Neodrop
Create
FeedDiscoverMy ChannelsAuto-publishJoin Our Discord
FeedDiscoverCreateMy Channels
Optimal transport, noisy latents, and alignment resistance: ten top-conf papers from September 7–11, 2026
2026-09-14

Optimal transport, noisy latents, and alignment resistance: ten top-conf papers from September 7–11, 2026

Ten first-submission arXiv papers from September 7–11, 2026 cover optimal-transport planning, noisy-space diffusion policies, conditioned attention initialization, spectral diffusion geometry, alignment resistance in mid-training, compositional category discovery, conformal calibration transfer, adaptive vision propagation, pruning concentration bounds, and multimodal alignment evaluation.

Top-conf Paper Weekly Digest
Top-conf Paper Weekly Digest@NeoDrop Official
The papers in this issue were first submitted to arXiv from September 6 through September 11, 2026, corresponding to the announcement batches of September 7 through September 11, 2026. I selected papers whose title or submission record carries an explicit conference affiliation for NeurIPS, ICML, ICLR, or CVPR, and excluded workshop-only entries. Confirmed conference acceptance or proceedings publication is labeled as such; records with a bare conference name or an under-review status remain marked as preprint.
PaperAreaStatusReading signal
Second-Order Smooth Planning with Optimal-Transport Bellman SmoothingReinforcement learning & planningPublished at ICML 2026Optimal-transport smoothing yields an unbiased Hessian estimator and cuts worst-case sample complexity from to . 1
Noisy-Space Policy Gradient for Diffusion Policies in Offline Reinforcement LearningOffline reinforcement learningAccepted at ICML 2026A noisy-space Q-function enables single-step actor updates without unrolling diffusion trajectories, scoring 64.9 on HalfCheetah-medium at 3.4 ms per update. 2
Conditioned Initialization for AttentionAttention architecture & visionPreprint; the record names ICLR 2026Spectral singular-value conditioning on query-key projections boosts ViT-B ImageNet accuracy from 80.3% to 80.9% without extra training parameters. 3
Revisiting Spectral Representations in Generative Diffusion ModelsGenerative diffusion modelsPreprint; the record names ICML 2026Proves that representation alignment acts as closed-form score distillation, dropping ImageNet-64 DiT-L/4 FID from 9.44 to 7.99. 4
Everything in Moderation: Per-Domain Coverage Optima and Alignment-Resistant Domain Gaps in Multi-Domain Mid-TrainingLLM mid-training & alignmentPreprint; about to commit to ICLRReasoning domains exhibit interior coverage optima (18%–28%); downstream SFT and RL fail to close multi-domain mid-training deficits. 5
CoGe-GCD: Reframing Generalized Category Discovery with Compositional GeneralizationOpen-world visual discoveryAccepted at ICML 2026Primitive competition and geometric induction raise CUB-200 clustering accuracy to 63.9% and drop class-count error to 2.0%. 6
Conformal Calibration TransferUncertainty quantificationAccepted at ICML 2026Optimal-transport quantile correction maintains 0.91 coverage under severe CIFAR-100-C image corruptions where standard calibration drops below nominal level. 7
TailProp: Content-Adaptive Light- and Heavy-Tailed Propagation for VisionVision foundation modelsPreprint; under review at ICLR 2027Fractional Laplacian dynamics dynamically interpolate between Gaussian and Cauchy diffusion, reaching 82.8% ImageNet Top-1 with 28.5M parameters. 8
Theoretical Guarantees for One-Shot Magnitude Pruning and Compute-Adaptive Early ExitModel compression & pruningPublished in ICML 2026Proves representation concentration under pruning and shows that simple scalar weight renormalization restores feature alignment without retraining. 9
UFO: Chain-of-Evaluation for Omni-Condition Alignment in Multi-Modal Image GenerationMultimodal generation evaluationAccepted at ICML 2026Decomposing generation prompts into atomic evaluation units with tool-assisted VQA reaches 0.6889 Pearson correlation with human visual alignment. 10

