Original, structured explainers for the AI papers worth your time — from arXiv, Hugging Face daily papers, and the papers researchers are actually discussing. Each one covers the problem, the method, the numbers and the limitations.
CohAlign predicts, from the circuit and noise model, the first-order decay of one QNN gradient's second moment. At eight qubits the alignment RMSE across eight channels is 0.009.
HAL trains 10 rat cortical cultures with random stimulation until the next burst if ignition is off-goal. Goal-side bursts reach about 70-80% in two days, between sessions.
SpecTr-GBV checks K independent drafts as blocks. On DeepSeek-33B, mean speedup is 2.12× versus 1.64× for standard speculative decoding, and verification drops by over half.
CodedLang span-annotates coded wording in 900 of 7,744 Chinese Google reviews. Best detection F1 is 0.78, and GPT-5-mini rating error rises from 0.52 to 0.68.
Tohoku's DigHyd mines papers with LLMs, then humans recompute enthalpy and entropy from PCT curves. Capacity test R² is 0.868; pressure error still exceeds 10×.
Zhanshuo Cao, Jiakai Tang, Yuqi Zhou et al. (5 authors) · cs.IR
LoopRec reuses one block and exits each item alone. All 48 comparisons beat fixed-depth models; mean NDCG@20 rises 11.14% on sparse Patio, at 0.41-0.76x compute.
Chenxu Zhou, Zhiyuan Yu, Yudong Jin et al. (7 authors) · SIGGRAPH Asia 2026 Conference Pa · cs.CV
RealMe adapts FLUX into multi-view diffusion so eight 720p captures reconstruct 2K volumetric video. On RealVV, novel-view PSNR rises 8.22 dB over FTGS, from 19.05 to 27.27.
Arnon Mazza, Elad Levi · cs.AI
SCOPE compiles policies into neural predicates plus Python composition code, letting cheap non-reasoning models match frontier judges at 9-47x lower inference cost.
Hyunji Alex Nam, Michael Dennis, Natasha Jaques, Minsuk Chang · cs.AI
Using a proposer, solver, and feedback model, TURTLE evolves AppWorld tasks until Gemma 4 31B's hard-task share rises from 19.8% to 52.5% and solve rates cluster at 40-60%.
Ming Zhong, Stacy Li, Nicholas Carlini, Matthew Jagielski · NeurIPS 2026 E&D Track · cs.AI
Four agents with 8 H200s and 12 hours beat the published best on all 20 science benchmarks. Opus averages a 10.4% gain, 20.1% with plug-in skills, mostly via ML engineering.
Rujun Han, Yufan Zhuang, Arijit Ray et al. (13 authors) · cs.AI
PLE scores agents on 85 billable electrical, civil, structural and mechanical jobs. Claude Opus 5.5 in Claude Code averages 50.6%, at 27M tokens and 63 minutes a task.
Jonathan T. Barron, Dan Gnanapragasam, Michael Tschannen · cs.LG
DeepMind rewrites softmax and cross-entropy in log space to fix two rounding bugs; Slingshot spikes vanish and grokking reaches 99-100% test accuracy in bf16/fp32/fp64.
Stanford injected strings into Pythia pretraining. One exposure peaks at 12±3.7 tokens on 6.9B; later checkpoints memorize more, and unlearning still leaks 10 to 15 tokens.
Zhenyu Zhang, Kewei Fu, Yongqi Tong et al. (7 authors) · cs.CL
Heartbeat finishes tool calls while the user still speaks. On FDBench v3, coverage rises from 34.67% to 50.67% and response latency falls from 6.90s to 4.39s.
Hita Kambhamettu, Bhavana Dalvi Mishra, Andrew Head et al. (7 authors) · cs.HC
LitPivot clusters papers around one facet of a draft and suggests rewrites. Expert grounding rose from 2.59 to 5.18 versus chat-with-papers, in a 17-person study.
