AI Papers

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.

1. Gated DeltaNet-2 splits the delta gate, lifting 1.3B multi-key retrieval from 54.0 to 72.6

   Ali Hatamizadeh, Yejin Choi, Jan Kautz · cs.AI

   Gated DeltaNet-2 splits erase from write with channel-wise gates. At 1.3B/100B it leads KDA and Mamba-3, lifting 1K multi-key retrieval from 54.0 to 72.6.

2. Attention-as-classification: 1% labels match full-data PredCls on Visual Genome

   Sahand Sharifzadeh, Sina Moayed Baharlou, Volker Tresp · AAAI Conference on Artificia · cs.CV

   Schemata turns scene-graph classification into attention over class embeddings and injects them as prior. With 1% labels plus BYOL, PredCls R@100 hits 65.68 vs 65.7 on full data.

3. Linear associative memory has a sharp top-1 threshold at d² = 2 n log n

   Nicholas Barnfield, Juno Kim, Eshaan Nichani et al. (5 authors) · stat.ML

   A linear memory retrieves all n pairs top-1 iff d²/(n log n)>2. The log n is an extreme-value tax; without it only a percentile TAM guarantee survives at load n/d².

4. Sleep-time compute matches accuracy with ~5× fewer test-time tokens, +18% if you scale sleep

   Kevin Lin, Charlie Snell, Yu Wang et al. (7 authors) · cs.AI

   Sleep-time compute pre-reasons over context before queries, matching accuracy with ~5× fewer test-time tokens on GSM/AIME; extra sleep-time lifts accuracy up to 13% and 18%.

5. 9 of 428 LLM API routers injected malicious tool calls, including one paid proxy

   Hanzhi Liu, Chaofan Shou, Hongbo Wen et al. (6 authors) · cs.CR

   UCSB measured 28 paid and 400 free LLM routers: 9 rewrote tool calls, 17 touched AWS canaries; weak decoys soaked up 2B tokens and 440 Codex sessions.

6. Alibaba's 6B Swift-Image Leads Open-Source Unified Editing After 243K GPU Hours

   Taihang Hu, Zhao Wang, Zuan Gao et al. (21 authors) · cs.CV

   Alibaba's 6B single-stream DiT unifies T2I and image editing in 243K GPU hours. With API Prompt Enhancer it scores 4.41 overall, beating 20B Qwen-Image-Edit; 3B barely drops.

7. OpticalDNA prints genomes as OCR pages and beats Enformer with 20× fewer tokens

   Hongxin Xiang, Pengsen Ma, Yunkang Cao et al. (7 authors) · ICML 2026 · cs.CV

   OpticalDNA treats DNA as OCR documents. On 450kb eQTL it hits 0.852 AUROC with a 256k probe, beating Enformer (0.681) with nearly 20× fewer tokens.

8. Self-styled Depth Anything V4 withdrawn in two days; claimed F1 +0.044 does not hold

   Jiaming Fan, Jian Lu, Jinling Jia, Chenbin Zhang · cs.CV

   NJUPT applies Riemannian Flow Matching to 4D Gaussians and claims +0.044 F-score over a matched MLP. Withdrawn two days later for major research errors; key tables contradict.

9. Task fMRI steers LLM representations, with up to 13-point reasoning gains

   Task fMRI from 10 subjects steers ten LLMs (1.5B-72B): +2.2 points over language labels on average, up to 13.2 on propositional reasoning.

10. MC-AIXI turns incomputable AIXI into a workstation agent that nears optima on 9 POMDPs

    Joel Veness, Kee Siong Ng, Marcus Hutter et al. (5 authors) · cs.AI

    MC-AIXI approximates AIXI via ρUCT planning and FAC-CTW prediction; it matches or beats U-Tree and Active-LZ on 9 POMDPs, 3.97 vs -0.97 for 1-ply on Extended Tiger.

11. MemFail: Stronger Models Barely Help; Hard Conditional Facts Stay Under 40%

    Ishir Garg, Neel Kolhe, Dawn Song, Xuandong Zhao · cs.AI

    MemFail stress-tests Mem0, A-MEM, SimpleMem and StructMem. None wins all tasks; all stay under 40% on Conditional-Facts Hard, and stronger internal models barely help.

12. Frontier labs: CoT monitoring is a useful safety layer, and easy to lose

    Tomek Korbak, Mikita Balesni, Elizabeth Barnes et al. (41 authors) · cs.AI

    A multi-lab memo argues CoT monitoring can catch misbehavior in reasoning models, is incomplete, and can be destroyed by ordinary choices such as process supervision.