Reinforcement learning and control

1. Second-Order Smooth Planning with Optimal-Transport Bellman Smoothing

Area tag: Reinforcement learning and planning theory
arXiv: 2609.06484, first submitted September 6, 2026. 1
Authors / institutions: Tuan Dam is affiliated with Hanoi University of Science and Technology, Hanoi, Vietnam. 11
Status: Published at ICML 2026. The submission record confirms publication in the proceedings of the 43rd International Conference on Machine Learning. 1
Problem: Planning with a generative model seeks to estimate the optimal value of a state using minimal simulator queries. While the classical Sparse Sampling Algorithm requires non-polynomial samples in 1/ε, SmoothCruiser established problem-independent complexity by smoothing the Bellman backup via entropy regularization. However, SmoothCruiser relies exclusively on a first-order Taylor expansion, leaving an open theoretical question: can higher-order smoothing reduce the sample complexity exponent further?
Method: The author establishes a curvature-driven planning framework proving that the sample complexity exponent is governed by the order β of the local Taylor remainder, yielding an oracle complexity bound of . When (first-order), this recovers SmoothCruiser's rate. When (second-order with cubic remainder), the complexity drops to . The paper realizes this second-order regime using an optimal-transport-smoothed Bellman backup over action distributions:
This operator possesses an explicit closed form, a valid policy gradient, and a Lipschitz-continuous Hessian whose quadratic correction admits an unbiased Monte Carlo cross-product estimator. The resulting algorithm, SecondOrderSmoothCruiser, combines this second-order estimator with adaptive multi-level sampling. Furthermore, an elimination variant (OT-GapE) matches the optimal gap-dependent rate of . 11
Comparison: The paper compares analytical guarantees against Sparse Sampling (SSA, Kearns et al., 2002), SmoothCruiser (Grill et al., 2019), and multi-level Monte Carlo formulations for regularized Markov Decision Processes.
Results: In the fixed-confidence value-based setting, SecondOrderSmoothCruiser improves total worst-case oracle calls from non-polynomial (SSA) and (SmoothCruiser) to . Under favorable action-value sub-optimality gaps, OT-GapE provably matches . 11
Resources: The paper provides theoretical proofs and pseudocode for SecondOrderSmoothCruiser and OT-GapE. No public software repository is linked in the text. 11
Takeaway: Essential reading for researchers in theoretical reinforcement learning, Monte Carlo tree search, and generative-model planning. The current formulation applies to discrete action spaces with Wasserstein regularization; continuous action domains remain an open direction.

2. Noisy-Space Policy Gradient for Diffusion Policies in Offline Reinforcement Learning

Area tag: Offline reinforcement learning and continuous control
arXiv: 2609.06882, first submitted September 7, 2026. 2
Authors / institutions: Mahmoud Selim is affiliated with TRATON CV AB and KTH Royal Institute of Technology, Stockholm, Sweden. Cristina Cipriani and Karl H. Johansson are affiliated with KTH Royal Institute of Technology. 12
Status: Accepted at the 43rd International Conference on Machine Learning (ICML 2026). 2
Problem: Diffusion policies express complex, multi-modal action distributions in continuous control. However, training them via reinforcement learning has proven difficult: standard policy gradient approaches require backpropagating gradients through the full multi-step reverse diffusion chain. This unrolling causes vanishing or exploding gradients, high memory consumption, and severe training instability.
Method: The authors introduce a noisy-space action-value function, , which assigns expected cumulative returns directly to noisy intermediate latents at diffusion step k. Leveraging this formulation, they derive the Noisy-Space Policy Gradient (NSPG), which updates the score network at individual reverse steps without unfolding the denoising chain. They integrate NSPG into Noisy-Space Actor-Critic (NSAC), which pairs single-step reverse updates with pessimistic Q-learning in the clean action space, enforcing distributional conservatism against out-of-distribution actions. 12
Comparison: Evaluated on 15 continuous control tasks across D4RL (Gym Locomotion, AntMaze, Adroit Pen/Door) and OGBench visual manipulation environments. Baselines include Behavior Cloning (BC), Onestep-RL, Conservative Q-Learning (CQL), Implicit Q-Learning (IQL), In-Sample Value Iteration (IVR), XQL, Diffuser, Decision-Time Q-Learning (DTQL), AlignIQL, SfBC, Diffusion Q-learning (DQL), IDQL-A, FQL, and Diffusion Actor-Critic (DAC).
Results: On D4RL locomotion benchmarks, NSAC achieves top normalized scores: 64.9 ± 0.6 on HalfCheetah-medium-v2 (compared to 59.1 for DAC and 47.4 for IQL), 101.4 ± 3.4 on Hopper-medium-v2 (vs. 100.9 for DAC), and 111.4 ± 1.5 on HalfCheetah-medium-expert-v2. On OGBench visual tasks, NSAC achieves an 88 ± 7% success rate on visual-cube-single-play (vs. 83.7% for DAC and 70% for IQL) and 34 ± 23% on visual-cube-double-play (matching DAC at 34%, while ReBRAC achieves only 4%). Computationally, NSAC with 4 Monte Carlo samples requires 3.4 ms per update—a 0.98x wall-clock speedup relative to DAC (3.6 ms). 12
Resources: Algorithmic details and hyperparameter schedules are provided in the paper; no public code repository is listed. 12
Takeaway: Practitioners training diffusion policies for robotics or continuous control should examine this paper. NSAC bypasses backpropagation through time while maintaining state-of-the-art policy quality. Performance depends on task-specific tuning of the guidance weight parameter.

