15.2% of agent runs take an unauthorized action, 81% collision avoidance from coupled guidance, and a 6.24× log-sum-exp kernel: twelve top-conf papers from October 24–30, 2026
Twelve first-submission arXiv papers from October 24–30, 2026 — ten accepted at NeurIPS 2026, one at ICML 2026 and one at CVPR 2026 as a Highlight — covering speculative decoding on sparse mixture-of-experts models, log-sum-exp kernels, steering vectors as a training signal, calibration by head intervention, agent side effects, offline goal-conditioned RL, causal explanations of policies, 3D reconstruction and generation, flow-matching guidance, publicly verifiable watermarking and Newton boosting.
Twelve papers in this issue were first submitted to arXiv between October 24 and October 30, 2026, the week that follows the October 17–23 window of the previous issue. An entry qualifies only when its own submission record states an accepted NeurIPS, ICML, ICLR or CVPR 2026 venue.
Ten entries are NeurIPS 2026 papers. One was accepted by CVPR 2026 as a Highlight, and one by ICML 2026 as a main-track paper with a PMLR volume line in its record.
The window held 193 unique records that named a top conference somewhere in their submission record. Of those, 94 carry explicit acceptance language; 77 are workshop tracks, 17 name a conference without saying it was accepted, and 5 are review-stage or withdrawn. One ICLR 2026 poster fell inside the window and was read; it did not make the final twelve.
Of the 94 papers whose records name an accepted venue, these twelve were picked to spread across directions and to keep papers that report their headline numbers against named baselines. Each first-submission date was checked against the paper's own arXiv record.
At a glance
Entries are grouped by area and, inside each section, ordered by posting date. The reading signal gives one verifiable number per paper rather than a verdict.
| Paper | Area | Status | Reading signal |
|---|---|---|---|
| Why Speculative Decoding Works Better Than Predicted on Sparse MoE Models | LLM inference | Accepted at NeurIPS 2026 main track | MoE layers are 67% of Qwen3-30B's weights but 33% of forward-pass time at batch size 1, and the L2 hit rate rises from 5.72% to 27.79% under verification. 1 |
| Speeding up Log-Sum-Exp: Kernel Fusion at the Memory Wall, Integer Arithmetic at the Compute Wall | Numerical kernels | Accepted at NeurIPS 2026 | 6.24× faster than torch.logsumexp at N = 128,256 on a V100, and 10–11× at N ≤ 1024. 2 |
| Steering Vectors as a Training Signal in LLM Post-Training | LLM post-training | Accepted at NeurIPS 2026 | Distilling from a thinking-mode teacher into Qwen2.5-1.5B lifts MATH from 18.13% to 33.47%. 3 |
| Why Are LLMs Confidently Wrong? Correcting Overconfident Errors via Causal Head Intervention | Calibration | Accepted at NeurIPS 2026 | Mistral-7B on MedMCQA: ECE 0.0604 to 0.0383 with accuracy held at 0.4847 against 0.4867. 4 |
| Task Success Is Not Enough: Side-effect-Aware Evaluation of Tool-Using Language Model Agents | Agent evaluation | NeurIPS 2026, Evaluations and Datasets Track | 15.2% of 2,211 runs take an unauthorized side-effecting action; on ambiguity tasks the rate is 38.8%. 5 |
| Escaping Path Mirages in Offline Goal-Conditioned Reinforcement Learning | Offline RL | Accepted to NeurIPS 2026 | OGBench antmaze mean success 91% against 70% for the strongest baseline, GAS. 6 |
| CausalXRL: Explainable Reinforcement Learning through Causal Graph Reasoning | Explainable RL | Accepted to ICML 2026 main track (PMLR 306) | MicroRTS-Lite behaviour F1 85.9% against 82.2% for a decision tree, and 0.5 s to learn the graph against 20.9 s for PC. 7 |
| Learning Scene Coordinate Reconstruction from Unposed Images via Pose Graph Optimization | 3D reconstruction | CVPR 2026 Highlight | Tanks and Temples Advanced split: 16.8 dB against 14.5 for ACE-Zero and 17.3 for COLMAP, in 30 minutes against ACE-Zero's hour. 8 |
| Generative Structure from Motion with Native 3D Diffusion | 3D generation | Accepted to NeurIPS 2026 | Median Chamfer distance 2.8–3.8× lower than TRELLIS-M and novel-view PSNR 3.2–4.5 dB higher from four uncalibrated views. 9 |
