The Archive

SpaceNum benchmark tests whether VLMs genuinely ground numerical outputs in spatial perception via dynamic and static reasoning tasks.

ETCHR decouples image editing from understanding in MLLMs to improve visual reasoning without predefined toolkits or noisy generation.

Complete-muE enables hyperparameter transfer between dense FFN and MoE architectures via normalized router scaling and active-width μP bridges.

Token selection strategy reduces quadratic attention cost in visual geometry transformers for 3D reconstruction by restricting key/value interactions.

Developer built text-to-speech mobile app using Claude Code, supporting PDFs, web articles, and image text with privacy-first design.

CHRONOS three-layer architecture handles temporal decay, dynamic Shapley pricing, and shared differential-privacy budgets in evolving knowledge-graph marketplaces.

Lexical intervention technique improves cross-lingual knowledge transfer for low-resource languages without parallel data or auxiliary models.

PGT generates procedural geometric tasks to improve MLLM fine-grained visual grounding and diagnose perception failure sources.

Perturbation theory for spherical Hellinger-Kantorovich gradient flows with dimension-free stability bounds.

BeeLlama v0.2.0 achieves 4-5x token throughput gains on RTX 3090 via DFlash optimizations for Qwen 27B and Gemma 31B models.

Inference-time layer looping retrofit for frozen transformers improves efficiency without retraining or architecture changes.

Two persona prompts, identical content, same model (gpt-5.2). Only difference is formatting: one prose, one bullet points. In a 10-round Prisoner’s Dilemma the prose version cooperated \~96% of the time, the bullet version \~20%. A 76pp gap, p < 0.001. Same meaning, opposite behavior. Authors call it the butterfly effect in LLM simulations. The part that matters here: CLAUDE.md, system prompts, and memory are mostly declared self-description. If formatting alone moves behavior this much, two people with the same intent get different Claudes based on how they happened to write it up. Any...

Hamiltonian gradient flow interpretation of Muon optimizer via regularized orthogonalization and Fenchel duality.

Large-scale human study on how LLM narrative explanations affect decision-making accuracy in classification tasks.

FM-CGM modular framework leverages foundation models for visual causal reasoning and counterfactual generation.

There is a harsh truth about Elon Musk's "truth-seeking" AI chatbot Grok: It's not very good, and not many people are using it. That's the takeaway of a new Reuters report, which found that Grok barely appears in federal records of how the US government used AI last year. It's not the only sign xAI's signature chatbot is in trouble, even as Musk puts it at the heart of what could be the biggest IPO in history. Reuters reviewed more than 400 examples of government AI use where specific vendors were named. Grok or xAI, it found, appeared in only three - each of those for basic uses like documen...

LLM distillation empirical finding: weak teachers improve larger students with proper loss mixing, challenging strong-teacher assumption.

Stratechery weekly roundup covering data center policy tensions, agent economics models, and tangential topics from May 2026.

Entry-wise error analysis of spectral ranking (Bradley-Terry-Luce) under semi-random adversarial edge sampling.

User reports Claude effective for UX strategy and product decision pressure-testing rather than design generation.

Trump delays AI safety testing EO, claiming it would be an innovation “blocker.”

ToolMerge: LLM-based query decomposition for keyframe retrieval in long-video QA with learned ranking merging.

Geopolitical bias in LLMs originates post-training, not pre-training; amplified by prompt language across seven labs.

Distributional theory: hypernymy geometry in word2vec emerges from co-occurrence, with eigenvectors encoding taxonomic hierarchy.

Joint Energy-Based Models isolate discriminative vs. generative objective effects on human visual alignment using controlled architecture experiments.

Claude Opus 4.7 refused to continue work on a project after detecting a potential cross-tenant logging vulnerability, raising questions about model safety behavior in production scenarios.

After Google Search's AI update, the word "disregard" now effectively breaks the search interface.

Dual-Brain architecture combines LLM orchestration with deterministic inference for O-RAN service provisioning and xApp/rApp deployment.

ByteShape releases optimized quantization for Qwen3.6-35B achieving 30% faster inference than Unsloth on 6GB VRAM.

Community fork of llama.cpp optimizes MoE inference on 12GB VRAM by loading only active experts rather than full layers.