13. Opus 4.6 scores 23.2% on PostTrainBench; official Instruct models sit at 51.1%

    Ben Rank, Hardik Bhatnagar, Ameya Prabhu et al. (7 authors) · cs.SE

    Frontier agents post-train 4B bases for 10 hours on one H100. Best is 23.2% (Opus 4.6) vs 51.1% official Instruct; on BFCL an agent hits 89% against 67%.

14. LLMs re-identify HN users at 67% recall and 90% precision; classical matching stays near 0

    Simon Lermen, Daniel Paleka, Joshua Swanson et al. (6 authors) · cs.CR

    A web-search agent re-identifies 67% of 338 HN users at 90% precision. Closed-world matching reaches 68% recall at 90% precision; classical methods stay near 0%.

15. EchoChange's dual-pass diffusion revises disaster captions, lifting RSCC ROUGE-L from 15 to 26

    Dongwei Sun, Bowen Yao, Yujie Zhang et al. (6 authors) · cs.AI

    EchoChange treats disaster change captions as discrete diffusion with dual-pass remasking, reaching 26.18 ROUGE-L and 65.34% single-error Strict Hit on RSCC.

16. Two clarifications lift product-description compliance from 4% to 38%

    Sanchayan Dutta, Sai Niranjan Ramachandran, Suvrit Sra · cs.AI

    Context actions scored by expected free energy. Models shrink OQA sets but still trail a DP oracle; two clarifications lift description compliance from 4.2% to 37.5%.

17. INDUCTION: bloated FOL formulas collapse held-out match from 93% to 25%

    Serafim Batzoglou · cs.AI

    INDUCTION: emit one FOL formula that explains a labeled concept across finite worlds. Grok 4 leads FullObs at 50.7%; bloated formulas fall from ~90% to ~20% held-out match.

18. Variance-weighted sampling beats full-set harness search at 20% of the evals

    Atsuyuki Miyai, Kiyoharu Aizawa, Toshihiko Yamasaki · cs.CL

    Samples validation tasks by Bernoulli variance and scores via inclusion probability. Hits 49.3% vs 48.6% full search at 20% budget; Terminal-Bench within 1.1 points.

19. HealthFormer matches 41 RCT directions and tracks individual DBP change at r=0.78

    Guy Lutsker, Gal Sapir, Jordi Merino et al. (10 authors) · cs.AI

    A 139M decoder-only model trained on 15k phenotyped people matches 41 RCT directions zero-shot and predicts individual 6-month diastolic BP change at r=0.78.

20. 6.12B LLM requests over a year: outputs shrink, prefix hits are bimodal, FIFO matches SOTA caches

    William Nixon, Jon Durbin, Florian Standhartinger et al. (5 authors) · cs.AI

    A year of 6.12B LLM requests: median output falls under 100 tokens, prefix hits are bimodal with 99% of repeats in 15 min, and FIFO matches SOTA caches.

21. Untrained nets beat backprop at V1 RSA by 0.044 at 224px; the gap is ~0 at 32px

    Nils Leutenegger · q-bio.NC

    CNNs trained at 32px reverse V1 RSA ranking: the untrained-minus-backprop gap grows from ~0 at 32px to +0.044 at 224px, from image detail above training resolution.

22. Coding-agent residual streams linearly encode test outcomes 25 steps ahead

    André Silva, Han Tu, Martin Monperrus · cs.LG

    Linear probes on coding-agent residual streams decode parse, tests, progress and regressions (AUC up to 0.83). Signal stays above chance about 25 steps before the edit hits disk.

23. Typed wreath products of Z decide Transformer length generalization on regular languages

    Andy Yang, Blerta Veseli, Corentin Barloy et al. (8 authors) · cs.FL

    Poly-time check decides which regular languages sit in C-RASP via typed Z wreath products. On 125 languages, GPT-2 stays near-perfect to 10x train length inside, collapses outside.

24. TMI recovers interleaved computer-use tasks as goal+control-flow trees, 74.9% step fidelity

    Yucheng Jiang, Zora Zhiruo Wang, Ruishi Chen, Diyi Yang · cs.CL

    TMI builds goal+control-flow task models from interleaved computer-use traces, recovering 74.9% of steps vs 30.3% baseline; skills from them raise held-out accuracy 30% relatively.

