The Archive

LLMs can audit and repair clinical neural networks by reading/writing raw weights, tested on radiotherapy screening models.

MedPhysBench evaluates LLMs on medical physics calculations across radiotherapy, imaging, and nuclear medicine with protocol-based grading.

Fine-tuning LLMs on data they believe false induces broad misalignment across domains; demonstrates realistic fine-tuning safety risk absent explicit harm.

Analysis of language-subspace interventions in open-weight multilingual LLMs using MATH and MGSM; specialized technique for improving multilingual reasoning.

Method using activation-space geometry (anon-vectors) to detect suppressed knowledge in instruction-tuned LLMs; advances interpretability and evaluation of model internals.

Evaluates representation validity of text, audio, video inputs across six open-weight MLLMs on social science video analysis tasks.

Yesterday's hot paper was Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights by Rikiya Takehi, Jayden Teoh, Alex Zhang, and Omar Khattab. The paper found that frontier models, in particular GPT-6 Astra, can both interpret and modify the raw weights of a machine learning model - literally a text description of the model containing a bunch of floating point numbers. I've been playing with one of the examples from the paper and I'm delighted to share that Qwen 3.8 27B running on my laptop can do this as well! Weight-Space Programming Here's the highlight fr...

MobileSpec trains specialist LLMs on-device via forward-only optimization and RLVR, enabling adaptation without backpropagation memory overhead.

Study exposes flaws in SAE-based PII suppression evaluation; shows logprob reduction alone fails to distinguish feature suppression from magnitude artifacts.

Study of input-dependent activation steering in LLMs shows Contrastive Activation Addition effectiveness varies by class-dependent vs. class-agnostic application.

Study challenges BLEURT-based automatic hallucination labels in LLMs, identifying bias and methodological issues in ground-truth construction for detector evaluation.

Evaluation of 12 open-weight LLMs as judges for counterspeech quality against online hate speech, assessing scalability and reliability of automated assessment.

Weight Space Programming enables LLMs to directly edit neural network weights; demonstrated on Qwen and DeepSeek for reasoning optimization.

FlowTranslator uses text-to-image models and LLMs for training-free in-image machine translation via flow-based visual editing and OCR.

DiMuRe framework for robotic agents uses diagnosis-guided recovery with MLLMs to select task reuse or predefined skills when disturbances occur during execution.

Open-source LLMs fail on long retrieval tasks due to attention competition, not length; neutral filler preserves accuracy to 32K tokens.

GPT-6 Astra directly read and edited Qwen3.5-0.8B weights to improve math capability, demonstrating LLM weight manipulation.

Comments

Chain-of-Thought (CoT) generation has substantially improved the performance of Large Language Models (LLMs) on complex reasoning tasks, including code generation, data analysis, and exam-style question answering. Despite these advances, purely neural LLMs continue to struggle with elementary logical reasoning problems and lack the determinism, soundness, and reliability characteristic of symbolic reasoning systems. Conversely, classical symbolic methods such as SAT solving and exact cover guarantee correctness and completeness, but require problems to be expressed in highly specialized forma...

Large language models (LLMs) show growing promise for scientific discovery, yet basic questions remain about what enables reliable discovery. How can we trace the scientific discovery capacity of LLMs to their training data? We study this question as one of missing probability mass and develop a theoretical account in which monofacts-the facts observed once during training-set the discovery capacity. In materials science and regulatory genomics, we directly manipulate monofact rates in training sets and find that more monofacts consistently yield more discoveries across two open-weight model ...

Analysis of vision model latency trade-offs and Genie 4 generative environment capabilities in relation to LLM inference.

Experimental study comparing hybrid retrieval, NLI models, and open-source LLMs for scientific claim verification using conservative fusion.

VERS steers reasoning LLMs by editing natural language explanations of activations at doubt points, improving GSM8K accuracy.

Kyle Kingsbury warns LLMs undermine trust in software docs and quality signals, prompting debate on verification and credibility for maintainers.

Hybrid narrative framework balances static content with real-time LLM generation for interactive storytelling coherence.

VagueBench benchmark and training method show LLMs ask clarifying questions <11% of time; identifies proactive elicitation gap in 8B-frontier models.

Malware operators deploy LLMs to prioritize targets by financial value, escalating threat sophistication and attack targeting.

TopoQuant uses topological data analysis to measure structural drift in quantized LLMs, revealing compression-performance tradeoffs.

Log-Driven Semantic Tuning uses LLMs to automate MIP solver parameter tuning by interpreting performance logs.