The Archive

mdok-style system finetuned Qwen3-32B using data augmentation for SemEval-2026 conspiracy detection task, ranking 8th of 52 submissions.

mdok-style system applied QLoRA finetuning on mid-size LLMs for SemEval-2026 multilingual polarization detection across detection, type, and manifestation subtasks.

LLMSearchIndex: open-source Python library for local, offline web search with 200M indexed pages, enabling RAG without paid APIs.

It's interesting how there is no advertised push to replace CEOs. LLMs are incredibly powerful but they have also been marketed very powerfully too.

Relevance propagation method at inference reduces hallucinations in multimodal LLMs by rebalancing modality utilization.

Geometric unlearning method removes specific content from LLMs without full training corpus access, balancing privacy and model utility.

LLMs exhibit genre-dependent credibility assessment bias, misclassifying entertainment news as fake more often than hard news.

Probe-Geometry Alignment surgically removes memorization traces from unlearned LLMs via cross-sequence detection without capability loss.

LLMs solve ARC-AGI-3 puzzles efficiently when allowed to search game logs; hill-climbing with tool use narrows performance gap to humans.

Diagnostic benchmark reveals LLMs fail at faithful procedural execution despite strong reasoning scores, with failures increasing with algorithm length and variable dependencies.

Framework to assess when LLMs should call external tools via decision theory, focusing on web search tool use and knowledge integration.

ML-Bench&Guard: policy-grounded multilingual safety benchmark (14 languages) aligning LLMs with region-specific regulations and cultural context.

MathArena: continuously-maintained evaluation platform aggregating mathematics benchmarks to track LLM progress; successor to static math benchmarks.

SC-Taxo generates hierarchical scientific taxonomies using LLMs with semantic consistency constraints across hierarchy levels.

Task vector analysis reveals structural conflicts (magnitude, sign, module-wise) preventing SFT-RLVR integration in LLMs.

April 2026 open-weights model landscape shows significant activity; MiniMax-M2.7 shifted to non-commercial license, limiting adoption.

Study shows LLMs can strategically resist RL training by manipulating exploration, revealing alignment vulnerability.

LLMs refine graph structure for EEG seizure diagnosis by removing redundant edges in noisy signal processing.

TopBench benchmark (779 samples) evaluates LLMs on implicit predictive reasoning over tabular data, addressing latent intent recognition beyond retrieval.

PLOS and DataSeer use LLMs to measure research data reuse in scholarly publications, finding 43% reuse rate via AI-based detection.

The San Francisco–based startup Goodfire just released a new tool, called Silico, that lets researchers and engineers peer inside an AI model and adjust its parameters—the settings that determine a model’s behavior—during training. This could give model makers more fine-grained…

Study finds persona prompting in multimodal LLMs produces stable but limited behavioral variation in urban sentiment judgment tasks.

SpecVQA benchmark evaluates multimodal LLMs on spectral understanding with 620 expert-annotated scientific images across 7 spectrum types.

MLLMs fail on circuit-to-Verilog translation due to 'Mirage' phenomenon; visual perturbations cause hallucinated code despite correct diagram interpretation.

Language model agents design mechanical linkages via symbolic lifting: LLMs explore topologies while numerical optimizers tune parameters, validated on six motion targets.

LLM+ASP framework translates natural language to Answer Set Programming for nonmonotonic reasoning, enabling defeasible logic without SMT limitations.

Reddit discussion recounting a conversation between a local LLM developer and a European government AI policy leader about domestic AI adoption and research priorities.

User demonstrates local LLM deployment via LiteLLM with Prometheus/Grafana monitoring for Frigate video AI summaries.

TIDE framework enables cross-architecture knowledge distillation for diffusion LLMs, reducing parameter requirements while maintaining parallel decoding capabilities.