25. Persona SFT lifts Llama 70B truth probes by 0.05; Emergent Misalignment lifts them by 0.28

    Benjamin Sturgeon, David Africa, Sid Black · cs.CL

    Prompting, ICL and SFT change Llama 70B outputs with a +0.05 truth-probe lift; Emergent Misalignment lifts historical-evil claims by 0.28, OCT in between.

26. Fake -cillin names elicit antibiotic answers; patching pins the shortcut to layers 2-10

    Kaijie Mo, Thomas Yang, Chantal Shaib et al. (9 authors) · cs.CL

    Nine LLMs treat fake affix-bearing names as real drugs. Of 653 drugs, Qwen marks 206 as affix-driven; medical fine-tunes overgeneralize most. The shortcut sits in layers 2-10.

27. LoopsBench: Opus-4.7 with Claude Code outer continuation still solves only 25% of tasks

    Han Li, Zhemin Fang, Rili Feng et al. (11 authors) · cs.SE

    LoopsBench turns 112 real coding tasks into dependency DAGs. Opus-4.7 with Claude Code and outer continuation resolves 25%; plans miss prerequisites and regressions persist.

28. Aletheia writes a publishable math paper unassisted; 13 of 700 Erdős solves are real

    Tony Feng, Trieu H. Trinh, Garrett Bingham et al. (28 authors) · cs.LG

    Aletheia generates, verifies, and revises proofs in natural language, producing one unassisted publishable paper; only 13 of 700 Erdős answers hit the intended statement.

29. AVO Makes the Coding Agent the Variation Operator, Beating FA4 by 10.5% on B200

    Terry Chen, Zhifan Ye, Bing Xu et al. (23 authors) · cs.LG

    NVIDIA's AVO uses a coding agent as the variation operator. In 7 days on B200, MHA kernels hit 1668 TFLOPS, beating cuDNN by 3.5% and FA4 by 10.5% on causal attention.

30. Half of agent patches never touch learning; those that do score 0.226 vs 0.126 on AI4AI-Bench

    Yizhe Chi, Wenyi Li, Deyao Hong et al. (10 authors) · cs.AI

    Agents get 4 hours to rewrite 10 frozen trainers; patches rerun up to 12 hours. Mean score 0.166 (shipped=0.1). Learning-side patches average 0.226 vs 0.126 for run-side only.

31. A latency-aware interconnect FoM, and 3D optical vias aiming past 10 TB/s/mm²

    UC Davis and TAMU add latency to the interconnect FoM and propose TSOV 3D photonic chiplets; they project >10 TB/s/mm² vs UCIe-3D and report 496 fJ/bit.

32. ICL Ciphers: a bijective token shuffle beats irreversible noise by 7.6 points

    Zhouxiang Fang, Aayush Mishra, Muhan Gao et al. (5 authors) · cs.CL

    JHU recodes ICL inputs with substitution ciphers; a bijection beats irreversible shuffle by 7.6 points on Amazon 20-shot, small but consistent on four datasets and six models.

33. CPO review maps energy from ~10 pJ/bit 2D packages to sub-100 fJ monolithic photonics

    Nature Electronics review maps CPO energy from ~10–20 pJ/bit in 2D to <100 fJ/bit monolithic; compute scales ~3×/2yr, interconnect bandwidth 1.4×.

34. SAFFRON: only Novae among six spatial FMs recovers global gene-expression gradients

    Johns Hopkins' SAFFRON SAE finds only Novae captures spatial gradients (median |ρ|=0.72 vs 0.39 gene SAE); no spatial FM beats PCA on local plaque microenvironments.

35. FlowEvo compiles successful workflows into skills, hitting 85.6% on ALFWorld at one-third the tokens

    Zeyu Ren, Ling Yue, Ran Li et al. (8 authors) · COLM) 2026 · cs.AI

    FlowEvo compiles successful workflows into skills and suppresses negative transfer. GPT-4o-mini leads 8 baselines on five full splits: 85.6% on ALFWorld at one-third the tokens.

36. TinyCast emits a forecast distribution from 146K params by computing seasonality

    Armin Steinhauser · cs.LG

    TinyCast is a 146K-parameter, attention-free zero-shot forecaster that computes periods and emits nine quantiles, posting nWQL 0.545 on GIFT-Eval.

37. When models stay silent or lie, RelP beats black-box by 5 points; SAE does not

    Ziqian Zhong, Aashiq Muhamed, Mona T. Diab et al. (5 authors) · cs.LG

    CMU plants hidden trees in 720 Gemma-2-2B models. When they stay silent or lie, RelP beats a 10-sample black-box by 3-5 points; SAE and circuit tracing do not.

