Top-conf paper digest — week of August 10–14, 2026

What is in this batch

LLM and model systems

How Can Rhetoric Reward-Hack AI Reviewers?

• Area tag: LLM evaluation and scientific peer review.
• arXiv: 2608.08975v1, submitted August 10, 2026. 1
• Authors / institutions: Ming Li, Chenguang Wang, Xirui Li, Xinyue Zeng, Dianqi Li, Peng Shi, Dawei Zhou, and Tianyi Zhou; the HTML version lists University of Maryland, Virginia Tech, MBZUAI, and University of Waterloo. 2
• Status: Preprint. The paper studies 120 anonymized ICLR 2026 submissions; the arXiv record does not state that this preprint itself was accepted to ICLR. 1
• Problem: An AI reviewer may change its assessment when the underlying scientific content stays fixed but the rhetorical presentation changes. Natural comparisons between unrelated papers cannot isolate that effect because stronger papers are often also better written.
• Method: The authors create 4,200 full-paper manuscripts: 120 originals and 4,080 rewrites. Two LLM rewriters alter six rhetorical dimensions in opposing directions, and five LLM reviewers score the variants under standard and strict protocols. The study also tests joint, recursive, and reviewer-guided rewriting. 2
• Comparison: Relative to prior observational analyses of LLM-as-a-judge behavior, this is a matched intervention design. It changes claim stance, scope, evidence framing, contribution structure, technical register, and lexical or syntactic complexity while preserving the manuscript's scientific content and structural anchors.
• Results: Evidence framing and novelty stance produce the largest score contrasts, with scope framing in a weaker second tier. Positive evidence framing raises Overall Assessment by up to 0.93, negative novelty stance lowers it by up to 0.73, and evidence framing changes weak-accept probability by 13 percentage points on average. Strict review lowers mean Overall Assessment by 1.36 points. More elaborate rewriting does not reliably produce larger gains. 2
• Resources: The project repository is Dissecting_AI_Reviews. 3

• Takeaway: This is a useful stress test for evaluation pipelines, not evidence that rhetorical clarity is illegitimate. The actionable question is whether a reviewer can separate legitimate exposition quality from content-preserving changes in evidence and novelty framing.

Hyperbolic Multimodal Continual Learning

• Area tag: Multimodal representation learning and continual learning.
• arXiv: 2608.09572v1, submitted August 10, 2026. 4
• Authors / institutions: Jiahong Liu, Ming Shen, Xiaohao Liu, Rex Ying, Menglin Yang, Tat-Seng Chua, and Irwin King; the paper front matter lists the Chinese University of Hong Kong, National University of Singapore, Yale University, and the Hong Kong University of Science and Technology (Guangzhou). 5
• Status: Published in ICML 2026 proceedings, PMLR 306. The arXiv comment itself says ICML 2026, while the paper front matter identifies the ICML proceedings volume. 4
• Problem: In a continual multimodal system, updates can preserve pointwise task accuracy while distorting the geometry that encodes cross-modal similarity and semantic hierarchy. Euclidean continual-learning constraints do not directly preserve hyperbolic invariants.
• Method: The paper derives a geometric condition for forgetting prevention: updates should preserve cross-modal relations under a shared hyperbolic isometry. HMCL projects parameter updates onto the canonical subspace compatible with the Lorentz geometry, and combines that update rule with hyperbolic multimodal contrastive learning. The evaluation uses up to 15 sequential classification and retrieval datasets, including CIFAR-10/100, Caltech-101, Food-101, COCO, and Flickr30k. 5
• Comparison: The baselines include vanilla continual training, EWC, GEM, and C-FLAT. The comparison is not simply an accuracy contest: it measures overall performance, classification and retrieval backward transfer, and four types of geometric drift.
• Results: On MERU-L, HMCL raises the overall metric from 38.22 to 41.36, a relative gain of 8.2% over vanilla training. Classification backward transfer improves from -6.46 to -0.74, an 88.5% relative reduction in forgetting. The paper reports that HMCL also reduces radial, angular, cross-modal, and paired-distance drift across most datasets. 5
• Resources: Code is linked from the arXiv record at HMCL_ICML. 6

• Takeaway: The paper is worth opening when a multimodal model's failure mode looks like relation drift rather than ordinary task forgetting. Its strongest claim is geometric: preserving the right invariants can be more informative than preserving weights or logits alone.

