The last six months in LLMs in five minutes
Simon Willison summarizes six months of LLM developments (Nov 2025 onward) including a November inflection point and rapid shifts in top model performance, especially for coding.
Search the full wire by company, model, lab, or keyword. Every story we have ever aggregated.
Simon Willison summarizes six months of LLM developments (Nov 2025 onward) including a November inflection point and rapid shifts in top model performance, especially for coding.
Vision-OPD: on-policy self-distillation training for MLLMs to improve fine-grained visual detail recognition via regional crops.
Benchmark and attribution method auditing value pluralism in medical LLMs, revealing ethical stance biases in clinical advice.
Scaling law: LLM factual recall sigmoid-scales with model size and topic frequency, explaining 60-94% variance across model families.
Framework uses LLMs as OR experts to dynamically re-optimize operational models via natural-language interaction, addressing real-world constraint evolution.
GIM benchmark of 820 problems evaluates LLMs on integrated multi-domain reasoning grounded in practical contexts, addressing saturation of existing benchmarks.
SCICONVBENCH benchmarks LLMs on multi-turn clarification dialogues for ill-posed scientific task formulation.
CrossView Suite adds dataset, model, and benchmark for cross-view spatial reasoning in MLLMs with object-level consistency.
Proxy metrics from token-level statistics predict downstream LLM capabilities faster than cross-entropy loss.
Community discussion on sustainability of local LLMs if free model releases cease; proposes knowledge-retrieval tooling as mitigation for model staleness.
DystopiaBench evaluates 42 LLMs across 6 dystopian scenario types with escalating compliance tests; finds closed-source models may mask unsafe behavior.
Three months ago I pressure-tested which LLMs would cave and help build the apocalypse. Claude was the only one that consistently said no. Since then I've tested 30 more models across 6 dystopia modules (Orwell, Huxley, Petrov, Basaglia, LaGuardia, Baudrillard). The gap…
MLLMs fail at grounded 3D spatial reasoning and multi-agent Theory of Mind due to text-based probability distributions lacking topological understanding.
Generative Visual Grounding uses EEG-to-image models to ground brain signals visually instead of via text, improving MLLM-based brain foundation models.
Elastic-dLLM reduces computational redundancy in diffusion language models via context compression and position-aware augmentation.
Vision Inference Former addresses visual consistency in MLLMs by elevating visual features above text-token parity in connector-based architectures.
BanglaMedVQA: first clinically validated medical visual QA benchmark for Bangla language with LVLM evaluation.
Safety Geometry Collapse: multimodal LLMs fail safety transfer due to representation drift; proposes adaptive correction mechanism.
PARAM-Delta method upcycles dense models to MoE architecture for efficient multilingual LLM expansion via post-training without full alignment.
Hardware sizing reference chart for Strix Halo mini PCs running local LLMs.
Reddit commentary arguing that characterizing LLMs as 'next token predictors' oversimplifies their capabilities and architecture.
Analysis of LLM-generated synthetic identities used to sell unvetted medical content online, raising concerns about agentic content at scale.
It might be the fact that as you use AI more, you quickly learn that **being** **direct**, making **specific requests**, and **giving constraints** will get you the best results. At the same time it has me thinking carefully about exactly what my *intentions* and *wants* are…
Empirical study shows LLMs mediating human-to-human communication can steer collective opinion formation.
Fully Open Meditron: first end-to-end auditable clinical LLM pipeline with published training data, curation, and generation procedures.
Formal methods + ML for auditing and runtime monitoring of AI systems against behavioral constraints like safety rules and regulations.
Equibles: open-source MCP server enabling local LLMs to query SEC filings, 13F holdings, insider trades, and FINRA short data without cloud dependencies.
SGR framework enhances LLM reasoning via external subgraph generation to reduce hallucination and improve logical inference in complex tasks.
DebiasRAG mitigates social bias and hallucinations in LLMs through retrieval-augmented generation without fine-tuning.
Hybrid architecture combining LLMs and GNNs for relational database foundation models that preserve relational context.