Tingwen Du, Haoyu Guo, Hanxue Zhang et al. (12 authors) · cs.RO
SERA freezes the VLA and edits a harness around gpt-5.6-sol. On 16 RoboMME tasks, selected harnesses reach 51.38% success, up 7 points from 44.38%.
Zhengye Bao, Shengtao Zhang, Haokai Zhang et al. (6 authors) · cs.AI
Dense retrieval plus reranking over raw chat history beats 12 agentic memory systems in 6 of 7 benchmark-model settings, at an order of magnitude lower ingestion cost.
Dresden neurosurgery randomized 300 adults to a checked AI lay summary or usual care. Median visit time fell from 33 to 25 minutes; mean interaction score rose from 4.25 to 4.42.
Chuanmeng Bian, Daren Chen, Peixin Chen et al. (48 authors) · cs.SD
Qwen-Audio-3.1-ASR cuts seven-language macro error from 22.50% to 15.09%; Next reaches 12.93% cpWER on diarized meetings.
Zhenwei Tang, Di Bai, Jintao Liu et al. (6 authors) · cs.AI
Five seller models, told only to be picked more often, inflate tool-field claims. Gemini 3.8 Flash drops Amazon success from 89% to 30%; the other four mostly keep it.
ProtTale writes function text from sequence and scores trust with pJay. On 1,031 unseen enzymes the filter lifts short-protein precision from 25.8% to 45.9%.
A latent neural ODE names Ctcf, Bcl6 and Stat3 as Arid1a rescues in germinal-center B cells; all three double-het mouse lines restore the dark-zone ratio.
Alejandro H. Artiles, Martin Weiss, Levin Brinkmann et al. (9 authors) · NeurIPS 2026 Main Track · cs.AI
Maximizing coherence while minimizing availability to author communities, an idea-atom sampler out-diversifies frontier LLMs at equal quality and hits 2025 papers at 35.9x chance.
Tianming Liang, Junyi Pan, Yuxi Li et al. (8 authors) · cs.CV
WeVisOne-VL (2B/4B on Qwen3-VL, plus a 9.6M decoder) unifies detection, segmentation, and matting. The 4B scores 71.3 mean F1 on Dense200 and 67.4% absent-object rejection.
Wen Xue, Yong Liu, Xueheng Li et al. (8 authors) · cs.CV
LayerEditor emits an edited canvas plus ordered RGBA layers in one pass. On ScaleEdit, SSIM goes from 0.8925 to 0.8839 once stacked (~1%), vs 11.6% for Seedream edit-then-split.
A. Jian Yang, An Huang, Bo Dan et al. (177 authors) · cs.CL
An open 320B/15B MoE built for CLI agents: 1.93x pre-training efficiency vs a Qwen3-style baseline on identical data, and 84.5 on CyberGym, ahead of closed models.
Tengju Ru, Weitong Lian, Zecong Tang et al. (9 authors) · cs.CV
GAFU compresses native generation features into 256 G tokens for question-time retrieval. On three UMMs it beats same-data SFT by 3.56-5.98 points and wins all 15 settings.
Meng Luo, Zechen Bai, Shanqing Xu et al. (12 authors) · cs.RO
A five-institution survey grades robot self-improvement claim by claim: behavior, later learning, and procedure improvement have hardware evidence; the recursive links have none.
Mosh Levy, Aly Lidayan, Yoav Goldberg · cs.LG
JEPA picks the feature set with the most predictive information, and reconstructing only C (β=0.36) lifts unsupervised D from 61.3% to 85.5% at copy probability 0.60.
Junha Song, Geonmo Gu, Jeongeun Park et al. (8 authors) · cs.CV
LaLa-World writes actions as text and trains a 0.8B LM for latent futures: 72.1% MPC success on five sim tasks, 7.5 points above JEPA-WMs, and 13.9× faster than five-step RAEv2.
Two rotating hexagons with four sticky ends fully bond only as the Walton-Whittaker aperiodic tiling, at 1:1. Forced growth of 2484 particles extends to a valid 2048×2048 grid.
