throughput gains.
LLMs' Ability to Verbalize Math Formulas Put to the Test
A new arXiv study systematically evaluates how well large language models can translate mathematical formulas into accurate written words.
Benchmark Tests Whether LLMs Keep to a Requested Perspective
The new evaluation focuses on perspective alignment, a failure mode where models answer correctly but mix in viewpoints the user did not ask for.
DeGMatch: A Gradient-Matching Method for Efficient Reasoning Data Selection
A new arXiv paper proposes DeGMatch, a deconfounded gradient-matching approach to select target-aware reasoning data and cut the cost of long-trajectory post-training.
MemBubble Benchmark Measures How Agent Memory Cuts Response Diversity
A new benchmark quantifies how much LLM agents' use of memory for personalization reduces the variety of their responses.
AI agents need memories that expire when their reasons do
A new arXiv paper argues that LLM agents should tether stored memories to the verifications they depend on, so revoked keys or changed premises invalidate old conclusions.
Hybrid MT Still Needs Symmetric Language Resources, Paper Argues
A new arXiv paper highlights a persistent bottleneck in machine translation: heavy reliance on balanced data from both source and target languages.
Deep Fusion Memory: A Parametric Approach to Adding Knowledge to LLMs
An arXiv paper proposes chunk-level parametric memory to help large language models use external knowledge without the drawbacks of continued pretraining or retrieval.
Dual-Slot: Parallelizing Looped Language Models
A proposed architecture reduces the sequential bottleneck of looped language models by processing iterations in parallel slots.
P2N: Reusing Deep Representations for Greater Reasoning Depth
A new arXiv paper proposes reusing deep representations to increase effective computational depth in language models for reasoning tasks.