Top-conf paper digest — week of July 6-10, 2026

The July 6-10 arXiv batch is mostly about controlled failure analysis: extracting hidden model state, detecting stealthy multi-agent attacks, diagnosing formalization errors, and stress-testing visual geometry. I kept the issue to papers whose arXiv records explicitly state ICML 2026 acceptance, and excluded workshop-only entries and older cross-list candidates whose first submission date fell outside this issue's window.

Selection basis

Window: arXiv submissions or current versions dated July 6-10, 2026.
Status rule: include confirmed NeurIPS / ICML / ICLR / CVPR main-conference papers. This issue's selected set is all ICML 2026-confirmed; workshop-only papers are not included.
Ranking rule: prioritize papers that expose a method mechanism and at least one decision-useful result, then keep area coverage across LLMs, agents, vision, and tabular ML.
Affiliations: the arXiv abstract pages for the selected entries list authors but do not expose institutional affiliations, so the entries below report authors and mark affiliations as not exposed rather than inferring them.

Area	Paper	Status	Why it is in this issue
LLM / auditing	Overthinking	ICML 2026	Uses reasoning-weight amplification to elicit hidden information from 2B-32B models, with up to 10x more secret surfacing than the base reasoning model. 1
Agents / security	When Agents Go Rogue	ICML 2026	Detects stealthy malicious behavior in asynchronous multi-agent systems from local activation states, beating graph baselines by +0.55 F1 in the asynchronous setting. 2
LLM / KG reasoning	RSF-GLLM	ICML 2026 full paper	Keeps knowledge-graph retrieval differentiable through recurrent soft-flow, then uses extracted paths to ground LLM answer generation. 3
Formal reasoning	FormalRx	ICML 2026	Turns autoformalization evaluation from a binary verdict into diagnosis, localization, and correction over a 28-category error taxonomy. 4
Vision-language	Scene Graph Thinking	ICML 2026	Builds 120K scene-graph reasoning traces and trains MLLMs to navigate explicit object-relation structure. 5
Vision theory	A Theory of Contrastive Learning with Natural Images	ICML 2026	Derives when contrastive learning with common augmentations should learn sinusoidal first-layer filters and partial whitening. 6
Vision robustness	Geometric Collapse	ICML 2026	Shows depth predictors can be pulled off course by physically unsupported edge cues, with deviations up to 3.2x larger than matched noise. 7
Video generation	Geometric Reciprocity	ICML 2026	Gives a self-supervised route for monocular-to-stereo video by analytically deriving disocclusion masks from cycle consistency. 8
Tabular ML	TabPack	ICML 2026	Packs many MLP hyperparameter choices into one efficient ensemble run, reducing reliance on exact manual tuning. 9

LLMs, agents, and formal reasoning

Overthinking: reasoning-weight amplification as a black-box audit tool

Area tag: LLM / safety auditing arXiv: 2607.08173 Authors / institutions: Jack Hopkins, Dipika Khullar, and Fabien Roger. The arXiv abstract page lists authors but does not expose institutional affiliations. Peer-review status: ICML 2026, confirmed in the arXiv comments field. 1

Problem: Standard black-box audits can miss hidden information or subtle misalignment when a model has learned something but does not reveal it in ordinary prompting.

Method: The paper defines an overthinking model by taking the parameter difference between a non-reasoning instruct model and a reasoning-distilled model, then amplifying that reasoning direction with a coefficient greater than one. It also introduces layer-wise attenuation so the amplification does not simply destroy output quality.

Comparison with prior work: Instead of changing prompts or adding external audit tasks, the method perturbs model weights along a reasoning task vector. That makes it closer to representation-level auditing than ordinary red-teaming.

Result / takeaway: Across 2B-32B models and four experimental settings, the abstract reports that reasoning amplification surfaces secrets or unintended behavior up to 10x more often than the original reasoning model. Code or a project page is not listed on the arXiv abstract page. 1

When Agents Go Rogue: activation-space detection for multi-agent attacks

Area tag: Agents / security arXiv: 2607.06807 Authors / institutions: Haowen Xu, Xue Tan, Lei Ma, Zhihao Zhang, Chao Wang, Qingze Wang, Ping Chen, Jun Dai, and Xiaoyan Sun. The arXiv abstract page lists authors but does not expose institutional affiliations. Peer-review status: Accepted to ICML 2026. 2

Problem: Multi-agent LLM systems can be attacked at the agent level or through interactions, but graph-based defenses assume explicit topology and temporal synchronization. Those assumptions are weak when agents run asynchronously.

