LLMs as Feature Engineers for Text-and-Tabular Prediction
Iterative framework using generator and extractor LLMs to synthesize interpretable categorical features for tabular prediction.
Search the full wire by company, model, lab, or keyword. Every story we have ever aggregated.
Iterative framework using generator and extractor LLMs to synthesize interpretable categorical features for tabular prediction.
Novel membership inference detector using entropy-corrected loss boundaries to distinguish pretraining data exposure from learned generalization in LLMs.
Fine-tuning Qwen2.5-7B and Ministral-8B on personality-labeled corpora improves consistency vs. instruction prompting for social agents.
L0-MoE accelerates dense LLMs via L0-regularized Mixture-of-Experts, achieving 2.5x speedup with minimal performance loss.
Open-weight LLMs tested on Norwegian Moral Foundations Questionnaire; persona steering and activation-level intervention techniques partially shift moral profiles.
Opinion: LLMs should be used for editing, fact-checking, and grammar—not phrase generation—to maintain authentic human voice.
SynthID can cause models to follow harmful instructions they would otherwise refuse.
Paper proposes 'summarization bias'—LLMs collapse narrative structure into abstract summaries rather than reconstructing inferential meaning—with formal testing protocol.
Study shows explicit sense inventories improve LLM performance on Word-in-Context lexical disambiguation vs. traditional WSD.
I have a lot of mixed feelings about AI and LLM technology. I’m fascinated by its effect on our profession, excited by the potential gains in productivity - and thus the products we could rapidly build. On the other hand, I’m fearful of the damage AI might cause: agent swarms…
MERIT-Rank uses multi-trajectory reasoning with LLMs to improve document reranking robustness beyond single reasoning paths.
MaSCoD multi-agent framework uses LLMs for causal discovery by organizing candidate variables before edge judgment; shows dataset-dependent performance on Auto-MPG, DWD, Sachs.
Infinite-parameter LLMs adapt weights from live interaction data via MoE, enabling models to learn facts and corrections post-deployment without retraining.
Evaluates multi-agent tool-augmented deductive reasoning via Clue board game environment, testing LLM consistency across extended reasoning chains.
Weakening neurons in GLU-based LLMs suppress residual stream directions via negative input-output cosine similarity, concentrated in late layers across models.
Population genetics framework interprets AI model generations as sexual/asexual reproduction across LLMs and neural networks.
Aligned Continuous Integrate-and-Fire framework compresses speech tokens efficiently for zero-shot SpeechLLM alignment without full-model fine-tuning.
MiST suite (8B/32B models) applies mid-training on curated cybersecurity corpus for strong domain performance without raw continual pre-training.
TypeSafe releases Jev, a lightweight classification/routing model claiming 100x+ speed and 200x+ cost reduction vs. frontier LLMs.
Chain-of-Self-Questioning framework reduces LLM hallucinations by conditioning answer commitment on explicit information-requirement assessment across eleven model families.
LLMs form linear belief-state geometry for HMM hidden states during in-context learning; probed across 6 models on 40 HMMs.
Empirical comparison of context windows, fine-tuning, and latent representations for knowledge adaptation in LLMs across five benchmarks.
Zero-shot narrative detection in social messages using LLMs; human-written narrative descriptions improve classification without training examples.
Study of counterfactual self-explanations across LLaMA-3 and Qwen-2.5 shows model scale predicts explanation faithfulness in sentiment analysis and NLI tasks.
Tool-assisted workflow identifies common student mistakes in mathematical modeling (formulas, equations, regex) via LLM analysis for targeted feedback.
EviScope: Counterfactual benchmark (960 tests) evaluating whether LLMs ground answers in correct evidence vs. producing correct outputs from wrong sources.
UNRESTSENT200K dataset enables crisis sentiment analysis for Bangla during July 2024 Bangladesh uprising with 200K annotated social media comments.
Nameless tokenization defends open-weight LLMs against control-token forgery, fixing vulnerability in all 256 audited chat tokenizers.
Lit3R combines retrieval, reranking, and LLMs for evidence-grounded QA over scientific literature without task-specific training.
Gavel extracts skill routing signals from frozen LLMs using linear probes, eliminating context overhead and enabling larger agent skill libraries.