| Coupled Guidance for Flow Matching | Generative modeling | Accepted at NeurIPS 2026 | Robotic collision avoidance reaches 81% success with coupled guidance against 36% for naive guidance and 21% unguided. 10 |
| PP-Mark: Provable and Publicly Verifiable Watermarking for Generative AI | Watermarking | Accepted at NeurIPS 2026 | Forged images pass at 0.33% under a 1% false-positive operating point, against 3.67% for TrustMark and 4.00% for RingID. 11 |
| Affine-invariant Cubic Newton with Weak Learners | Optimization theory | Accepted at NeurIPS 2026 | The global rate's worst-case dependence on the weak-gradient edge drops from γ¹² to a path-averaged cosine angle Θ̄k. 12 |
Inference, kernels and post-training
1. Why Speculative Decoding Works Better Than Predicted on Sparse MoE Models
Area tag: LLM inference and speculative decoding
arXiv: 2610.38981, first submitted October 28, 2026. 1
Authors / institutions: Ekagra Ranjan, Komal K. Teru, Bharat Venkitesh and Acyr Locatelli, all at Cohere. 13
Status: Accepted at the NeurIPS 2026 main track, as stated in the arXiv record. 1
Problem: The standard account of speculative decoding on mixture-of-experts models predicts a concave speedup curve, peaking at intermediate batch sizes where expert activation saturates. Recent fine-grained models with small experts do not behave that way, and the model that predicted the curve cannot say why. 13
Method: Ten sparse MoE targets from 30B to 1T parameters and 8 to 512 experts are profiled on H100 and B200 at batch sizes from one upward. Verification cost is decomposed into expert bytes, routing overlap across draft tokens and kernel-level bandwidth use, and a maximum-entropy iterative proportional fitting scheme predicts the unique expert count from overlap statistics. 13
Comparison: The activated-expert-bytes cost model and the published concave speedup prediction, against measured time-per-output-token curves for EAGLE, EAGLE-3, multi-token-prediction and draft-model speculative decoding across MT-Bench and three further datasets. 13
Results: At batch size 1, MoE layers hold 67% of Qwen3-30B's weights but take 33% of forward-pass time, and 35% on B200. Adjacent positions share 20–45% of their activated experts against 4–12% under independence, which cuts the unique experts activated during verification by 21–34%. The L2 hit rate rises from 5.72% under autoregressive decoding to 27.79% under verification, and the fitting model predicts unique expert counts within 4% error. 13
Resources: The record and the paper name no public code or data release. 1
Takeaway: Fine-grained models peak at batch size 1, while coarse models with large experts keep the concave shape, so the scheduling decision depends on expert granularity rather than on the model's size. All of it is measured on two GPU generations, and a reader serving on a different accelerator is extrapolating. 13
2. Speeding up Log-Sum-Exp: Kernel Fusion at the Memory Wall, Integer Arithmetic at the Compute Wall
Area tag: Numerical kernels and systems
arXiv: 2610.33647, first submitted October 24, 2026. 2
Authors / institutions: Lingyun Yao at Aalto University, Martin Andraud at UC Louvain, Niki Loppi at NVIDIA Helsinki, Andrea Pilzer at NVIDIA Bologna, Anji Liu at the National University of Singapore, Guy Van den Broeck at UCLA and Martin Trapp at KTH. 14
Status: Accepted at the 40th Conference on Neural Information Processing Systems (NeurIPS 2026), as stated in the arXiv record. 2
Problem:
torch.logsumexp dispatches a chain of nine sub-kernels per call and round-trips the whole input through high-bandwidth memory roughly six times, which makes it about 6× slower than a plain sum over the same tensor. On a CPU, where no special function unit exists, the exponentials and logarithms dominate the cost. 14Method: FuseLSE evaluates LSE in a single pass, computing exp and log on the GPU's special function unit and reducing with warp shuffles. IntLSE replaces the floating-point evaluation with integer arithmetic and a 64-entry lookup table in constant memory for CPU use, and both kernels drop in where the PyTorch call sat. 14
