The Archive

Differential privacy text rewriting via LLMs preserves privacy but alters linguistic style and register beyond lexical level.

Human-in-the-loop benchmarking framework evaluates LLMs on automated competency assessment for secondary mathematics using rubric-based evaluation.

Domain-adaptive pipeline fine-tunes small LMs for crisis translation via data retrieval, filtering, and preference optimization toward A2-level English.

Benchmark evaluating 72 LLMs on 270 harmful instructions for robotic health attendant safety; mean violation rate 54.4% across models.

Hierarchical cascading approach for Speech Sound Disorder classification using fine-tuned speech representation models, outperforms multimodal LLMs on SLPHelmUltraSuitePlus benchmark.

Reddit discussion questioning why LLMs use language-based chain-of-thought reasoning instead of latent vector space operations for faster, more compressed inference.

A recent paper published in *JMIR Mental Health* (Csigó & Cserey, 2026) caught my attention. The researchers administered the 10 standard Rorschach inkblot cards to three multimodal LLMs (GPT-4o, Grok 3, Gemini 2.0) and coded their responses using the Exner Comprehensive…

Carbon-Taxed Transformers proposes green compression pipeline addressing computational cost, memory overhead, and environmental impact of LLMs in software engineering.

Sparse autoencoders reveal three-phase information flow in LLM emotion recognition with both shared and emotion-specific features.

Agora-Opt combines decentralized multi-agent debate with memory-augmented LLMs for automated optimization modeling from natural-language requirements.

FoodBench-QA 2026 competition compares TF-IDF, DeBERTa-v3, and LLMs for nutrient estimation from recipe text under EU regulatory constraints.

Language-agnostic jailbreak detection using semantic codebooks mitigates cross-lingual prompt attacks on multilingual LLMs without retraining.

Dataset and design guidelines for assessing cultural alignment in LLMs, addressing limitations of prior cultural bias evaluation approaches.

Experimental study demonstrating LLMs can be manipulated to prioritize fringe scientific material and generate misleading fluent responses contradicting scientific consensus.

Study examines LLM integration in hybrid work environments to adjust spatial experiences and collaboration dynamics.

First systematic study of uncertainty estimation in audio-aware LLMs; benchmarks five methods addressing hallucination and confidence calibration.

Opinion piece: MBSE models used with LLMs function as prompts rather than knowledge bases; co-design methodology needed.

Empirical study showing LLMs lack reliable architectural reasoning for networked systems design despite architectural complexity.

Dependency-driven multi-stage prompt pipeline for coherent RPG world and narrative generation using LLMs.

User reports local LLMs (Qwen 27B, Gemma 4 31B) underperform Claude for coding tasks due to poor tool-calling and decision-making.

For years, the AI/ LLM critics had the same reasoning: LLMs don't reason and they just predict the next token Recently, it reasoned better than 50 years of mathematicians on an open erdos problems by applying a basic phd level formula Chat gpt conversation…

ProHist-Bench: LLM historical reasoning benchmark grounded in Chinese Imperial Examination system, evaluating evidentiary reasoning over lexical knowledge.

BMW case study: adapts LLMs to generate multi-file domain-specific language code for Xtext-based DSL with Java/TypeScript output.

Measures sycophancy (agreement bias) in LLMs deployed for agentic financial tasks under user contradiction.

Evaluates whether LLMs track Turkish evidential morphology conditioned on source trustworthiness in cloze tasks.

K-MetBench: expert-level multimodal benchmark for Korean weather forecasting LLMs, exposing gaps in visual reasoning and domain knowledge across 55 models.

LLMs derive traffic law requirements for autonomous vehicles, scaling beyond manual formal-logic encoding of legal compliance.

Layerwise Convergence Fingerprinting: tuning-free runtime defense detecting backdoors, jailbreaks, and prompt injections in LLMs.

ElementsClaw combines Large Atomic Models with LLMs via agentic orchestration to automate materials discovery workflows.

AIPsy-Affect releases 480-item clinical stimulus battery without emotion keywords to enable mechanistic interpretability of emotion circuits in LLMs.