The archive · Page 4 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.
73–96 of 129 articles · Newest first
An SLM ensemble caught 10% of papers human reviewers missed, yet no model matched human accuracy. Now journals must define how much SLM-screened evidence they will accept.
Both MiniMax M3 and GLM-5.2 advertise 1M context. Only Z.ai ships benchmarks, and neither publishes independent needle-in-haystack data, so the spec sheet is not the budget.
DeepSeek V4.1 Flash led Hugging Face trending in June 2026 within one week, and five of the top ten slots went to Chinese labs. Trending measures velocity, not installs.
125 Wikipedia edits produced significant attribution signals in Llama 3.1 8B and Llama-3.2-1B, showing pretraining does not average out targeted high-weight-source edits.
A June 2026 preprint anchors continual-learning retention in sparse autoencoder features instead of weights, removing replay buffers but tying quality to SAE coverage.
A 30B Nemotron model post-trained itself to 8th of 4,000 on NVIDIA's leaderboard, then detected its own internal metric lying and rewrote its evaluation frame mid-run.
DataLearner's June 2026 snapshot ranks GLM-5.2 seventh by HLE at 54.70 and places no Chinese flagship in the overall top three, undercutting launch-day claims.
Alibaba framed Qwen3.7-Max as globally top-ranked at launch. The Artificial Analysis index scored it 56.6 and ranked it fifth overall, behind Claude and GPT.
ScaleToT distills tree-of-thought reasoning into a profile encoder that serves billion-user recommenders without per-user LLM calls, lifting LT30 by 6.738% in an A/B test.
VeriPilot wraps an LLM in a four-phase Verilog debug loop with a golden reference, lifting GPT-4o repair on CVDP from 54.3% to 85.71% in an unreproduced v1 preprint.
No preprint backs decomposed annotation as an LLM labeling-cost fix. Where measured, gains come from shrinking a model's output space, not from splitting annotation work.
A June 2026 preprint argues U-Net skip connections leak noise into flow-matching speech decoders, and a codec-supervised skip-free backbone can replace them at parity.
TND proposed per-neuron dynamics as a sequence-modeling primitive, then was withdrawn from arXiv for accuracy errors the day before coverage called it a Transformer.
PV-TAM, proposed in a June 2026 arXiv preprint, corrects two structural biases in VLM localization scoring: decoding drift and modality boundary-marker contamination.
Meituan's General 365 benchmark caps Gemini 3 Pro at 62.8% and most of 26 models below 60%, exposing how saturated public leaderboards inflate reasoning scores.
A 2026 preprint proposes LLMs as A/B test proxies for human subjects. Raw outputs captured 39% of the treatment effect, and surrogacy bias does not average out.
Token pricing decouples from GPU cost via cached discounts, reasoning markups, and per-model multipliers. No single per-token rate works as a budget unit for real workloads.
An LLM judge's self-preference claim holds only if the study fixed generation quality. Without that control, the bias number conflates real preference with a quality gap.
CADBench evaluates eleven systems on 18,000 CAD samples and finds vision-language models unreliable, a gap SWE-bench and LiveCodeBench were never built to measure.
ByteDance's Doubao-Seed-2.1 Pro launch deck claims parity with GPT-5.5 on benchmarks ByteDance graded itself, and no independent source has reproduced the scores yet.
The RoboSSM preprint swaps the transformer backbone of in-context robot imitation for a Longhorn state-space model and claims LIBERO gains. The full paper is still pending.
Routing architecture decides whether MoE experts can be pruned safely: hard routers keep calibration per expert, soft routers need an aggregate guard.
GLM-5.2 and Kimi K2.7 Code shipped a day apart as open-weight coding models. One leads the open-weights index; the other wins on price and a claimed token-efficiency gain.
Per-block linear recoverability in transformer FFNs swings from near-linear to strongly nonlinear between layers, so compression tools must probe each block per checkpoint.