Method: AcMAS reads internal reasoning states from local agent activations and detects malicious behavior without relying on an explicit interaction graph. The same activation signals guide recovery of compromised agents, rather than isolating them outright.

Comparison with prior work: The contrast is with graph-propagation defenses that model the multi-agent system as a synchronized interaction graph. AcMAS works in activation space, so it can handle semantically stealthy attacks and asynchronous execution.

Result / takeaway: The reported F1 score is 0.94 versus 0.72 for graph baselines in synchronous settings, and 0.93 versus 0.38 in asynchronous settings. The arXiv page does not list code or a project page. 2

RSF-GLLM: differentiable soft-flow before LLM generation

Area tag: LLM / knowledge-graph QA arXiv: 2607.06527 Authors / institutions: Sambaran Bandyopadhyay and Ananth Muppidi. The arXiv abstract page lists authors but does not expose institutional affiliations. Peer-review status: ICML 2026 full research paper, confirmed in the comments field. 3

Problem: Multi-hop KGQA often breaks into retrieve-then-read stages. Once retrieval is discrete, the retriever cannot learn through semantically dissimilar bridge nodes that lack lexical overlap with the query.

Method: RSF-GLLM separates differentiable graph reasoning from answer generation. A recurrent soft-flow module uses a GRU-guided query updater and dynamic gating to propagate continuous relevance scores; sparse flow regularization pushes those scores toward discrete paths, which are then textualized for LLM fine-tuning.

Comparison with prior work: The paper positions this against computationally expensive LLM-only reasoning and against non-differentiable retrieve-then-read pipelines.

Result / takeaway: The abstract reports competitive performance on WebQSP and CWQ with better inference efficiency than LLM-heavy approaches, but it does not expose headline numeric scores. No code link is listed on the arXiv abstract page. 3

FormalRx: error taxonomy for autoformalization

Area tag: Formal reasoning / evaluation arXiv: 2607.04655 Authors / institutions: Haocheng Wang, Baiyu Huang, Yingjia Wan, Xiao Zhu, Xiaoyang Liu, Yinya Huang, and Zhijiang Guo. The arXiv abstract page lists authors but does not expose institutional affiliations. Peer-review status: Accepted at ICML 2026. 4

Problem: Autoformalization benchmarks usually say whether a natural-language-to-formal-language translation passed, but not where the semantic error sits or how to repair it.

Method: FormalRx introduces the SCI Error Taxonomy with 28 error categories and a diagnostic pipeline for verdicts, categorization, localization, and correction. The authors instantiate it with FormalRx-8B, trained on 56,287 annotated natural-language / formal-language pairs.

Comparison with prior work: The evaluation target shifts from scalar pass/fail scoring to actionable diagnosis. That matters for systems that need to improve a failed formalization, not merely reject it.

Result / takeaway: FormalRx-8B reports 0.88 F1 for alignment verdicts, 0.71 F1 for error categorization, 0.75 accuracy for localization, and 0.73 accuracy for correction. The abstract says the paper releases FormalRx-Test, but the arXiv page does not list a direct repository link. 4

Vision, multimodal reasoning, and generation

Scene Graph Thinking: explicit relation structure for MLLMs

Area tag: Vision-language / MLLM reasoning arXiv: 2607.05716 Authors / institutions: Zhiwei Yang, Yuanchen Wu, Nan Zhang, Yucong Meng, Ke Yan, and Shouhong Ding. The arXiv abstract page lists authors but does not expose institutional affiliations. Peer-review status: ICML 2026; v1 was submitted on July 7, and the current v2 was revised on July 9. 5

Problem: Many MLLMs can recognize objects but handle structured object relations less efficiently, especially when a task requires navigating visually dense scenes.

Method: SaGe converts flat image-text corpora into scene graphs with hierarchical entity nodes and relation edges. It then samples 120K reasoning traces and applies two graph-aligned post-training stages: supervised fine-tuning for structured reasoning and reinforcement fine-tuning with node-as-proxy graph rewards.

Comparison with prior work: Instead of treating image-text data as unstructured supervision, SaGe turns spatial and relational structure into the training target.

Result / takeaway: The abstract reports improvements across eight multimodal benchmarks but does not expose a single headline numeric score. Code is listed at zwyang6/SaGe. 5

A Theory of Contrastive Learning with Natural Images: why simple filters emerge

Area tag: Vision theory / representation learning arXiv: 2607.07470 Authors / institutions: Antonio Torralba and Yair Weiss. The arXiv abstract page lists authors but does not expose institutional affiliations. Peer-review status: ICML 2026. 6

Problem: Contrastive learning with simple image augmentations often learns useful representations, but the mechanism is usually studied empirically rather than solved analytically.

