Peak and see
Building an open dataset to study what humans actually consider AI slop. SlopFinder presents random AI-generated text samples with a single rating control, streaming crowdsourced human evaluations directly to Hugging Face.
Read article →Six real architecture changes and nearly 4x the capacity: MMDiT joint attention, T5 conditioning, 2D RoPE, QK-norm + SwiGLU, and REPA. FID drops to 23.62 at cfg 3.0, beating v5's 34.06 by over ten points at every guidance scale tested.
Read article →Same DiT as v4, 36x more training data: FID drops 39.54 → 34.06 and CLIP score climbs to 29.27, within 0.78 of the real-photo ceiling. Nothing in the network changed.
Read article →Extending implicit coordinate conditioning to (x, y, z) space to generate resolution-independent 3D shapes with 148k parameters.
Read article →v4 replaces bag-of-words mean-pooling with a 2-layer text Transformer and spatial cross-attention, breaking FID 320 at native 256px.
Read article →v2: scaled up, v3: a different architecture
Read article →All three tiers remade: 300 balanced items each, one unified schema scored by generation and log-likelihood in a single run, machine-verified Math, blind cross-checked answers, full per-item provenance. 6-2026 stays frozen; the leaderboard now defaults to the latest generation.
Read article →Deduped Easy (269→238) and Mid (155→143), fixed a broken barber-paradox item, unified the Effortless taxonomy, and corrected two metric labels. Existing leaderboard scores are flagged stale pending re-eval.
Read article →The universal evaluator is live: download script.py from the leaderboard Space, point it at any HF model, get every benchmark scored with full category/subcategory breakdowns — and a ready-to-paste models.json entry.
8.5× fewer parameters, better FID and CLIP Score, any-resolution output — and all 23,747 weights still live in one PNG, now the size of a thumbnail.
Read article →the model is not being taught how to reason step by step. Instead, it learns which reasoning mode to activate before solving the problem. a lightweight direction module to predict a latent reasoning vector using a regression-style latent-space objective.
Read article →We hear you. Every time. We decided its time to talk about it.
Read article →Finally, a great test for AI.
Read article →Evaluation of Qwen/Qwen2.5-0.5B on bench-labs/bench-mid-6-2026 using lm-eval multiple-choice
loglikelihood scoring (CPU, bf16).
It was good, but not perfect. An easy-tier benchmark. Bench-labs/bench-easy-6-2026 tested on Qwen 1.5B
Read article →The code is present.
Read article →We seek colaborators to utilize our benchmarks and work with us. More information at spaces/Bench-labs/Members
Read article →Open the article to see who's new..
Read article →An effortless-tier benchmark, 240 rows total, see at Bench-labs/bench-effortless-6-2026
Read article →A small post
Read article →A neural network where the weights are literally encoded in the pixels of a PNG image.
Read article →Why we started Bench Labs, what we mean by generalization, and what we're planning to build.
Read article →Throwaway post exercising every styled element.
Read article →All rights reserved to @bench-labs