Niharika Hegde, Subarnaduti Paul, Lars-Joel Frey et al. (7 authors) · cs.CL
Chronoberg is 2.7B tokens of English books from 1750-2000, with era VAD scores for 335,804 words. Swapping in a period synonym flips 50-75% of hate labels in 3 of 4 LLMs.
Rubin Wei, Siyuan Sun, Hao Sun et al. (8 authors) · cs.CL
DFM writes predicted chunk memory into a frozen LLM and skips retrieval at inference. On Mistral-7B the six-QA mean rises from 24.14 to 27.21, 12.71% over the best baseline.
Songlin Yang, Zhe Wang, Xiaotong Zhao et al. (11 authors) · cs.CV
On UniVideo's 13B DiT, three FAPL stages make MLLM latents drive generation. Text-only motion order rises from 48% to 61%, and single-factor edit accuracy from 42% to 58%.
One model plays five doctors and works up cases from the chief complaint. GPT-5 scored 83.2% on NEJM cases vs 36.1% for open-book doctors, and 80.3% vs 46.8% on ED admissions.
Xiaomin Yu, Hongliang Chen, Zhe Huang et al. (25 authors) · cs.IR
On 200 live-web tasks that chain a video anchor across formats and interfaces, GPT-6 Astra scores 39.0%; 1,000 human attempts score 21.1%.
Zhenghan Chen, Zekai Shao, Xin Lin et al. (5 authors) · cs.CV
A 2,104-task benchmark has 15 MLLMs write code that renders full videos from text, images or clips. Claude Opus 5.5 ranks first in all three settings; render success spans 0.52 to 0.97.
Ermo Hua, Xiangyu Hong, Baiting Wu et al. (14 authors) · cs.AI
Cross-layer Meta Experts keep per-token activation and expert width fixed, and match vanilla MoE loss with 70-80% of the experts. Deeper models save more stored parameters.
Mankeerat Sidhu, Gangmuk Lim, Hetarth Chopra et al. (13 authors) · cs.DC
Koi plans heterogeneous LLM inference with a causal graph and deterministic scoring. On 104 GPUs it delivers 1.18× the best baseline's goodput and 20% lower cost per token.
Federico Bianchi, Yongchan Kwon, Aneesh Pappu, James Zou · cs.CL
By May 2026, EinsteinArena agents set 11 new math records. The 11D kissing-number lower bound rose from 593 to 604 via shared verifiers and public discussion.
Yuhua Zheng, Xueyi Chen, Xinze Zou et al. (6 authors) · cs.LG
EdgeKernelBench: on phone CPU, Jetson, and RK3588, GPT-5.6-sol one-shot scores 64.5% Correct@1 but 25.5% [email protected] vs tuned TVM. LEKA raises that to 43.5% on 69 tasks, from 13.0%.
Ruhan Wang, Kishan Panaganti, Dongruo Zhou · cs.AI
Anti-Persistence stores trajectories, synthesizes guidance, then drops it. Nine settings average 59.9%, vs 53.5% trajectory memory and 49.0% stored rules, at lowest LLM-call cost.
Naman Vats · cs.AI
A pre-specified controller swap on 315 Terminal-Bench trials: Jev 1.13 raised Sonnet 5.5 cost per attempt 19% while the pass rate stayed 73.0%.
Jiashuo Sun, Hao Zhang, Xueqiang Xu et al. (12 authors) · cs.AI
A2D scores questions by student-teacher disagreement and compatibility, with no answers. Three math pairs beat real-bank OPD; a 1.5B student recovers 99.9% of the gap.
On an otherwise ideal simulated 256x256 PCM tile, sparse recovery over random probes localizes up to 64 stuck cells in 4 to 12 reads at 1% noise, versus 256 for a full readout.
Arnav Chavan, Kusha Sareen, Rishabh Tiwari et al. (9 authors) · cs.LG
Project every optimizer step onto the bottom 1% of each weight matrix's spectrum: forgetting drops from up to 57 points to ~1 while target-task accuracy holds or improves.