Comparison:
torch.logsumexp, torch.sum as the memory-bound reference, TensorFlow's reduce_logsumexp, and the LSE cost inside a GPT-2 language-model head, an RLHF sequence-log-probability stage, a PyJuice circuit and a particle filter. 14Results: On a V100-PCIe, both kernels beat
torch.logsumexp by about 10–11× for N ≤ 1024 and settle at 6.24× and 6.18× for N = 128,256, with HBM traffic falling from about six times the input to about one. On a Xeon Gold CPU, IntLSE is 3.1–6.1× faster than FuseLSE, and 2.8–5.2× faster than PyTorch in the particle-filter workload. FuseLSE holds median relative error below 10⁻⁴ %, while IntLSE stays under 0.2% on the U(−100, 0) and U(−500, 0) ranges but rises to 2.38% on U(−1, 0). 14Resources: Code is public at the authors' repository. 14
Loading content card…
Takeaway: The GPU gain comes from layout rather than approximation, so FuseLSE is the safe swap. IntLSE is the accuracy-and-throughput trade, and its error grows exactly where log-values cluster near zero, which is the case to check against a specific workload. 14
3. Steering Vectors as a Training Signal in LLM Post-Training
Area tag: Post-training and distillation
arXiv: 2610.40998, first submitted October 29, 2026. 3
Authors / institutions: Tiejin Chen, Huaiyuan Yao and Hua Wei at Arizona State University, with Maunil Vyas at Amazon. 15
Status: Accepted at NeurIPS 2026, as stated in the arXiv record. 3
Problem: Steering vectors are normally added to a frozen model's hidden states at inference time. The paper asks whether the same direction can steer the parameter update itself during post-training, and where that stops working. 15
Method: Two intervention regimes. Supervised training gets a hidden-state hook that injects the vector during teacher forcing, on a per-sample direction computed by spherical interpolation. RLVR gets an advantage-repair step that supplies a correctness-contrast direction to wrong rollouts, including all-wrong groups that would otherwise yield zero gradient. 15
Comparison: Plain SFT and plain distillation with LoRA, three students, four teachers, four datasets; a random Gaussian control and a PCA-based variant isolate which part of the direction does the work; RLVR against GRPO with entropy-aware advantage shaping. 15
Results: Distilling from thinking-mode teachers is where the gain sits: on MATH into Qwen2.5-1.5B, accuracy rises from 18.13% to 33.47% with the per-sample spherical injection, and into the 7B student from 42.47% to 46.20%. GSM8K into the same 1.5B student goes from 49.58% to 60.42% with the Gemma-4-26B thinking teacher. No-thinking teachers are flat across the grid, and the random Gaussian control hurts. 15
Resources: Code is public at the authors' repository. 15
Loading content card…
Takeaway: The gain tracks the off-policy gap, which is why a thinking teacher pays and a no-thinking teacher does not. Three boundaries are reported: code generation breaks the direction, cross-family layer transfer breaks the placement, and keeping the vector at inference on a vector-trained checkpoint breaks the regime. 15
Calibration and agent evaluation
4. Why Are LLMs Confidently Wrong? Correcting Overconfident Errors via Causal Head Intervention
Area tag: Calibration and model internals
arXiv: 2610.34994, first submitted October 26, 2026. 4
Authors / institutions: Jing Ren, Bowen Li, Ziqi Xu, Xuechao Yang and Feng Xia, at RMIT University. 16
Status: Accepted by NeurIPS 2026, as stated in the arXiv record. 4
Problem: Post-hoc calibration adjusts the output distribution without asking which internal components produced the miscalibration. The paper separates a model's wrong-but-confident answers from its right-but-confident ones and looks for the heads responsible. 16
Method: Predictions are sorted into four correctness-by-confidence quadrants. Attention heads are then scored by necessity and sufficiency on the wrong-high and correct-high inputs, and a ranked union of error heads and noise heads is masked at inference, with no weight updates. 16
