EMBERS stress-tests Deep Research systems via defect-guided fuzzing, probing how question and search-input variations propagate into wrong answers across stages. Useful for teams shipping or evaluating deep-research agents that need robustness guarantees.

Import AI 478 covers how math could contribute to alignment, the ProgramBench release, and a concrete example of recursive self-improvement from Mirendil. Strong pick for researchers following alignment and RSI discourse.

Argues reasoning text has simpler co-occurrence topology than non-reasoning text, with proposed applications to AI safety and evaluation. A provocative, compact hypothesis likely to spark debate among interpretability and eval researchers.

LHM is a training-free method for merging long chain-of-thought reasoning models with concise base models, arguing that linear interpolation and first-order activation stats miss how the loss landscape responds to parameter shifts. Relevant if you do model merging and want better accuracy on hard reasoning tasks without retraining.

Nathan Lambert revisits his own position on open models and cyber risk after a wave of critical vulnerability disclosures he dubbed 'Black October.' A self-critical take worth reading for anyone weighing open-weight release policy.

Probes LMs on chemistry to show functional-group knowledge is encoded internally before it surfaces in outputs — a lens into what models 'know' vs. say. Relevant for anyone using LLMs as chemistry/materials reasoning cores or studying interpretability of factual knowledge.

Unsatz benchmarks LLM agents on scientific discovery where they must construct the model ansatz themselves rather than being handed variables and candidate terms. A notable step toward testing genuine scientific reasoning beyond constrained equation-fitting.

A ready-made Claude Code setup with 23 opinionated tools spanning CEO, designer, eng manager and QA roles. Interesting as a real-world template for multi-role agent orchestration from a prominent VC.

Qualcomm details a 150W data-center AI accelerator with 128GB of memory at 548GB/s, positioning against NVIDIA and other inference chips. Interesting for teams evaluating inference hardware economics, though a vendor product page limits depth.

Simjecture is an open-source framework where LLM agents pursue physics conjectures, applied to excess radiated energy in a Z-pinch. Interesting as a transfer of AI-assisted math discovery to empirical physics research workflows.