38. Activation oracles and SAEs give no uplift on wild LLM bugs; CHIVE predictors match Opus

    Adam Karvonen, Euan Ong, Subhash Kantamneni, Samuel Marks · cs.LG

    CHIVE explains wild LLM bugs with counterfactual edits. Activation tools add no uplift over the transcript; training on those labels lets models match Opus on held-out hint tests.

39. Retrieving thinking traces beats web corpora, lifting Flash 56.3% on AIME

    Negar Arabzadeh, Wenjie Ma, Sewon Min, Matei Zaharia · cs.IR

    Berkeley RAG over thinking traces (T³) lifts Gemini-2.5-Flash on AIME 2025–2026 from 53.3 to 83.3 (+56.3%), beating web-corpus retrieval.

40. Fine-Grained Negative Queries Trip MLLMs; FINER-Tuning Recovers Up to 24.2% Paired Accuracy

    Rui Xiao, Sanghwan Kim, Yongqin Xian et al. (5 authors) · CVPR 2026 · cs.CV

    FINER probes MLLMs with fine-grained negative queries. InternVL3.5-14B paired accuracy on CompreCap Multi-rel rises from 47.0% to 71.2% (+24.2).

41. PCTree Turns DSpark Chains Into Parent-Conditioned Trees, Adding Up to 29.5% Extra Speedup

    Zixian Li, Tong Li, Chi Xie et al. (5 authors) · cs.CL

    PCTree turns DSpark's chain into a parent-conditioned tree, no retraining. Extra AR speedup at B=7 is 3.1%-29.5%; Qwen3-4B GSM8K at B=16, acceptance 9.41 to 11.16.

42. τ0-VLA Searches Subtasks with a World Model, Lifting Long-Horizon Success to 45%

    Xiaowei Cai, Yunuo Cai, Bingao Chen et al. (39 authors) · cs.RO

    τ0-VLA searches subtasks with a world model, then a 40k-hour VLA executes. Hierarchy lifts long-horizon success from 27.5% to 45%; TTC lifts Book Organization from 6/10 to 9/10.

43. FireRedTTS3 Distills a Frozen Audio Encoder, Hits 3.04% Avg Error on Seed-TTS-Eval

    Feiyu Shen, Kun Xie, Yichen Wu et al. (11 authors) · cs.SD

    FireRedTTS3 injects semantics into continuous latents via a frozen audio encoder. Base hits 3.04% avg error and 78.8% SIM on Seed-TTS-Eval; Instruct unifies design and editing.

44. QuoteBench: GPT-5.6-sol's -3.6 matched gap hides -64.3 transport damage

    Shangao Li, Yao Zhang, Volker Tresp, Yuanyuan Yang · cs.AI

    QuoteBench splits contract from transport on 56 Bash tasks. A fixed reply through one added parser drops 55.4-73.2 points; disclosure recovers 30.4-60.7 for six configs.

45. One affine match plus IMU gravity solves upright PnPf, faster than 4-point solvers

    Marcus Valtonen Örnhag, Alberto Jaenal, Stefan Adalbjörnsson · cs.CV

    UP1PfAC solves upright pose and focal length from one affine match; UP2PfORI needs two orientation features. Both cut RANSAC sampling from four points to one or two.

46. RIPE++ trains keypoints from positive pairs only, lifting MegaDepth AUC@5 to 56.58

    Johannes Künzel, Peter Eisert, Anna Hilsmann · ECCV 2026 · cs.CV

    RIPE++ pays inliers and penalizes outliers inside positive pairs, dropping negatives. MegaDepth AUC@5 goes from RIPE's 53.47 to 56.58, then 59.65 with weakly supervised LightGlue.

47. S4 and Mamba sit in TC0: one-layer RNNs learn A5, they need growing depth

    William Merrill, Jackson Petty, Ashish Sabharwal · ICML 2024 · cs.LG

    S4 and Mamba, like transformers, only express TC0. One-layer RNNs and IDS4 learn A5 at any length; S4/Mamba need depth that grows with sequence length.

48. Frozen DeepSeek-V4-Flash rewrites its own harness, BabyAI 42 to 81 on BALROG

    Tailin Zhou · cs.AI

    HSI uses one frozen DeepSeek-V4-Flash to rewrite its own harness. BALROG BabyAI rises 42.0 to 81.3, Crafter 11.6 to 44.6; NLE stays near 0.