Comparison: Temperature scaling, ITI, and two further temperature-based baselines labelled Therm. and ATS in the paper, on LLaMA-3.1-8B, Gemma-2-9B and Mistral-7B-v0.1, over MMLU, MedMCQA, OpenBookQA and MedQA-USMLE, plus SciQ in the appendix, scored by accuracy, expected calibration error, negative log-likelihood and Brier score. 16
Results: On Mistral-7B, MedMCQA expected calibration error falls from 0.0604 to 0.0383 while accuracy holds at 0.4847 against 0.4867. On LLaMA-3.1-8B, MMLU error falls from 0.0569 to 0.0490 while accuracy rises from 0.6007 to 0.6120. On Gemma-2-9B, MedMCQA accuracy rises from 0.5607 to 0.5873 with error from 0.0556 to 0.0380. Attribution uses fewer than 50 labelled samples, and the mask size is 6 to 8 heads. 16
Resources: A project page is listed in the paper. 16
Takeaway: The interesting comparison is internal: masking the correct-high heads alone lowers error to 0.0694 but costs accuracy, while the class-conditioned union keeps both. The evaluation is multiple-choice question answering on three 7B–9B models, so nothing here speaks to long-form generation or to models above that size. 16
5. Task Success Is Not Enough: Side-effect-Aware Evaluation of Tool-Using Language Model Agents
Area tag: Agent evaluation
arXiv: 2610.34167, first submitted October 25, 2026. 5
Authors / institutions: Jiaju Huang, Xinyu Ma, Hao Yang, Wei Ke, Tao Tan and Yue Sun at Macao Polytechnic University, Shaobin Chen at Southwest Medical University, and Xinglong Liang at the Netherlands Cancer Institute and Radboud University Medical Center. 17
Status: Accepted at NeurIPS 2026 in the Evaluations and Datasets Track. 5
Problem: An agent that finishes the task can still write to the wrong object, widen the scope of the request or contact the wrong recipient. Final-success benchmarks score the first fact and miss the second. 17
Method: SABench holds 201 email, calendar and ticket tasks, crossing four state structures with four trap families, including ambiguity, mis-targeting, cascade harm and overreach. Each task carries a deterministic oracle that scores goal satisfaction and authorization compliance separately from the trace. 17
Comparison: Eleven frontier models from commercial and open families, 2,211 model-task pairs, with collateral damage measured against goal success per model rather than as a single aggregate. 17
Results: 15.2% of all runs take an unauthorized side-effecting action. Ambiguity tasks produce a 38.8% violation rate against 8.4% for the non-ambiguous traps, and 12.8% of goal-satisfied runs still write without authorization. Cross-tool and dependent states run at roughly twice the rate of independent states, and the damage rate varies 4.7× across models whose goal rates are comparable. 17
Resources: The task corpus is on Hugging Face and the traces, oracle and analysis code are in a public repository. 17
Loading content card…
Takeaway: Two models with comparable task-success rates differ 4.7× in how often they act without authorization, and the gap is widest on the ambiguous requests. The benchmark covers three office tools simulated deterministically, so an agent wired to real APIs meets a wider set of failure modes than these 201 tasks contain. 17
Reinforcement learning
6. Escaping Path Mirages in Offline Goal-Conditioned Reinforcement Learning
Area tag: Offline goal-conditioned RL
arXiv: 2610.40130, first submitted October 29, 2026. 6
Authors / institutions: Junhyeon Bae, Jaebak Hwang, Gwanwoo Choi, Minung Kim and Seungyul Han, at the Graduate School of Artificial Intelligence, UNIST. 18
Status: Accepted to NeurIPS 2026, as stated in the arXiv record. 6
Problem: In stochastic long-horizon offline data, some trajectories reach the goal by luck. Value learning treats them as evidence of a usable path, and the agent commits to a route that will not repeat. The paper calls this a path mirage. 18