LLM training, steerability, and attention architecture

3. Conditioned Initialization for Attention

Area tag: Deep learning architecture and attention mechanisms
arXiv: 2609.07086, first submitted September 7, 2026. 3
Authors / institutions: Hemanth Saratchandran and Simon Lucey are affiliated with the Australian Institute for Machine Learning at the University of Adelaide, Australia. 13
Status: Preprint. The arXiv submission record names ICLR 2026 without an explicit acceptance decision. 3
Problem: Standard transformer weight initializations (such as uniform, normal, or Xavier) treat attention projection matrices identically to feedforward layers, ignoring the mathematical structure of the softmax attention operator. At initialization, query and key weight matrices and often exhibit ill-conditioned singular value spectra, leading to premature rank collapse in attention maps and sluggish early training dynamics.
Method: The authors present a theoretical framework characterizing the condition number of the query-key product matrix under input sequences of length N and token dimension D. They propose Conditioned Initialization, which constrains the singular values of and to lie within a mathematically derived spectral interval, preventing the condition number of the attention logits from growing unboundedly with sequence length. The method operates as a parameter-free, drop-in replacement for default initialization at epoch zero. 13
Comparison: Evaluated against default PyTorch initializations and Mimetic initialization across Vision Transformers (ViT-B, DeiT-B, Swin-B, XCiT-M, DaViT-B) trained on ImageNet-1K, fine-grained image transfer benchmarks (Oxford-IIIT Pets, Flowers-102, CIFAR-10, CIFAR-100), Mask R-CNN object detection on MS COCO 2017, Long Range Arena sequence tasks, and causal language modeling with Cramming.
Results: On ImageNet-1K, conditioned initialization consistently improves Top-1 classification accuracy across all architectures: ViT-B reaches 80.9% (vs. 80.3% default, 80.5% mimetic), DeiT-B reaches 82.1% (vs. 81.6% default), Swin-B reaches 83.8% (vs. 83.4% default), and DaViT-B reaches 84.7% (vs. 84.3% default). On small-scale datasets where overfitting is acute, ViT-T pretrained with conditioned initialization achieves 58.2% on Oxford Pets (vs. 26.7% default, 47.7% mimetic) and 78.4% on Flowers-102 (vs. 64.5% default, 71.6% mimetic). On MS COCO Mask R-CNN, bounding-box AP improves by 0.5 to 0.8 points across backbones. 13
Resources: Experiments build on public repositories including timm, cramming, and Nystromformer. No dedicated standalone project repository is provided. 13
Takeaway: This paper is valuable for teams training vision transformers or sequence architectures from scratch, especially under constrained data budgets. The initialization costs zero additional FLOPs during training or inference. Gains narrow as pretraining dataset size scales into hundreds of billions of tokens.

4. Everything in Moderation: Per-Domain Coverage Optima and Alignment-Resistant Domain Gaps in Multi-Domain Mid-Training

Area tag: LLM mid-training, steerability, and post-training alignment
arXiv: 2609.09081, first submitted September 8, 2026. 5
Authors / institutions: Yunpeng Xu and Kun Zheng. Institutional affiliations are not specified in the paper header or arXiv record. 14
Status: Preprint. The author comment notes "About to commit to ICLR". 5
Problem: In multi-domain mid-training—the intermediate training stage between general web-scale pre-training and supervised fine-tuning—data mixtures are typically determined by raw token availability rather than principled optimization. Pretraining teams often assume that downstream alignment (SFT and RL) can compensate for any domain under-representation during mid-training. The authors examine whether this assumption holds, and how token allocations across reasoning domains interact.
Method: The study conducts controlled experiments on five distinct logical reasoning domains in KOR-Bench: Ciphers, Operations, Logic, Counterfactuals, and Puzzles. Across 12 interior simplex allocations (with token shares ranging from 12% to 34% per domain), models are evaluated post-mid-training, after remedial Supervised Fine-Tuning (SFT), and after Group Sequence Policy Optimization (GSPO) reinforcement learning. The study validates findings on Qwen3-4B-Base. 14
Comparison: Compares no-mid-training baselines (Base: 40.32 ± 2.7%, SFT only: 64.08 ± 1.8%, SFT+RL: 67.52 ± 1.9%) against 12 mid-training allocations, alongside compensatory SFT regimes that artificially oversample under-represented domains during post-training.
Results: First, per-domain performance curves exhibit interior coverage optima: accuracy peaks at moderate token allocations (typically 18% to 28%) and degrades when a single domain is over-allocated, due to cross-domain interference. Second, performance deficits from mid-training prove alignment-resistant: for a model trained with 12.9% Ciphers data, post-mid-training accuracy is 52.9 ± 1.6%, and remedial SFT with oversampled cipher data raises it only to 68.4%—failing to reach the 78.1% achieved by a model given a 24.4% mid-training allocation. Downstream RL (GSPO) improves overall scores but preserves relative domain gaps intact. 14
Resources: Training configs and KOR-Bench evaluation scripts are documented in the supplementary text; no public code URL is provided. 14
Takeaway: A vital read for LLM post-training and pre-training pipeline architects. The central takeaway is actionable: post-training alignment cannot fix structural omissions in mid-training data. Data mixtures must be balanced during mid-training rather than patched downstream.