Min Cai, Dan Zhang, Gao Xing et al. (6 authors) · cs.AI
A coding agent hot-swaps the user simulator, rewards, and hyperparameters during multi-turn RL training, beating the strongest baseline by ~15% overall and 36.7% on ColBench.
My Chiffon Nguyen, Tianyi Hu, Vinod Anbalagan et al. (158 authors) · cs.CL
167 native speakers wrote 6,191 riddles across 62 communities. Of 28 LLMs, the best reaches 79.4% overall, but none exceeds 22% on Lumasaba or Mongondow.
Xingsong Ye, Yongkun Du, Jiaxin Zhang et al. (5 authors) · NeurIPS 2026 · cs.CV
TextRIRO trains scene-text editing on real images, via column-shuffle and crop self-pairs. Variable-length accuracy is 47.6% versus ~12%; the 3M set adds 1.91 recognition points.
Max W. Shen, Mark Goldstein, Zichu Wang et al. (5 authors) · cs.LG
Stopgrad regression makes the true flow map the unique stationary point of MeanFlow and slim ESD/LSD, with exponential convergence and 2× lower training memory.
Yujian Liu, Li An, Zelin He et al. (8 authors) · cs.CL
UCSB builds a 16-agent AI research team with distinct scientific profiles; in 80 iterations it discovers CFDN, a linear attention variant that beats GDN2 on every eval at 1.3B.
Patrick Haller, Filipe Laitenberger, Alan Akbik · cs.CL
PACE quantizes linear-attention recurrent states to 4 bits by rounding once every 8 tokens, within a point of 8-bit DAMP on three larger models at two thirds of the memory.
Joel Jang, Seonghyeon Ye, Minjoon Seo · cs.CL
On nine tasks, negating the prompt makes OPT and GPT-3 worse from 125M to 175B. Original-plus-negated averages stay near 50%, and the best method trails 13-year-olds by 31.3%.
Liam Dugan, Callum McDougall, Matthew Denton et al. (6 authors) · EMNLP 2026 · cs.CL
SAEs on Gemma's residual stream, filtered by auto-interp score, feed an L1 logistic head that reaches 92.2 F1 on COLING test, 17 points above fine-tuned Gemma-2-2B.
Fei Bai, Huatong Song, Chuan Hao et al. (8 authors) · cs.LG
GRPO on 18 orthogonal harnesses lifts Qwen3-4B in-space Pass@1 by 9.8 over single-harness RL, and beats OpenClaw training on Claude Code.
Houde Qian, Zheng Jiang, Wenyi Li et al. (12 authors) · cs.CL
Nine frontier LLMs build A-share strategies from raw Level-2 data on Quant-Arena; on sealed 2024, 1 of 38 runs beats the 17.1% equal-weight market and the top scorer wins by holding cash.
Jiahui Yang, Jason Jingzhou Liu, Ray Muxin Liu et al. (7 authors) · cs.RO
A batched whole-body controller reaches, three local RL experts grasp, and one distilled point-cloud policy hits 91% zero-shot success across 24 real scenes.
Tianze Xia, Xiaoling Zhou, Ruya Lan et al. (6 authors) · cs.CV
ALLOT sets each layer-step density from final-video error, with no training. On HunyuanVideo, matched-time XAttention gains 11.8 dB PSNR; nine settings tie or beat their defaults.
Yanqing Liu, Letian Zhang, Chengzhi Liu et al. (8 authors) · cs.CV
ALPS trains a ViT-B from still images and text; one token set holds state, language, and pixels. Frozen, it hits 85.5% LIBERO success and matches 3.5× larger V-JEPA 2-L.
Tianle Chen, Pushkar Singh, Ketan Todi et al. (11 authors) · cs.AI
DeLoop runs DPO on repeated calls versus later recoveries, with no reward. On compressed CodeGym, Qwen3-8B loops fall from 57.3% to 6.1%, steps about one third, success flat.