Method: Branch-aware graph planning puts graph nodes at high-variability points in the data, then builds goal-conditioned edges with segmented learning so unreliable branches are cut and the remaining path carries less stochasticity. Planning runs on the resulting graph rather than on a single learned value function. 18
Comparison: HIQL, GCIQL, QRL, CRL, TMD, CGCIVL and GAS on OGBench, over the large, giant and teleport maze layouts on antmaze, pointmaze and visual-antmaze, with navigate, stitch and explore datasets, five seeds per cell. 18
Results: Mean success is 91% on antmaze against 70% for GAS and 43% for HIQL, 98% on pointmaze against 69% for CGCIVL, and 86% on visual-antmaze against 78% for GAS. The largest single gap is teleport-navigate, at 90% against HIQL's 42%. One cell goes the other way: on large-navigate, GAS scores 93% to BGP's 90%. 18
Resources: Code and project materials are at the authors' project page. 18
Takeaway: The gains come from the branching and teleport layouts, which are the ones built to produce mirages; on a layout without them the method is level with the strongest baseline. A reader with offline data that has no branching structure should expect less than these averages suggest. 18
7. CausalXRL: Explainable Reinforcement Learning through Causal Graph Reasoning
Area tag: Explainable RL
arXiv: 2610.42028, first submitted October 30, 2026. 7
Authors / institutions: Yanming Zhang, Eric Papenhausen and Klaus Mueller, at the Department of Computer Science, Stony Brook University. 19
Status: Accepted to the ICML 2026 main track; the arXiv record's journal reference gives PMLR 306. 7
Problem: Deep RL agents are trained on pixels and act for reasons nobody can inspect. Post-hoc saliency and reward decomposition explain individual decisions without giving a structured account of which state variables drive an action. 19
Method: State features and actions become variables in a structural causal model, fitted on a held-out replay buffer. Local causal models explain single decisions, a global graph summarises the policy, and the graph is learned by a structural learning algorithm that avoids conditional independence tests. 19
Comparison: Behaviour-prediction accuracy and F1 against a linear-model and a decision-tree explainer, and structure-learning time against the PC and FCI algorithms, on CartPole, LunarLander, PongDuel and MicroRTS-Lite, followed by a qualitative expert study on MicroRTS. 19
Results: On MicroRTS-Lite, behaviour accuracy is 86.0% against 80.3% for the decision tree and 75.3% for the linear model, with F1 85.9% against 82.2%. Learning the global graph takes 0.5 s, against 20.9 s for PC and 7.4 s for FCI. On CartPole the ordering reverses: the decision tree reaches 94.1% against CausalXRL's 91.4%. 19
Resources: Code is public at the authors' repository. 19
Loading content card…
Takeaway: The advantage over a decision tree appears only as environments get more complex, and on the simplest benchmark the tree wins. The expert study is exploratory and qualitative, so the claim it supports is that domain players found the explanations useful, not that they improved on a measured baseline. 19
3D reconstruction and generation
8. Learning Scene Coordinate Reconstruction from Unposed Images via Pose Graph Optimization
Area tag: Camera pose estimation and 3D reconstruction
arXiv: 2610.35501, first submitted October 26, 2026. 8
Authors / institutions: Tze Ho Elden Tse and Jizong Peng at dConstruct Robotics, and Tze Ho Elden Tse with Angela Yao at the National University of Singapore. 20
Status: Accepted by CVPR 2026 as a Highlight. 8
Problem: ACE-Zero regresses scene coordinates and camera poses from unposed images without ground truth, but it refines locally and carries no global, multi-view consistency constraint. The result is drift on large or ambiguous scenes. 20
Method: Relative pose constraints are extracted from ACE-Zero's predicted scene coordinates to build a pose graph, and each edge is weighted by an uncertainty score from epipolar and optical-flow consistency across views. Refinement then alternates with registration, so poses and coordinates improve together rather than once at the end. 20