Generative diffusion models

5. Revisiting Spectral Representations in Generative Diffusion Models

Area tag: Generative diffusion models and representation learning
arXiv: 2609.08253, first submitted September 8, 2026. 4
Authors / institutions: Yuehao Wang, Peihao Wang, Hanwen Jiang, Ziyi Yang, Qixing Huang, and Zhangyang Wang are affiliated with the University of Texas at Austin, USA. 15
Status: Preprint. The arXiv comment lists ICML 2026 without an explicit acceptance statement. 4
Problem: Aligning intermediate representations of diffusion networks with self-supervised feature spaces (such as DINO or REPA) accelerates training and enhances sample fidelity. However, this phenomenon has been treated purely empirically: the mathematical mechanism linking continuous diffusion perturbation kernels to latent spectral representation learning has remained unexplained.
Method: The authors establish a formal duality connecting diffusion perturbation kernels with spectral manifold learning. They prove that representation alignment acts as closed-form diffusion score distillation in the eigenspace of the forward diffusion operator. Based on this spectral geometry, they introduce Spectral Representation Alignment (SRA), which regularizes intermediate diffusion states using the spectral decomposition of the noise kernel, eliminating reliance on heuristic feature distillation. 15
Comparison: Evaluated on ImageNet generation at 64×64 pixel resolution (DiT-L/4) and 256×256 latent resolution (DiT-XL/2), 3D point cloud generation on ShapeNet (Chair, Airplane, Car), and non-Euclidean Riemannian flow matching for protein and RNA conformational backbones. Baselines include default DiT architectures and REPA.
Results: On ImageNet-64 (DiT-L/4), SRA reduces FID from 9.441 (baseline) to 7.994, with sFID improving from 7.653 to 7.372. On ImageNet-256 (DiT-XL/2), SRA achieves 1.682 FID (compared to 2.508 for baseline DiT and 1.745 for REPA). In 3D point cloud generation on ShapeNet Chair at 5K training iterations, 1-NNA Chamfer Distance drops from 0.850 to 0.583, while coverage (COV) increases from 0.295 to 0.488. On 7D RNA manifold generation, test negative log-likelihood reaches -4.852 ± 0.180 (vs. -4.780 ± 0.196 for Riemannian Flow Matching). 15
Resources: Code is publicly released at spectral-reg-diffusion. 4
Loading content card…
Takeaway: Diffusion researchers and generative model engineers should read this paper. It provides both a rigorous mathematical justification for representation alignment and an open-source training module that improves DiT sample fidelity across 2D images, 3D point clouds, and biological manifolds.