Diffract: Spectral View of LLM Domain Adaptation

• Area tag: LLM continual pretraining and model analysis.
• arXiv: 2608.10850v1, submitted August 11, 2026. 7
• Authors / institutions: Nikita Borodin, Maria Krylova, Artem Zabolotnyi, Dmitry Aspisov, Egor Shikov, Nikita Tyuplyaev, Oleg Travkin, Roman Alferov, and Dmitry Vinichenko; affiliations are not exposed in the arXiv HTML metadata.
• Status: Accepted at ICML 2026. 7
• Problem: Continual pretraining adapts an LLM to math, code, instruction, or natural text, but it is unclear which weight changes carry the domain-specific capability and which updates are redundant.
• Method: Diffract analyzes weight matrices with singular value decomposition across 1B, 7B, and 13B OLMo 2 models. The key finding is that CPT leaves singular-value spectra largely stable while changing singular vectors, especially in domain-dependent attention heads. The authors rank heads by their contribution to CPT quality, then rewind low-importance heads to their pretrained state. They also linearly interpolate checkpoints trained on different domains to test what they call domain connectivity. 8
• Comparison: Instead of treating CPT as a dense delta or applying generic pruning after training, the method compares the full CPT checkpoint with selectively rewound heads and sparse CPT deltas. The checkpoint interpolation experiment also tests whether model merging behaves like a smooth path between domain-specialized models.
• Results: Up to 60% of attention-head updates can be removed without measurable quality loss. Selectively rewinding low-importance heads improves benchmark accuracy by up to 4% over the fully trained baseline. Linear interpolation between domain checkpoints gives smooth domain-quality interpolation without notable degradation on either endpoint, although the paper notes that a mixed CPT checkpoint is worse than training on a data mixture. 8
• Resources: The open-source toolkit is Risk-AI-Research/diffract. 9

• Takeaway: The practical signal is that a CPT run may contain a smaller, diagnosable update than its parameter count suggests. The limitation is equally important: the evidence is concentrated in the OLMo 2 family and should not yet be read as a universal rule for modern LLMs.

LoKiFormer: Locality-aware Attention with Decoupled Knowledge Memory

• Area tag: Efficient LLM architecture.
• arXiv: 2608.12419v1, submitted August 12, 2026. 10
• Authors / institutions: Qiuwu Chen, Zimo Liu, Yuchen Li, Ying Sun, Yifan Zhang, Ryan Dong, and Simeng Ma at AIGCode; Zhijie Qiu, Zeng You, Yaofo Chen, and Mingkui Tan at South China University of Technology; Yaofo Chen is also affiliated with Pazhou Laboratory. 11
• Status: Accepted by ICML 2026. 10
• Problem: Standard self-attention spends capacity modeling local token interactions that could use an explicit locality bias. In MoE-style systems, global knowledge storage is also coupled to the computational pathway, which makes direct knowledge access and modification less flexible.
• Method: Local Fusion Attention (LFA) applies a convolutional fusion before attention, so attention receives representations that already encode local patterns. The Knowledge Memory Module (KMM) adds trainable key-value slots for sequence-external knowledge and decouples storage from the expert computation path. The two modules extend a standard decoder rather than replacing it with a retrieval system.
• Comparison: The comparison is against baseline decoder architectures and the implicit knowledge storage of MoE. LFA does not claim to reduce full sequence-level attention complexity; its stated benefit is a locality inductive bias. KMM differs from external retrieval because its slots are learned end to end and remain static at inference.
• Results: LoKiFormer reaches the same validation loss in 7.5K pretraining steps instead of 10K for the baseline, corresponding to 1.33x faster convergence. The arXiv HTML does not provide a single headline downstream accuracy table in the accessible summary, so the quantitative signal here is convergence efficiency. 11
• Resources: No public code link is listed in the arXiv abstract or HTML record.
• Takeaway: This is most relevant for researchers who want local structure and explicit parametric memory in the same decoder. The open question is how the added memory and convolutional path change inference cost and transfer across model scales.