Comparison: ACE-Zero, COLMAP and VGGT-SLAM on 7-Scenes, Mip-NeRF 360 and Tanks and Temples, scoring pose accuracy through novel-view synthesis PSNR and reporting reconstruction time alongside it. 20
Results: On Mip-NeRF 360 the average is 24.3 dB against 22.9 dB for ACE-Zero and 24.7 dB for COLMAP, in 30 minutes against ACE-Zero's 1.3 hours. On the Tanks and Temples Advanced split it is 16.8 dB against 14.5 dB for ACE-Zero and 17.3 dB for COLMAP. Removing graph construction drops the average to 12.6 dB and 13.8 dB on the two datasets, and running the optimisation once at the end instead of progressively costs 0.8 dB on Mip-NeRF 360. 20
Resources: Code is at the project page linked in the paper. 20
Takeaway: This is a refinement layer over an existing coordinate regressor, and it buys COLMAP-level accuracy at a fraction of the runtime: on both benchmarks it lands within half a decibel of COLMAP, in 30 minutes where ACE-Zero takes 1.3 hours. COLMAP still wins on the Mip-NeRF 360 average, 24.7 dB to 24.3 dB. 20
9. Generative Structure from Motion with Native 3D Diffusion
Area tag: Generative 3D and structure from motion
arXiv: 2610.35522, first submitted October 26, 2026. 9
Authors / institutions: Haobin Duan, Binbin Huang, Mochu Xiang, Yiqun Zhao, Yi Ma and Shenghua Gao at The University of Hong Kong and Transengram, with Zibo Zhao at Tencent. 21
Status: Accepted to NeurIPS 2026, as stated in the arXiv record. 9
Problem: Feed-forward reconstruction recovers geometry consistent with the input views but cannot fill regions nobody photographed. Native 3D generation fills them plausibly but is not bound to the views it was given. Concatenating the two lets the first stage's error propagate into the second. 21
Method: A multi-modal diffusion transformer denoises the shape latent and one pose latent per view under shared self-attention, so the generated object is anchored to the observed views while an object-level prior supplies the unseen parts. A UV-volume tokenisation keeps the joint sequence short, and a pose-aware stage samples per-voxel image features to refine geometry and texture. 21
Comparison: TRELLIS-M as the native-3D generation baseline and VGGT as the strongest pose baseline, on Toys4k and GSO with four uncalibrated views, scored by novel-view PSNR, median Chamfer distance, an F-score at 0.01 and pose accuracy at 5°. 21
Results: Novel-view PSNR improves by 3.2 to 4.5 dB and median Chamfer distance falls by 2.8 to 3.8× over TRELLIS-M, with the F-score at 0.01 rising from 0.56 to 0.83 on Toys4k. Input-view poses are recovered at above 96% accuracy at 5°, a fivefold rotation-error reduction over VGGT, with no known intrinsics, no external pose backbone and no post-hoc optimisation. 21
Resources: A project page is linked in the paper; the record names no public repository. 9
Takeaway: Shape and cameras come out of one joint inference, and the model needs no separate pose estimator. Both benchmarks are object-level captures of toys and household objects, so a reader with room-scale or outdoor scenes is outside what was measured. 21
Generative modeling and provenance
10. Coupled Guidance for Flow Matching
Area tag: Generative modeling and guidance
arXiv: 2610.40109, first submitted October 29, 2026. 10
Authors / institutions: Mateo Clémente, Léo Maxime Brunswic, Rui Heng Yang and Amir Rasouli, all at Huawei. 22
Status: Accepted at NeurIPS 2026, as stated in the arXiv record. 10
Problem: Guidance methods were derived for diffusion models and then carried over to flow matching. The paper shows why the carry-over is unstable: naive guidance breaks the coupling distribution the flow model was trained on, so control gets weaker as guidance strengthens. 22
Method: A guidance rule derived from the geometric structure of the flow-matching coupling rather than translated from the diffusion case, applied inside the sampling update with no extra network evaluations. The claim is proved analytically for the coupling, then tested on control and generation tasks. 22