Computer vision and representation learning

6. CoGe-GCD: Reframing Generalized Category Discovery with Compositional Generalization

Area tag: Computer vision, open-world discovery, and clustering
arXiv: 2609.10158, first submitted September 9, 2026. 6
Authors / institutions: Luyao Tang and Cheng Chen are affiliated with The University of Hong Kong. Jiewei Zheng, Kunze Huang, Chaoqi Chen, and Yue Huang are affiliated with Xiamen University, China. 16
Status: Accepted at ICML 2026. 6
Problem: Generalized Category Discovery (GCD) clusters an unlabelled image dataset that contains a mixture of known classes (present in a labelled set) and novel classes. Conventional GCD approaches rely on global Euclidean or cosine similarity in feature space. When applied to fine-grained categories, these methods fail because novel categories frequently share subtle visual primitives (e.g., wing shapes, beak patterns) with known categories, causing models to severely overestimate the number of distinct classes or fragment clusters.
Method: CoGe-GCD introduces a Latent Compositional Manifold framework. It divides discovery into two steps:
  1. Compositional Perception: Employs primitive competition and evidence consolidation to decompose visual features into sparse, reusable semantic primitives.
  2. Generalizing Induction: Uses a geometric structure-preserving calibration operator to ensure novel category assignments maintain the manifold relationships established across known classes. The method integrates into existing GCD pipelines without changing underlying vision backbones. 16
Comparison: Benchmarked across fine-grained datasets (CUB-200, FGVC-Aircraft, Stanford-Cars) and generic benchmarks (CIFAR-100, ImageNet-100). Compared against standard GCD, GPC, DCCL, SimGCD, PromptGCD, and CMS.
Results: On fine-grained classification, CoGe-GCD establishes new state-of-the-art results: on CUB-200, overall clustering accuracy reaches 63.9% (Known: 65.5%, Novel: 63.1%), significantly exceeding GCD (51.3% All, 48.7% Novel) and GPC (52.0% All, 47.5% Novel). On FGVC-Aircraft, accuracy reaches 57.9% All / 55.4% Novel (vs. 45.0% / 46.9% for GCD). On Stanford-Cars, accuracy reaches 62.9% All / 55.4% Novel (vs. 39.0% / 29.9% for GCD). Crucially, CoGe-GCD prevents category fragmentation: on CUB-200 (ground truth 200 classes), it predicts 204 categories (2.0% error), compared to 231 (15.5% error) for GCD and 304 (52.0% error) for DCCL. 16
Resources: Official implementation is available at CoGe-GCD. 16
Loading content card…
Takeaway: Open this paper if you work on open-world visual categorization, unlabelled data discovery, or fine-grained representation learning. The primary insight—reusing compositional primitives rather than clustering monolithic embeddings—solves the primary error mode of existing GCD systems.

7. TailProp: Content-Adaptive Light- and Heavy-Tailed Propagation for Vision

Area tag: Vision foundation models and continuous dynamical architectures
arXiv: 2609.11081, first submitted September 10, 2026. 8
Authors / institutions: Jiahao Kong and Zihan Li are affiliated with Shandong University, China. 17
Status: Preprint. The record indicates the submission is under review at ICLR 2027. 8
Problem: Science-inspired vision backbones replace quadratic multi-head self-attention with continuous partial differential equation (PDE) dynamics or diffusion propagation. However, current architectures fix the propagation dynamics to a single mathematical distribution (e.g., standard Gaussian diffusion). Gaussian propagation is light-tailed, making it effective for localized texture smoothing but incapable of capturing long-range semantic dependencies without deep recurrence. Conversely, purely heavy-tailed propagation (e.g., Cauchy jumps) blurs fine local edges.
Method: The authors introduce TailProp, governed by fractional Laplacian propagation dynamics:
TailProp incorporates a Content-Adaptive Tail Dynamics module that dynamically predicts the fractional exponent α per token or feature region. When , the operator behaves as light-tailed Gaussian smoothing, preserving sharp object boundaries; when , it acts as heavy-tailed Cauchy jump diffusion, bridging distant visual patches. The operator is computed in O(N) spectral complexity using separable Discrete Cosine Transforms (DCT) under Neumann boundary conditions. 17
Comparison: Evaluated on ImageNet-1K classification, Mask R-CNN object detection and instance segmentation on MS COCO 2017, and UPerNet semantic segmentation on ADE20K. Baselines include ConvNeXt, Swin Transformer, Vision Mamba (Vim), and existing PDE-based vision networks.
Results: On ImageNet-1K, TailProp-T achieves 82.8% Top-1 accuracy with 28.5M parameters and 4.6G FLOPs, outperforming Swin-T (81.3%, 28.3M params, 4.5G FLOPs) and ConvNeXt-T (82.1%, 28.6M params, 4.5G FLOPs). TailProp-S reaches 83.9% Top-1 accuracy with 50.1M parameters. On MS COCO object detection (Mask R-CNN 1× schedule), TailProp-T achieves 43.8 box AP and 40.1 mask AP (vs. 42.2 / 39.1 for Swin-T). On ADE20K semantic segmentation, TailProp-T reaches 45.3 mIoU (vs. 44.5 for Swin-T). Throughput on an NVIDIA A100 GPU is 1,024 images per second for TailProp-T. 17
Resources: Detailed mathematical derivations and DCT implementations are provided in the appendix; no public code repository is linked in the text. 17
Takeaway: This paper will interest researchers exploring linear-complexity visual backbones and physics-informed neural architectures. The content-adaptive tail index provides a principled mechanism for balancing local edge preservation and global contextual reasoning.

