The archive · Page 3 of 6
Where architecture, training tricks, and eval methodology meet the marketing layer — separating durable progress in foundation models from leaderboard theater that quietly falls apart under load.
49–72 of 129 articles · Newest first
A new arXiv paper ports Tree-of-Thoughts prompting to text-to-image in-context learning and reports CoBSAT gains, but reasoning runs on prompt hypotheses, not image states.
FourierQK filters query and key projections before attention, cutting TinyShakespeare character-level loss by 79%, but word-level, retrieval and long-context tests are absent.
Tencent's Hy3 is billed as a 295B agent model with a 21B active token footprint, but public materials omit benchmark tables and independent DeepSeek or Qwen comparisons.
A July 2026 arXiv paper argues that scarce HBM and DRAM bandwidth, not raw compute, will determine which labs and providers can profitably serve large language models through.
A July 2026 arXiv paper shows Self-Consistency voting can hide contradictory reasoning, and GraphEVAL's graph-based coherence metrics catch flawed paths output checks miss.
LongCat-2.0 is a 1.6T MoE trained on a 50,000-card domestic cluster without NVIDIA silicon, posting agent scores near Claude Opus 4.6 yet CUDA serving stacks still dominate.
A July 2026 arXiv break-even study finds pretrained time series foundation models win unconditionally on half of 30 datasets, but lose to ARIMA or XGBoost on a fifth at any.
VLA Grounder shows frozen vision-language-action models improve when language is an optimizable input. Success rates increased from 12.6% to 38.7% for pi0 and 24.4% to 63.9%.
NightVision recovers transformer hidden dimension to within 23% error using only single logprob output and timing, proving API restrictions meant to protect model IP instead.
InduceKV keeps multimodal LLM adaptation under a fixed KV memory budget, outperforming replay and PEFT while trading selection complexity for serving simplicity.
Sonnet 5 undercuts GPT-5.5 by 60% on input tokens and leads coding benchmarks, but a new tokenizer inflates effective costs and Anthropic did not publish GPQA scores.
LLMs beat classic truth-discovery on conflicting facts, but the same reasoning treats repeated low-credibility claims as corroboration. More sources do not mean more truth.
Flexformer makes linear attention's kernel learnable by training spectral frequencies, but the abstract offers no perplexity or accuracy numbers to back its gains.
Huawei ships CUDA-free AI compute on domestic silicon today, but specific OpenPangu quantization accuracy claims on Ascend NPUs lack any readable primary source.
Multimodal RAG readers lose 16 to 26 percentage points when the correct evidence sits at the end of context, and standard rerankers do not close the gap.
A Chalmers/Tampere paper trains a CNN on EM-simulated layouts to search Doherty amplifier combiners in milliseconds. EM simulation is amortized, not eliminated.
A 70-patient synthetic EHR pipeline shows believable prose is not clinical validity. Fabricated dosages and labs poison models and invite HIPAA and FDA scrutiny.
Doubao-Seed-2.1 Pro's parity claims are vendor-graded and unreproduced. Route the Chinese flagship tier by axis: Qwen for ZH to EN, Doubao for EN to ZH, DeepSeek for code.
LiMoDE freezes a pre-trained expert bank and adds low-rank experts per new task, forcing continual robot learning to choose between architecture and replay.
A June 2026 preprint feeds the PDE residual into a neural corrector as input, not an optimization target, shifting surrogate cost from inference loops to training capacity.
Vision-language models now reason over tables, charts, and documents, but detection-era benchmarks still rank them on box localization and undercount comprehension.
ArBG reframes equilibrium sampling as one forward pass and beats flow-based generators, but MD training data and importance-sampling reweighting remain in the pipeline.
The June 2026 preprint Look-Before-Move separates what a camera observes from how it moves in dynamic 3D story worlds, but its gains stay preprint-only with no numbers.
GLM 5.2 sits within 1% of Opus 4.8 on FrontierSWE, Qwen 3.7 Max owns agentic workflows, and DeepSeek V3.2 covers cheap bulk inference. Route by workload, not by rank.