Natural Language Processing

A projection-based, black-box method operating directly on output embeddings recovers meaningful chronological and spatial structure from language-model representations, according to arXiv paper 2609.05721v1. The method …

Researchers introduced MonoTM, an interpretable topic modeling framework that decouples document-topic mixture estimation from semantic interpretation, according to a paper posted to arXiv as 2609.09575v1. MonoTM estimat…

Researchers introduced AutoLexSteer, described as the first fully automated process for building steering vectors for large language models, alongside a new steering vector type called Hangman that operates using word se…

Researchers studying Looped GPT-BERT in the BabyLM 2026 Strict-small setting trained a 12.18M-parameter model on a preprocessed 7.48M-word English corpus, using four physical layers for twelve recurrent traversals, and t…

Researchers released BuzzASR, a collection of 102 language-specialized fine-tuned Whisper models for automatic speech recognition, according to an arXiv paper (2609.09554v1). BuzzASR models outperform Whisper-large-v3 on…

Researchers introduced jaROTE, a rule-based pipeline that reproduces omitted temporal expressions in Japanese news articles as concrete dates or intervals using the publication date as external context before indexing fo…

A study posted to arXiv (2609.09363v1) found only a weak context-memory conflict in large language models, with counterfactual RDF triple inputs scoring just 0.05 lower on a 1-5 faithfulness scale than factual inputs whe…

Researchers introduced Edu-QuRating, a pipeline that scores and curates educational text on multiple education-specific criteria by distilling LLM-judge pairwise preferences into reusable Edu-QuRaters, with the best scor…

A new arXiv paper, arXiv:2609.09496v1, traces three "mutations of machine speech" through which algorithmic outputs have reshaped expression: search engines redefining speech as data to be queried, social media platforms…

A new arXiv paper (2609.09462v1) reports that factorizing the dense CoOp prompt matrix P ∈ R^(m×d) as P = BA cuts trainable prompt parameters from md to r(m+d), and to rd once the token-side factor B is fixed, matching o…

A study of an author-facing LLM system that generates pre-submission peer review feedback found that independent sampling covered 44.9% of historical issues on a ten-paper diagnostic, while deduplication and refill reach…

AgentBrew, an offline training framework from researchers publishing on arXiv (2609.05837v1), learns tool-use policies from a single batch of raw interaction trajectories without task verifiers or iterative on-policy rol…

Researchers proposed EvoTree, a staged framework that generates evolution trees from citation graphs by decoupling conceptual backbone learning from temporal refinement, according to the arXiv paper 2609.09561v1. EvoTree…

A new arXiv paper introduces SWORD (Systematic Wikidata-based Object-Relation Distortion), a benchmark that tests whether large language models consistently reject factually incorrect statements across eight widely spoke…

A new arXiv paper (2609.07160v1) formalizes In-place Instruction Following (IIF) for diffusion large language models and introduces GRAFT, a post-training framework that raised the average IIF score from 57.75 to 73.10 (…

Researchers proposed Diverse Skill Routing (DSR), a diversity-aware reranking framework that uses a Determinantal Point Process to balance relevance and non-redundancy when routing user requests over large skill registri…

A new multi-stage benchmark called SciLitBench, spanning 42,981 retrieved records, 1,012 full texts, and annotations for 888 included papers, finds that explicit inclusion and exclusion criteria improve title and abstrac…

Researchers proposed X-CoSD, a cross-vocabulary collaborative speculative decoding framework for distributed LLM inference that works with heterogeneous small language model and server LLM vocabularies, according to an a…

A new arXiv paper, arXiv:2609.07153v1, introduces FreqBLiMP, a frequency-controlled extension of the BLiMP benchmark that regenerates all 67 paradigms under explicit Zipf-frequency regimes while preserving each minimal p…

A developer detailed how Whoosh's analyzer pipeline determines whether full-text search returns results, showing that mismatched index-time and query-time analyzers cause zero-result queries. The writeup demonstrates tha…