Uncertainty, efficiency, and multimodal evaluation

8. Conformal Calibration Transfer

Area tag: Uncertainty quantification and distribution shift
arXiv: 2609.10737, first submitted September 9, 2026. 7
Authors / institutions: Achref Doula is affiliated with Technical University of Darmstadt, Germany, and Tongji University, Shanghai, China. 18
Status: Accepted at the 43rd International Conference on Machine Learning (ICML 2026). 7
Problem: Conformal prediction provides mathematically guaranteed prediction sets with finite-sample coverage, but its validity relies strictly on the exchangeability of calibration and test data. In real-world deployment, labeled calibration data is frequently available only in a source domain, while the operational environment exhibits severe covariate shift or corruption where labels cannot be obtained.
Method: The author formalizes the problem of Conformal Calibration Transfer (CCT) and proves a fundamental impossibility result: exact finite-sample coverage in the target space is impossible without assumptions on label conditional distributions. The paper then develops Transported Conformal Calibration (TCC), mapping source non-conformity scores into the target covariate distribution via optimal transport. To address imperfect transport maps, the author introduces two robust calibration rules:
  1. TCC-KS: Uses a Kolmogorov-Smirnov one-sided supremum bound to inflate prediction thresholds, guaranteeing conservative coverage under bounded transport error.
  2. Weighted-TCC: Applies importance weights derived from the transport density ratio to maintain tight set sizes under moderate shift. 18
Comparison: Evaluated on CIFAR-100-C and Tiny-ImageNet-C across 20 synthetic corruptions and five corruption severity levels. Baselines include Oracle Conformal Prediction (with target labels), standard uncorrected Transported CP, and Weighted Conformal Prediction (WCP).
Results: At a target miscoverage level of (90% nominal coverage), on CIFAR-100-C, weighted-TCC achieves 0.89 empirical coverage with an average prediction set size of 37.15, matching the Oracle CP baseline (0.89 coverage, 37.63 set size). Under a severe stress test (shot noise and motion blur at severities 4–5), uncorrected transported CP coverage drops to 0.88, while TCC-KS maintains 0.91 coverage (set size 47.66). Across transport training epochs, as Wasserstein transport quality improves (effective sample size rising from 37.4% to 79.5%), weighted-TCC set sizes contract from 48.42 to 39.33 while strictly satisfying coverage. 18
Resources: Complete proofs and concentration inequalities are included in the paper's appendix. Standalone code has not been released publicly. 18
Takeaway: Recommended for practitioners deploying safety-critical classifiers under distribution shift. The paper clarifies when distribution-free coverage can be salvaged via optimal transport, and when conservative threshold inflation is mathematically necessary.

9. Theoretical Guarantees for One-Shot Magnitude Pruning and Compute-Adaptive Early Exit

Area tag: Model compression, neural network pruning, and early exit
arXiv: 2609.12337, first submitted September 11, 2026. 9
Authors / institutions: Erdem Koyuncu is affiliated with the Department of Electrical and Computer Engineering at the University of Illinois Chicago, USA. 19
Status: Published in ICML 2026. The submission record notes publication in the proceedings of the 43rd International Conference on Machine Learning. 9
Problem: Magnitude pruning and adaptive early exit are widely used to reduce neural network computation during inference. However, their practical adoption relies on empirical trial-and-error: theoretical guarantees quantifying exact output fidelity, cosine similarity, and signal-to-noise ratio degradation across deep layers have been missing.
Method: The author develops an asymptotic mathematical analysis unifying static compute reduction (one-shot unstructured and semi-structured magnitude pruning) and dynamic compute reduction (early exit). In a single-neuron model with independent Gaussian weights, the paper proves concentration theorems establishing that the dot product between original and pruned outputs converges almost surely to an analytical constant . The author proves that scalar weight renormalization—multiplying remaining weights by —restores expected activation scale and significantly improves cosine similarity. The analysis extends to multi-layer deep sign and ReLU networks via recursive covariance tracking. 19
Comparison: Theoretical predictions are compared against finite-width random networks, multi-layer perceptrons (L=5, width 512), and trained Vision Transformers (ViT on CIFAR-100). The evaluation tests structured patterns (1:2, 2:4, 4:8, 8:16) at a constant 50% keep ratio.
Results: Theoretical constants match empirical simulations within 0.002 across all block sizes: for 1:2 pruning , for 2:4 , for 4:8 , and for 8:16 . Renormalization improves correlation from to , raising empirical cosine similarity on width-512 networks from 0.637 to 0.798 for 2:4 pruning. On trained ViTs, the predicted class-wise coupling correlations align closely with empirical features (–0.94 across intermediate layers). For adaptive early exit, the formulation derives closed-form compute-accuracy Pareto frontiers. 19
Resources: Detailed mathematical proofs and simulation scripts are provided in the paper's extensive appendix. 19
Takeaway: Hardware acceleration engineers and model compression teams should study this paper. It provides exact, closed-form equations for post-pruning weight rescaling that can be applied instantaneously without fine-tuning data, alongside rigorous bounds on block size selection.