Method: The paper analytically computes optimal representations for contrastive loss under basic augmentations and stationary image statistics. For some augmentations, the optimum can be implemented by a CNN with sinusoidal first-layer filters, pointwise nonlinearity, global average pooling, and partial whitening in the final linear layer.

Comparison with prior work: The contribution is theoretical rather than another larger contrastive pretraining recipe. It connects learned filters to the expected power spectrum of the dataset through a waterfilling algorithm.

Result / takeaway: Experiments across image datasets and augmentations show that SGD-trained CNNs empirically learn sinusoidal first-layer filters and partial whitening, but the abstract does not report a single benchmark score. Code is not listed on the arXiv abstract page. 6

Geometric Collapse: unsupported edges can break dense prediction

Area tag: Vision robustness / depth prediction arXiv: 2607.06871 Authors / institutions: Wentao Zhang, Jinhu Qi, Weiqiang Jin, Yifei Zhang, Chan-Tong Lam, and Irwin King. The arXiv abstract page lists authors but does not expose institutional affiliations. Peer-review status: ICML 2026; the arXiv page also lists poster #1310. 7

Problem: Dense geometric predictors may use edge evidence without checking whether the implied geometry is physically plausible.

Method: The authors introduce Scrambled Edges, a controlled counterfactual that injects salient edge-like cues while breaking surface continuity, illumination coherence, and occlusion ordering. Energy-matched and structure-matched controls separate unsupported edge evidence from ordinary high-frequency noise.

Comparison with prior work: The paper tests CNN, ViT, and self-supervised depth predictors, then extends the analysis to diffusion and flow-matching depth estimators.

Result / takeaway: On NYU Depth v2 and KITTI, Scrambled Edges produce up to 3.2x larger deviation from clean predictions than energy-matched noise. Even with oracle knowledge of the corrupted region, output-level repair recovers only 47%. Code is not listed on the arXiv abstract page. 7

Geometric Reciprocity: self-supervised stereo from monocular video

Area tag: Vision / stereoscopic video generation arXiv: 2607.05354 Authors / institutions: Jingyi Lu and Kai Han. The arXiv abstract page lists authors but does not expose institutional affiliations. Peer-review status: Accepted to ICML 2026. 8

Problem: Monocular-to-stereo conversion depends on inpainting disocclusions. Training-based methods can work well, but stereo video pairs are scarce and synthetic data can shift away from real monocular video.

Method: The paper introduces the Geometric Reciprocity Theorem: under nearest-neighbor depth-image-based rendering, the disocclusion mask for target-view synthesis equals the set of pixels lost when warping back from the target to the source. That makes test-time disocclusion masks computable from monocular inputs.

Comparison with prior work: The self-supervised setup learns from monocular videos instead of requiring stereo pairs or synthetic stereo supervision.

Result / takeaway: The abstract reports substantial improvements over training-free and supervised state-of-the-art methods, but it does not expose the exact benchmark numbers. A project page is listed at visual-ai.github.io/grt. 8

ML methods and tabular systems

TabPack: hyperparameter ensembles without heavy tuning

Area tag: Tabular ML / efficient ensembles arXiv: 2607.05380 Authors / institutions: Yury Gorishniy, Akim Kotelnikov, Ivan Rubachev, and Artem Babenko. The arXiv abstract page lists authors but does not expose institutional affiliations. Peer-review status: ICML 2026. 9

Problem: Deep tabular models can depend heavily on hyperparameter tuning, and efficient MLP ensembles often use the same hyperparameters across all ensemble members.

Method: TabPack trains many MLPs with sampled hyperparameters in parallel and selects ensemble members on the fly. Users specify sampling ranges rather than a single tuned configuration.

Comparison with prior work: The baseline target is extensively tuned tabular deep learning methods. TabPack aims to keep competitive performance while reducing tuning effort.

Result / takeaway: The abstract reports that default TabPack matches extensively tuned prior methods on medium-to-large public datasets, and that one default run on a modern MacBook took less time than tuning some baselines on an industry-grade GPU. Code is listed at yandex-research/tabpack. 9

Reading order

Start with Overthinking if you audit LLMs or care about hidden model behavior. Read When Agents Go Rogue next if your work touches agent orchestration or multi-agent security. For vision researchers, Geometric Collapse is the sharpest empirical stress test, while A Theory of Contrastive Learning with Natural Images is the theory-first entry. If you build applied tabular systems, TabPack is the fastest paper to triage because its practical question is simple: can one ensemble run replace a tuning loop?