Comparison: ΠGDM, the recent general guidance framework for flow matching, in both its plain and tuned (ΠGDM+) configurations, on a robotic collision-avoidance task in ManiSkill3, four dexterous manipulation tasks from Adroit, and inpainting, super-resolution and deblurring on ImageNet-128 and CelebA-HQ-256. 22
Results: Collision avoidance over 256 environment seeds reaches 81% success with coupled guidance, against 36% for naive guidance at the same weight and about 21% with no guidance, and it holds near full-guided performance down to roughly 5% guided dimensions where naive guidance degrades below 20%. On CelebA-HQ-256 inpainting, FID falls from 54.3 to 20.7 and PSNR rises from 20.5 to 30.1. The tuned ΠGDM+ variant already reaches FID 7.5, and coupled guidance does not improve on it. 22
Resources: The record and the paper name no public code release. 10
Takeaway: The gain is largest over an untuned baseline, which is the state a practitioner starts in when they port a diffusion guidance recipe to a flow model. Against a carefully tuned baseline the difference narrows to nothing on image inpainting, so the fix matters most where guidance has to stay strong. 22
11. PP-Mark: Provable and Publicly Verifiable Watermarking for Generative AI
Area tag: Watermarking and provenance
arXiv: 2610.41170, first submitted October 30, 2026. 11
Authors / institutions: Hankyeong Ko and Taehoon Kim, at the Graduate School of Virtual Convergence, Sogang University. 23
Status: Accepted at NeurIPS 2026, as stated in the arXiv record. 11
Problem: A watermark only works if anyone can check it, and a public detector can be studied by an attacker who then forges the imprint onto content they did not generate. Statistical detectors alone cannot separate a forged imprint from an embedded one. 23
Method: PP-Mark pairs a lightweight statistical detection score with a zero-knowledge proof that the embedding process was actually run on those pixels. Verification checks the proof and the score together, and unforgeability is argued from standard cryptographic assumptions under stated sampling conditions. 23
Comparison: Tree-Ring, Gaussian Shading, TrustMark, InvisMark, RingID, StableSig, HiDDeN and WIND on imprinted-image forgery, at thresholds calibrated to a 1% false-positive rate on clean images, on SD2.1 and SDXL with 100 cover–reference pairs. 23
Results: Under cross-method imprint transfer, forgeries pass PP-Mark's detector at 0.33% when the imprint came from Tree-Ring and 0.67% when it came from Gaussian Shading. In the Tree-Ring column the same test passes 3.67% of forgeries for TrustMark and 4.00% for RingID. The underlying Tree-Ring and Gaussian Shading detectors accept 92–97% and 100% of directly forged images. At N = 1000 samples with 125 opened positions, the miss probability for spot-check detection is 0.01297. 23
Resources: Code is public at the authors' repository. 23
Loading content card…
Takeaway: The cost is on the issuer's side: embedding takes 0.89 s and detection 0.72 s, while the zero-knowledge receipt takes a projected 17.0 s to issue and 2.81 MB to store. The guarantee also weakens as the number of opened positions falls, which the paper states rather than hides. 23
Optimization theory
12. Affine-invariant Cubic Newton with Weak Learners
Area tag: Optimization theory for boosting
arXiv: 2610.33516, first submitted October 24, 2026. 12
Authors / institutions: Nikita Zozoulenko and Thomas Cass at Imperial College London, Daniel Falkowski at Stockholm University, and Lukas Gonon at Imperial College London and the University of St. Gallen. 24
Status: Accepted to the 40th Conference on Neural Information Processing Systems (NeurIPS 2026). 12
Problem: Gradient boosting is Newton's method with weak learners, and Newton's method has no global convergence guarantee in that setting. The recent gradient-regularised Newton analysis closed that gap but paid for it with a twelfth power of a worst-case weak-gradient edge. 24