10. UFO: Chain-of-Evaluation for Omni-Condition Alignment in Multi-Modal Image Generation

Area tag: Multimodal evaluation, text-to-image customization, and alignment
arXiv: 2609.12397, first submitted September 11, 2026. 10
Authors / institutions: Danning Zhang, Yijing Lin, Shuhan Zhuang, Mengqi Huang, Shaojin Wu, Shancheng Fang, and Zhendong Mao are affiliated with the University of Science and Technology of China (USTC), Hefei, China. 20
Status: Accepted at the 43rd International Conference on Machine Learning (ICML 2026). 10
Problem: Customized text-to-image generation requires an image to satisfy multiple heterogeneous conditions simultaneously: reference subject identity, textual scene prompts, spatial arrangements, and stylistic constraints. Current automated evaluation metrics either compute global embedding similarities (such as CLIP-I or DINO) or query generic Multimodal Large Language Models (MLLMs). These metrics fail to isolate individual condition failures and exhibit low correlation with human perceptual evaluations.
Method: The authors introduce UFO (Unified omni-Condition Alignment), an evaluation framework that implements a Chain-of-Evaluation:
  1. AEU Decomposition: An LLM decomposes user prompts and conditions into fine-grained Atomic Evaluation Units (AEUs).
  2. Modality-Relevance Classification: Assigns each AEU to its relevant input modality (e.g., text prompt, reference image, or layout bounding box).
  3. Specialized Tool Scoring: Evaluates each AEU using specialized visual tools—such as ArcFace for human identity similarity, visual grounding detectors for spatial relationships, and targeted Visual Question Answering (VQA) queries for semantic attributes.
  4. Weighted Aggregation: Combines granular scores using weights aligned with human perceptual importance. The paper also introduces UFO-Bench, spanning non-editing, single-dimension editing, and multi-dimension editing tasks. 20
Comparison: Evaluates 12 text-to-image and customization models, including PhotoMaker, InstantID, IP-Adapter, Flux, and proprietary commercial platforms. Automated scores are compared against CLIP-I, CLIP-T, DINO, ImageReward, and standard MLLM prompting, validated across 1,200 human expert evaluations.
Results: UFO achieves a Pearson correlation of 0.6889 with human rankings, substantially exceeding CLIP-I (0.2140), DINO (0.2482), ImageReward (0.3315), and baseline MLLM prompting (0.4120). Across task sub-categories, UFO correlation reaches 0.7012 on non-editing customization and 0.6745 on multi-dimensional editing. In ablation studies, removing AEU decomposition lowers correlation from 0.6889 to 0.5752, while removing weighted aggregation drops correlation to 0.6253. 20
Resources: Evaluation code and UFO-Bench dataset assets are described in the paper; references standard model baselines including Flux. 20
Takeaway: Generative AI researchers and benchmark creators should open this paper. UFO demonstrates that modular, tool-assisted decomposition is required to reliably evaluate multi-condition image generation, replacing unreliable holistic CLIP or MLLM scores.

Reading prioritization guide

Readers can prioritize this week's papers based on their primary technical bottlenecks:
  • For reinforcement learning and control engineers:
    • If you are building diffusion actors for continuous control or robotic manipulation, open Paper 2 (NSAC) first for a scalable policy gradient that eliminates unrolling diffusion trajectories.
    • If you work on the theoretical limits of generative-model planning or tree search, read Paper 1 (SecondOrderSmoothCruiser) for the optimal-transport smoothing proof and sample complexity bounds.
  • For LLM pre-training and alignment teams:
    • If you are designing post-training or mid-training data recipes, read Paper 4 (Multi-Domain Mid-Training) to understand why mid-training domain imbalances resist downstream SFT and RL correction.
    • If you are initializing transformer models from scratch, test Paper 3 (Conditioned Initialization) for a zero-cost spectral initialization that accelerates early training.
  • For generative modeling and vision researchers:
    • If you work with Diffusion Transformers (DiTs), inspect Paper 5 (Spectral Representations) for theoretical score distillation proofs and open-source training modules.
    • If you study open-world categorization or fine-grained discovery, read Paper 6 (CoGe-GCD) for its compositional primitive formulation and released code.
    • If you explore non-attention linear vision backbones, review Paper 7 (TailProp) for fractional Laplacian propagation dynamics.
  • For deployment reliability, safety, and evaluation engineers:
    • If you require distribution-free uncertainty intervals under domain shift, read Paper 8 (Conformal Calibration Transfer) for transport-based coverage corrections.
    • If you deploy sparse models on hardware, use Paper 9 (Pruning Guarantees) for closed-form weight rescaling formulas that avoid retraining.
    • If you evaluate customized text-to-image or multimodal generation pipelines, adopt Paper 10 (UFO) for its atomic, tool-assisted evaluation chain.

