Natural Language Processing

A developer's guide to RAG chunking argues that the document-splitting step is the most commonly rushed and most damaging part of a retrieval pipeline, since errors there propagate downstream into embeddings and retrieva…

A user of Japanese-English machine translation rated ChatGPT at roughly 80 out of 100 and DeepL at about 60 for one-shot paragraph translation, arguing the technology is very good but not "solved." The user, posting on L…

A developer who built InsNet, a ~21,000-line C++14 deep learning library with 4,828 lines of hand-written CUDA kernels, starting in 2018, describes how its padding-free dynamic batching design anticipated the flat, paddi…

Researchers submitted LensVLM, an inference framework and post-training recipe built on Qwen3.5-9B-Base, to arXiv on 7 May 2026, enabling vision language models to scan compressed rendered text and selectively expand onl…

A developer building a Japanese decision model measured the multilingual checkpoint of the 20k-star open-source model laya and found its ordinal head effectively ignores content, always selecting the first-listed option:…

MINER, a training-free inference framework from researchers publishing on arXiv (2609.27142v1), improves text-to-image retrieval with frozen dual encoders by augmenting a single global image embedding with a small bank o…

A new arXiv paper (2609.27037v1) presents a wake-up system for voice assistants that adds contextual trigger detection to traditional direct keyword detection, using reasoning after an initial wake word to separate user …

Applying Minimum Risk Training to power outage report generation raised Qwen2.5-7B-Instruct's overall accuracy from 16.20% to 68.95%, according to an arXiv paper (2609.27197v1) on the Outage Data Initiative Nationwide (O…

Researchers proposed COPE (Continual Optimization with Personalized embedding and self-Evaluation), a framework that assigns learnable personalized embeddings to each user and uses self-evaluation to generate proxy rewar…

A study on arXiv (2609.26926v1) found that having experts label cases of cross-LLM disagreement with rationales produced the highest LLM-labeling accuracy at 64.9% against expert labels, beating an expert-revised codeboo…

A new arXiv paper (2609.26945v1) introduces a sentence-level benchmark for evaluating how well large language models classify interpretive canons, drawing on Larenz's conception of interpretation in the Savigny tradition…

A dual-module framework called LEGO, combining Legal Expert GraphRAG and expert Chain-of-Thought, reached 40.53% exact-match accuracy on the LawExamQA_Civil benchmark using a Qwen3-8B backbone, according to the arXiv pap…

LexLattice, an extractive summarizer that models a legal act's hierarchy as a two-dimensional semantic lattice and consolidates over it with a masked 2D neural cellular automata, attained state-of-the-art ROUGE across al…

Researchers proposed Tempered Evidence Fusion (TEF), a training-free decision-fusion rule that weights each sentence's log-odds by its normalized information gain, for measuring value orientations in long social media po…

Researchers introduced EduBehaviors, an interpretable annotation framework that uses large language models to measure repeated observable behaviors in educational dialogues and then learns a classifier for a target const…

A new arXiv paper (2609.27173v1) introduces R3Con, a harness that builds context representations for large-scale reasoning using principles drawn from the cognitive theory of relevance realization, and reports it beats t…

Researchers introduced UGTPhon, described as the first grapheme-to-phoneme (G2P) benchmark for user-generated text in English, Vietnamese, and Korean, alongside an inference-grounded taxonomy for fine-grained diagnosis. …

Researchers introduced LOCKR, a hidden-state trajectory-guided planner that detects and repairs "stable-but-wrong lock-in" in diffusion language models, according to the arXiv paper 2609.27220v1. Across two diffusion lan…

A new arXiv paper (2609.27233v1) introduces a method for distilling sequential computation in Transformer language models by replacing spans of input tokens with collapsed surrogate embeddings generated on the fly by a l…

Researchers proposed LLMAE, a method that repurposes a pretrained decoder-only language model as a continuous text autoencoder by exposing a fixed-length latent bottleneck within its internal activations. Instantiated wi…