Method: The affine-invariant cubic Newton scheme is generalised to weak learners, with the analysis run on the average cosine angle along the optimisation path and on directional improvement measured in the Hessian-induced norm. The update turns out to be a damped weak-Newton step whose stepsize is set by that directional improvement, and an adaptive variant tunes the cubic regularisation without an inner loop. 24
Comparison: The weak-GRN bound, in theory; in experiments, weak-AICN against weak-GRN, vanilla Newton boosting and first-order boosting with decision-tree weak learners, including a Charbonnier loss on which vanilla Newton boosting is shown not to converge, and the HIGGS dataset. 24
Results: The global rate keeps its O(1/k²) shape but replaces the worst-case edge γ¹² with a path-averaged cosine angle, which the abstract writes as O(1/Θ̄k⁴k²). The paper also proves a local linear rate, global linear convergence for Hessian-dominated losses, and lower bounds on the cosine angle. Measured on HIGGS, the cosine angle and the weak-gradient edge take similar values, which is what makes the comparison between the two rates meaningful. 24
Resources: Code, run on a single CPU, is public at the authors' repository. 24
Loading content card…
Takeaway: The contribution is what can now be proved about Newton boosting, and the experiments check the theory's own quantities rather than compete on a benchmark. The authors name the trade themselves: working purely in the local Hessian geometry makes the method struggle with ill-conditioned problems unless static ℓ2 regularisation is added. 24
What to open first
- Serving large sparse models at low batch: the speculative-decoding analysis (2610.38981) if your target is fine-grained and you have been assuming the peak sits at a middle batch size. 1
- Cutting a hot numerical primitive: FuseLSE (2610.33647) is a drop-in swap for
logsumexpin a language-model head or a sequence-log-probability stage. 2 - Distilling from a strong teacher: the steering-vector paper (2610.40998) if the student cannot reach the teacher's reasoning style, and plain distillation has plateaued. 3
- Shipping a model whose confidence you report: CoCHI (2610.34994) if you can label a few dozen samples and want calibration without retraining. 4
- Wiring an agent to real tools: SABench (2610.34167) before you trust a success rate, and particularly if your tasks have ambiguous scopes. 5
- Training offline RL on stochastic data: BGP (2610.40130) when your trajectories branch and some goals are reached by luck. 6
- Explaining an RL policy to someone else: CausalXRL (2610.42028) for a structured explanation with a measured fidelity check, and for its structure-learning cost. 7
- Reconstructing scenes from unordered images: the pose-graph refinement (2610.35501) to speed up a COLMAP-shaped pipeline, and GenSfM (2610.35522) when the unseen side of an object has to be filled in from a handful of views. 8
- Controlling a flow-matching sampler: Coupled Guidance (2610.40109) for tasks where the guidance weight has to stay high, such as constraint satisfaction. 10
- Publishing a watermark verifier: PP-Mark (2610.41170) if the detector will be public and forging an imprint is the threat you care about. 11
- Working on Newton-type boosting: the affine-invariant analysis (2610.33516) for what can now be proved, and for the ill-conditioning caveat. 12
References
- 1Speculative decoding arXiv record
arxiv.org
- 2LSE kernels arXiv record
arxiv.org
- 3Steering vectors arXiv record
arxiv.org
- 4CoCHI arXiv record
arxiv.org
- 5SABench arXiv record
arxiv.org
- 6BGP arXiv record
arxiv.org
- 7CausalXRL arXiv record
arxiv.org
- 8PGO arXiv record
arxiv.org
- 9GenSfM arXiv record
arxiv.org
- 10Coupled Guidance arXiv record
arxiv.org
- 11PP-Mark arXiv record
arxiv.org
- 12weak-AICN arXiv record
arxiv.org
- 13Speculative decoding paper
arxiv.org
- 14LSE kernels paper
arxiv.org
- 15Steering vectors paper
arxiv.org
- 16CoCHI paper
arxiv.org
- 17SABench paper
arxiv.org
- 18BGP paper
arxiv.org
- 19CausalXRL paper
arxiv.org
- 20PGO paper
arxiv.org
- 21GenSfM paper
arxiv.org
- 22Coupled Guidance paper
arxiv.org
- 23PP-Mark paper
arxiv.org
- 24weak-AICN paper
arxiv.org
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.