References

  1. 1
    Second-Order Smooth Planning arXiv record

    arxiv.org

  2. 2
    Noisy-Space Policy Gradient arXiv record

    arxiv.org

  3. 3
    Conditioned Initialization arXiv record

    arxiv.org

  4. 4
    Spectral Representations in Diffusion arXiv record

    arxiv.org

  5. 5
    Multi-Domain Mid-Training arXiv record

    arxiv.org

  6. 6
    CoGe-GCD arXiv record

    arxiv.org

  7. 7
    Conformal Calibration Transfer arXiv record

    arxiv.org

  8. 8
    TailProp arXiv record

    arxiv.org

  9. 9
    Pruning and Early Exit Guarantees arXiv record

    arxiv.org

  10. 10
    UFO Alignment Evaluation arXiv record

    arxiv.org

  11. 11
    Second-Order Smooth Planning paper

    arxiv.org

  12. 12
    Noisy-Space Policy Gradient paper

    arxiv.org

  13. 13
    Conditioned Initialization paper

    arxiv.org

  14. 14
    Multi-Domain Mid-Training paper

    arxiv.org

  15. 15
    Spectral Representations in Diffusion paper

    arxiv.org

  16. 16
    CoGe-GCD paper

    arxiv.org

  17. 17
    TailProp paper

    arxiv.org

  18. 18
    Conformal Calibration Transfer paper

    arxiv.org

  19. 19
    Pruning and Early Exit Guarantees paper

    arxiv.org

  20. 20
    UFO Alignment Evaluation paper

    arxiv.org

Top-conf Paper Weekly Digest
Top-conf Paper Weekly Digest@NeoDrop Official

This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.

Related content

  • Reinforcement learning and control
  • 1. Second-Order Smooth Planning with Optimal-Transport Bellman Smoothing
  • 2. Noisy-Space Policy Gradient for Diffusion Policies in Offline Reinforcement Learning
  • LLM training, steerability, and attention architecture
  • 3. Conditioned Initialization for Attention
  • 4. Everything in Moderation: Per-Domain Coverage Optima and Alignment-Resistant Domain Gaps in Multi-Domain Mid-Training
  • Generative diffusion models
  • 5. Revisiting Spectral Representations in Generative Diffusion Models
  • Computer vision and representation learning
  • 6. CoGe-GCD: Reframing Generalized Category Discovery with Compositional Generalization
  • 7. TailProp: Content-Adaptive Light- and Heavy-Tailed Propagation for Vision
  • Uncertainty, efficiency, and multimodal evaluation
  • 8. Conformal Calibration Transfer
  • 9. Theoretical Guarantees for One-Shot Magnitude Pruning and Compute-Adaptive Early Exit
  • 10. UFO: Chain-of-Evaluation for Omni-Condition Alignment in Multi-Modal Image Generation
  • Reading prioritization guide
O(ε−4)
O(ε−3)
N:M
O(ε−4)
O(ε−(2+2/(β−1)))
β=2
O(ε−4)
β=3
O(ε−3)
FsOT​(Q)=π∈Δ(A)max​{⟨π,Q⟩−τWλ​(π,μs​)}
ε−2
O(ε−4)
O(ε−3)
O(ε−2)
Qπ(s,x(k),k)
x(k)
WQ​
WK​
WQ​WKT​
WQ​
WK​
github.com · GitHub repository
yuehaowang/spectral-reg-diffusion
https://github.com/yuehaowang/spectral-reg-diffusion
github.com · GitHub repository
lytang63/CoGe-GCD
https://github.com/lytang63/CoGe-GCD
∂t∂u(x,t)​=−κ(−Δ)α/2u(x,t),0<α≤2
α→2
α→1
α=0.10
N:M
τN:M​
1/τN:M​​
τN:M​
τ=0.570
τ=0.637
τ=0.702
τ=0.760
τN:M​
τN:M​​
ρ=0.81
N:M
N:M