Search the full wire by company, model, lab, or keyword. Every story we have ever aggregated.
Controlled evaluation of GPT-5.4 and gpt-oss-120b on knowledge-driven grammar engineering for Cantonese ParGram resources with varied prompting conditions.
Analyzes how LLMs fail at tracking logical exclusions and proposes Logic-Augmented Generation with POLANYI++ for structured output.
Characterizes low-resource language underperformance in LLMs via representational geometry, finding degeneration in final layers.
This article is from Making AI Work, MIT Technology Review’s limited-run newsletter examining how to apply LLMs across industries. To receive it in your inbox, sign up here. Chatbots took many schools by surprise upon their release a few years ago. Suddenly, students carried an…
One of the memes on Twitter is that “programming is solved now.” I’m not sure to what degree it is, but one thing is pretty clear: the act of familiarizing yourself with a language no longer matters and some of the friction that mattered for humans does not matter for agents…
Analysis of frontier LLMs conducting psychotherapy finds models over-use certain moves (reassurance, validation) vs. human clinicians; maps therapeutic behavior with ontology validated by licensed psychologists.
Empirical study across 7 LLMs shows affective context significantly amplifies sycophancy in subjective evaluations, relevant to alignment.
RARE framework enables representation engineering steering for MoE LLMs by decoupling routing from behavior-directed latent modifications.
LLM augmentation with semi-structured data for predictive political QA; actor stance and vote prediction from external resources.
P3Bench: benchmark for personalized privacy control in LLMs via attention head intervention; user-specific disclosure preferences.
NSPIN neurosymbolic framework combines LLMs and symbolic methods to induce probabilistic planning models from surgical narratives.
Thermo-FL adds thermal-aware control and Byzantine robustness to federated LoRA fine-tuning of LLMs on edge devices.
COEC proposes alternating orthogonal compensation for training-free structured pruning of LLMs to reduce inference cost.
RuleMaze benchmark evaluates MLLMs on rule-compliant spatial planning in mazes with natural-language constraints.
Recommender systems (RecSys) are one of the most ubiquitous machine learning problems in the consumer internet industry yet notoriously difficult to train and... Recommender systems (RecSys) are one of the most ubiquitous machine learning problems in the consumer internet…
FormalTCS benchmark evaluates LLMs on frontier theoretical computer science research with 175 expert-validated instances from STOC/FOCS/SODA/COLT 2025–2026.
Hybrid multi-agent framework orchestrates LLM reasoning with RL/rule-based control for autonomous driving to reduce latency and hallucination risk.
Large language models (LLMs) are increasingly used in settings where textual summaries, numerical observations, and external tool outputs may provide conflicting evidence. We study how LLMs arbitrate between such sources when they support opposing decisions. To do so, we…
Jeremy Morrell proposes LLMs + sandboxing enable extensible web apps where users safely customize core functionality via AI-generated extensions.
SPADE: self-play RL framework where LLMs generate adaptive training environments and learn from them, enabling continuous goal-distribution scaling for language agents.
Intel AI PCs with integrated GPUs/NPUs serve 70B LLMs via layer-wise pipeline parallelism and pre-compiled OpenVINO shards over local networks.
Frontier LLMs saturate on accuracy; precision (output consistency) rather than capability now determines practical advantage, reframing the metrics.
LLMs extract scientific data via self-authored prompts and cross-model consensus, matching expert curation while reducing human annotation burden.
Automated Sigma rule generation from cyber threat intelligence using LLMs reduces manual detection engineering for APT mitigation.
Multi-agent AI pipeline for radiology report structuring and QA using local LLMs deployed retrospectively on 638 CT reports from board-certified radiologists.
Chain-of-Experience framework enabling LLMs to improve through test-time iterative feedback and self-interaction.
TSN4PI framework tracks political ideology evolution on social media using LLMs with style transfer and unsupervised domain adaptation to address annotation scarcity.
BEAR-Bench evaluates multimodal LLMs on bilingual text-dense document reasoning (English/Russian) across enterprise and academic contexts.
Probing study reveals six frozen LLMs encode geometric constraints in hidden states but fail to steer behavior using them.
Novel white-box attack on LLMs using knowledge editing framework and associative context retrieval to induce unsafe outputs.