49. Baker lab's SAPP tests hundreds of designed proteins a day at about $24 each

    SAPP runs arrayed expression and SEC in 48 hours at ~$24/design, median 96% clonal purity on 929 reactions. DMX cuts gene cost ~5x and produced nM RSV binders (cb13).

50. Consumer health AI is taking over triage and fulfilment; accountability must follow the pathway

    ChatGPT Health, Amazon, Ant's Afu and Claude Healthcare wire LLMs into records, booking and pharmacy. Topol and Keane: public health must govern pathway control, not model scores.

51. Pandora's Router matches exhaustive LLM routing while opening far fewer expensive estimators

    Adam Fisch, Shubhendu Trivedi, Fantine Huot et al. (8 authors) · cs.AI

    Routing as Pandora's Box: cheap kNN always runs; costly scores fire only when VoI covers cost. Three domains track exhaustive scoring; EmbedLLM 0.386 vs 2.356 for g-only.

52. Gemini 2.5 Pro hits 85.6% on IOAA theory; geometry still costs all models 15-26 points

    Five models sat 57 IOAA problems (2022-2025). Gemini 2.5 Pro scored 85.6% on theory, GPT-5 84.2%, both gold; only GPT-5 held 88.5% on data analysis. Geometry cost 15-26 points.

53. GOAG Learns a Gripper Contact Manifold and Hits 86.93% on MultiDex

    Julien Merand, Boris Meden, Mathieu Grossard, Liming Chen · IROS) · cs.RO

    GOAG trains a CVAE on gripper kinematics and injects objects only at inference. MultiDex mean success is 86.93% across three hands; dataset generation takes one GPU-hour.

54. CoToGrasp Conditions Dexterous Grasps on Contact Topology Without Object Labels

    Julien Merand, Boris Meden, Liming Chen, Mathieu Grossard · ECCV) · cs.RO

    CoToGrasp generates 21-topology grasps with no object meshes in training. DexGraspNet precision success is 30.3% vs Dexonomy 10.5%; topology compliance 17.18% vs 14.28%.

55. SFT Puts Greek in the Trace; a 7.7-Point Seed Swing Swamps Accuracy

    Ayoub Kirouane, Christos Petrocheilos · cs.CL

    Three 3.6–4.0B MoEs barely move Greek accuracy after SFT; a seed change swings 7.7 points. Greek traces go from 0/1,000 to ~98%; RLVR cuts format fallback 24.1% to 2.5%.

56. DeepMind: AGI May Arrive as a Market of Sub-AGI Agents, Not One Model

    Nenad Tomašev, Matija Franklin, Julian Jacobs et al. (5 authors) · cs.AI

    DeepMind argues AGI may first emerge from coordinating sub-AGI agents, and proposes a four-layer safety stack around insulated virtual agent sandbox economies.

57. Prompting Hits a Model's Stated Year 94.6% of the Time, Not Its Verb-Tense Year

    Suze van Adrichem, Aditi Bhaskar, Diyi Yang et al. (5 authors) · COLM) 2026 · cs.CL

    Two probes split the current year: prompting hits the stated year on 94.6% of 351 targets, the verb-tense year on 1.7%. SFT and weight edits cannot move both at once.

58. GigaBrain-0.7's 37k-hour three-system VLA hits 74% on humanoid complex tasks

    GigaBrain Team, Angen Ye, Axiang Sun et al. (59 authors) · cs.RO

    GigaBrain-0.7 is a three-system VLA pretrained on 37k hours across 16 robots; post-training reaches 74.1% on Maker H01 complex tasks vs 45.2% for π0.5.

59. Specify the copula, then TDC still finds noisy functional dependence that Pearson misses

    Gautier Marti, Frank Nielsen, Philippe Donnat · cs.LG

    Empirical copulas plus EMD measure intra/inter MTS dependence; TDC stays near power 1 on noisy quadratic and step patterns until noise 30-40, while Pearson sits at 0.

60. On NARU's 146.8 hours of Japanese video, open MLLMs stay under 40% MCQ accuracy

    Yuheng Huang, Jianlang Chen, Jiayang Song et al. (9 authors) · cs.CV

    NARU tests Japanese long-video narrative and cultural QA on 1,481 items over 146.8 hours. Gemini-3-Flash scores 76.2%; the best open model hits 39.8%, Flash 57.4% on subtext.