What preferences can - and cannot - predict in multi-agent online learning

• Area tag: Multi-agent learning theory and game dynamics.
• arXiv: 2608.13810v1, submitted August 13, 2026. 12
• Authors / institutions: Omar Abbadi and Rida Laraki at Mohammed VI Polytechnic University / Moroccan Center for Game Theory; Panayotis Mertikopoulos at Universite Grenoble Alpes, CNRS, Inria, Grenoble INP, and LIG. 13
• Status: ICML 2026 oral presentation. 12
• Problem: A preference graph records the directions of profitable unilateral deviations, but it discards the cardinal payoff values. The paper asks when that ordinal information is enough to predict the long-run behavior of no-regret dynamics such as FTRL.
• Method: The authors characterize dynamically stable sets and preferentially stable sets, prove an inclusion relation, and then construct a three-player counterexample. They introduce resilience under aggregate deviations as a payoff-sensitive condition that restores asymptotic stability for arbitrary spans of pure strategies.
• Comparison: The main comparison is conceptual rather than empirical. Preference-only reasoning can characterize stability for subgames, but the equivalence breaks for general spans; two payoff realizations with the same preference ordering can therefore have different dynamic behavior.
• Results: The key evidence is a theorem and a counterexample, not a benchmark score. In the constructed three-player game, a preferentially stable set has a dynamically unstable span. The paper therefore shows a boundary on what preference graphs can certify, then supplies a checkable payoff-based condition for the missing cases. 13
• Resources: No code or dataset link is listed in the arXiv record.
• Takeaway: Open this if your multi-agent design starts from preferences or ordinal rewards. The practical warning is that an ordering of actions is not always enough to predict learning dynamics; the numerical reward realization can remain part of the system specification.

Vision, sensing, and adaptation

Gaze Target Estimation Anywhere with Concepts

• Area tag: Promptable gaze estimation and vision-language perception.
• arXiv: 2608.11367v1, submitted August 11, 2026. 14
• Authors / institutions: Xu Cao, Houze Yang, Vipin Gunda, Zhongyi Zhou, Tianyu Xu, Adarsh Kowdle, Inki Kim, and James M. Rehg; the HTML version lists University of Illinois Urbana-Champaign and Google. 15
• Status: Preprint. The arXiv comment says CVPR 2026 and points to code and a benchmark, but it does not state accepted, published, or Findings status. 14
• Problem: Traditional gaze-target pipelines require an explicit head box, pose estimate, or other intermediate cue. A detector error can therefore make the downstream gaze estimate fail, and the user cannot specify the person naturally at inference time.
• Method: The paper defines Promptable Gaze Target Estimation (PGE), in which a text description such as "the boy in the red shirt" or a visual coordinate identifies the subject. GazeAnywhere fuses frozen visual and text encoders with a trainable transformer detector, then jointly predicts the subject head box, in/out-of-frame status, and gaze heatmap. The accompanying Gaze-Co data engine converts GazeFollow, VAT, and ChildPlay into 119,525 prompt-annotated training samples. 15
• Comparison: The paper compares one end-to-end model with two-stage combinations of open-vocabulary detectors and gaze estimators. It also compares against zero-shot VLM prompting and tests text prompts against coordinate prompts. The central comparison is pipeline dependency: a two-stage baseline must first localize the correct person, while GazeAnywhere learns localization and gaze estimation together.
• Results: The paper reports state-of-the-art performance across the three public PGE benchmarks and an out-of-domain Child-SC set containing 326 clips from 40 children. The accessible HTML does not expose the main numeric table because it contains an extraction placeholder, so the headline metric values are not reported here. One concrete diagnostic is available: OWLv2 reaches only 70% child head/face detection accuracy on Child-SC, which illustrates the failure mode the end-to-end design targets. 15
• Resources: Code is available at GazeAnywhere, and the benchmark is linked at Gaze-Co-Benchmark. 14

• Takeaway: This is a strong candidate for readers building interactive or crowded-scene perception systems. The model's advantage is architectural and operational: it removes the need to hand off a precise subject box, while the missing extracted table means the original should be opened before making a performance decision.

Dual Anchors, Do It Better: Hierarchical Group Merging for Zero-Shot Anomaly Detection

• Area tag: Vision-language anomaly detection.
• arXiv: 2608.11933v1, submitted August 12, 2026. 16
• Authors / institutions: Jimin Roh, DongKyu Kim, and Suk-Ju Kang; affiliations listed in the HTML are Sogang University, LG Electronics, and NAVER Cloud. 17
• Status: Accepted to CVPR 2026 Findings. 16
• Problem: CLIP-based zero-shot anomaly detection commonly anchors normal and abnormal semantics on text prompts. That makes the decision boundary sensitive to prompt wording and leaves visual structure underused, especially when the target domain is industrial or medical and unseen during training.
• Method: Hierarchical Group Merging progressively aggregates DINOv3 patch tokens into image anchors through cross-attention, differentiable group assignment, and token merging. A Group-Gated Token Refiner uses normal and abnormal group tokens to refine the global image representation. Dynamic State Prompts then fuse those image anchors with text anchors.
• Comparison: The paper compares against WinCLIP, APRIL-GAN, AnomalyCLIP, AdaCLIP, AA-CLIP, and Bayes-PFL across eight industrial and six medical datasets. The contribution is not another prompt template: it moves the semantic anchor from a text-only design to a dual visual-text design and tests the result across object, texture, and medical domains.
• Results: Across the image-level benchmark average, the method reports 92.9 AUROC, 91.1 F1, and 94.6 AP, compared with Bayes-PFL's 91.2, 90.3, and 92.9. On MVTec-AD it reports 92.7 AUROC, 94.0 F1, and 97.1 AP; on VisA it reports 88.3, 85.0, and 91.8. Removing hierarchical merging, the group-gated refiner, or dynamic state prompts degrades both image- and pixel-level performance. 17
• Resources: No public project repository is linked from the arXiv abstract or HTML record.
• Takeaway: The paper is worth reading when prompt engineering is the unstable part of a zero-shot inspection system. The key implementation question is whether a DINOv3-centered visual anchor can compensate for the weaker text alignment of a pure vision backbone.

Curvature-Aware Zeroth-Order Optimization for Memory-Efficient Test-Time Adaptation

• Area tag: Test-time adaptation and memory-efficient optimization.
• arXiv: 2608.12279v1, submitted August 12, 2026. 18
• Authors / institutions: Junming Zhang, Shuyu Yin, Peilin Liu, Rendong Ying, and Fei Wen at Shanghai Jiao Tong University. 19
• Status: Published in CVPR 2026. The arXiv record includes the CVPR proceedings reference, pages 836-846. 18
• Problem: Backpropagation-based test-time adaptation is costly in memory, while zeroth-order methods use forward-only evaluations but suffer from high-variance gradient estimates.
• Method: CAZO first observes that the adaptation loss Hessian has a persistent, low-rank, slowly varying structure. It uses an exponential moving average of diagonal curvature to sample perturbations anisotropically, freezes the pretrained weights, and updates only a lightweight adapter through forward-only finite differences.
• Comparison: The evaluation includes non-backpropagation baselines LAME, T3A, FOA, ZOA, and a vanilla zeroth-order estimator, as well as backpropagation methods such as TENT, CoTTA, SAR, DeYO, EATA, and RoTTA. This makes the relevant comparison a joint accuracy-memory tradeoff rather than accuracy alone.
• Results: On ImageNet-C at severity 5, CAZO reaches 69.0% average accuracy, compared with 65.8% for FOA and 67.5% for ZOA, while avoiding backward passes. In continual TTA without model reset, it reaches 65.3%, above LCoTTA at 62.3%. Runtime CUDA memory is 1,695 MB for CAZO versus 6,404 MB for TENT; in 8-bit and 6-bit settings, CAZO reports 67.8% and 61.2% accuracy respectively. 19
• Resources: Code is available at CAZO. 20

• Takeaway: The paper gives a concrete route to forward-only TTA when memory is the hard constraint. Its result is strongest on ImageNet-C severity 5 and should be checked against the reader's device, adapter size, perturbation budget, and domain-shift pattern.

Can Language Models Understand mmWave Data?

• Area tag: Radar-language reasoning and multimodal sensing.
• arXiv: 2608.14179v1, submitted August 14, 2026. 21
• Authors / institutions: Jeongwan Shin, Jaehyeon Kim, Donguk Ko, and Jaeho Choi at DGIST; the HTML also lists KAIST InnoCORE LLM. 22
• Status: Accepted to CVPR 2026 Findings. 21
• Problem: mmWave sensing works under darkness and occlusion, but it lacks the large radar-language corpus and standardized encoder ecosystem that RGB and audio enjoy. Existing radar models are usually retrained for each dataset, sensor, and scenario.
• Method: The paper serializes each 5D mmWave point into compact text and lets off-the-shelf LLMs answer questions without mmWave-specific tuning. It builds mmWave-QA by harmonizing mmBody, MM-Fi, and mRI, reducing 139 action labels to 86 categories, six scene conditions, and five question types: action recognition, trajectory checking, action order, action count, and limb focus.
• Comparison: Instead of training a new radar encoder, the paper tests a minimal textualization interface against standard zero-shot, few-shot, chain-of-thought, and few-shot chain-of-thought prompting. It also compares mmWave and RGB under visual degradation and reports performance across heterogeneous radar hardware.
• Results: GPT-5 is best in the few-shot CoT setting, improving overall accuracy by 1.39 percentage points over GPT-4o and 1.86 points over Gemini 2.5-Pro. For trajectory checking, zero-shot GPT-4o gets 4.0% on static cases versus 92.0% on moving cases; few-shot examples raise static accuracy to 32.0%. Models perform best on arm motions, while leg and torso actions remain difficult. 22
• Resources: The arXiv HTML describes the benchmark construction and evaluation, but the abstract and accessible page do not list a public code repository.
• Takeaway: This is a benchmark and interface paper, not evidence that LLMs already understand raw radar in the same way they understand language. Its useful contribution is a controlled way to measure whether textualizing a non-RGB modality is enough to unlock zero-shot reasoning.

Agents and robotics

RoboSynChallenge: Mastering Real-World Dexterity via Generalizing Synthesized Manipulation Skills

• Area tag: Embodied AI, dexterous manipulation, and Sim2Real evaluation.
• arXiv: 2608.12416v1, submitted August 12, 2026. 23
• Authors / institutions: Runyi Zhao, Ruixin Wu, Chengkun Li, Hongrui Zhang, Ang Li, Ruixing Jin, Yueci Deng, Yingying Guo, Lihe Ding, Shaocong Dong, Tianfan Xue, Yanjun Gao, Yudong Luo, Pascal Poupart, Simo Wu, Kui Jia, Wei-shi Zheng, and Guiliang Liu; affiliations span the Chinese University of Hong Kong, Shenzhen Loop Area Institute, DexForce, the Hong Kong University of Science and Technology, LARK Lab at the University of Colorado Anschutz, Mila, the Vector Institute, the University of Waterloo, Fudan University, and Sun Yat-sen University. 24
• Status: NeurIPS 2026 Competition Track. This is a competition proposal and benchmark, not an accepted model paper claiming a new policy. 23
• Problem: Simulation scales data collection, while real-world manipulation exposes dynamics, sensing, lighting, object, and distractor shifts. Existing benchmarks often test only simulation or only fixed real-world datasets, so they do not directly measure how synthesized data transfers to unseen physical environments.
• Method: RoboSynChallenge combines procedurally generated simulated state-action trials with standardized real-world evaluation on a dual-arm platform. It covers rigid, articulated, deformable, and tool-use tasks, three difficulty levels, five real-world variation factors, and baselines spanning ACT, Diffusion Policy, pi0, pi0.5, and Motus. The protocol measures success rate, action steps, and inference time on the same NVIDIA A800 hardware.
• Comparison: The benchmark differs from simulation-only suites such as RLBench, CALVIN, and LIBERO by making unseen real-world evaluation part of the official test. It also differs from fixed real-world benchmarks by using generative simulation data and a standardized Sim2Real co-training setting.
• Results: The proposal generates 1,000 simulated trials per task and collects 60 real-world samples per task across five conditions, four positions, and three orientations. In the preliminary baseline table, Motus reaches 14/20 on click-bell when trained with real data, while pi0.5 reaches 12/20 on items hand-over and place; the paper reports that simulation-trained baselines are comparable to, and sometimes outperform, real-only baselines on the tested tasks. These are baseline results for the challenge setup, not a leaderboard result. 24
• Resources: The paper links the RoboSynChallenge website, GitHub repository, and tutorial. 24

• Takeaway: The paper is useful for deciding whether an embodied-learning result has been tested at the level where deployment fails. Researchers should read the task definitions and real-world protocol before comparing a model's score with results from simulation-only benchmarks.

Reading order