2026-09-08

TL;DR - This paper introduces a black-box intervention framework for testing whether factors cited in LLM explanations are necessary or sufficient for model decisions. Across eight Claude, GPT, and Gemini models, cited factors were informative but only moderately aligned with measured behavioral influence, limiting their reliability for agent oversight.

• Controlled interventions measured necessity by changing a cited factor and sufficiency by retaining it while removing other changeable information.
• Mean rank correlations ranged from 0.349 to 0.580 across advisor-recommendation and prompt-monitoring tasks.
• In advisor recommendations, an uncited factor outperformed the weakest cited factor in roughly 58% of responses under both measures.
• Prompt monitoring showed better alignment: the corresponding rates were 25.8% for necessity and 8.9% for sufficiency.

TL;DR - WearableQA is a benchmark testing whether LLMs can reason over noisy, longitudinal health data from real wearable users. Results across 14 models show substantial capability differences, but most remain below 60% accuracy.

• Includes 4,084 ten-option questions derived from wearable time series, blood biomarkers, and demographics for 200 users with up to 500 days of measurements.
• Covers 16 question types spanning data computation versus health interpretation and single-signal versus cross-signal reasoning.
• Uses literature-grounded findings and statistically validated population patterns to generate reliable questions at scale.
• Model accuracy ranges from 19.6% to 72.9%, compared with a 10% chance baseline.

TL;DR - Conserved Immune Topology (CIT) augments pathology foundation-model embeddings with spatial immune descriptors to improve cross-cancer MSI-H prediction. It enables stronger zero-shot transfer across colorectal and stomach cancer cohorts without annotations or target-domain data.

• CIT captures tertiary lymphoid structures, peritumoral immune reactions, multi-scale tumor-infiltrating lymphocyte density, and immune-tumor mixing.
• Immune-associated image tiles are identified through unsupervised clustering of frozen foundation-model embeddings and spatial coordinates.
• On zero-shot cross-cancer transfer, CIT raised TransMIL AUC from 0.6627 to 0.7161, a 0.0534 absolute gain (p=0.003).
• Improvements held across all three tested multiple-instance learning aggregators despite scanner, site, and organ-specific distribution shifts.

TL;DR - This study maps the ecosystem for removing safety guardrails from open-weight models and finds that quantization, mirroring, and cross-platform redistribution make uncensored models difficult to eliminate. The persistence matters because a substantial share of downstream applications were classified as explicitly malicious.

• Researchers identified 3,471 original uncensored models on Hugging Face from January 2024 through March 2026.
• Each model was repackaged 2.4 times on average, producing 8,164 compressed redistributions; three actors accounted for 52% of them.
• Copies distributed across accounts, formats, and registries such as Ollama remained available even after upstream removal.
• Of 1,643 identified GitHub applications integrating uncensored LLMs, 25% were classified as explicitly malicious.

TL;DR - WorldSculpt generates cluttered 3D scenes as collections of individually grounded object meshes in a shared coordinate frame. It scales to hundreds of heavily occluded objects without scene-level training, enabling editable worlds for gaming, AR/VR, simulation, and robotics.

• Extends the Pixal3D single-object generative prior with multi-view conditioning over posed observations.
• Fine-tunes only on canonical single objects but generalizes to complex, densely cluttered scenes.
• Introduces UE-MeshyScene, a photorealistic benchmark with per-object annotations and ground-truth meshes.
• Outperforms prior methods across single-object and multi-object evaluations, with larger gains under greater complexity and occlusion.

TL;DR - UniMate is a topology-aware diffusion transformer that generates text-guided motion for arbitrary rigged 3D skeletons without per-skeleton retraining or test-time optimization. It aims to remove topology-specific constraints from learned animation and enable zero-shot motion transfer across diverse asset types.

• Encodes skeletal structure through graph-aware attention biases, graph-Laplacian spectral rotary embeddings, and a global rest-pose topology conditioner.
• Trains on UniML3D, a curated dataset of 13,006 text-paired motion sequences spanning animals, articulated objects, and varied skeletal topologies.
• Reportedly outperforms existing baselines in motion quality, generalization, and efficiency.
• Supports zero-shot cross-topology transfer, motion in-betweening and expansion, and text-guided editing.

TL;DR - WeAgent-MMGenEdit is a full-stack framework for knowledge-intensive, agentic image generation and editing that combines multimodal retrieval, evidence management, verification, and post-training. Its compact 30B-total/3B-active policy approaches the reported performance of a 1T-parameter agent.

• WeAgent-Harness persistently manages retrieved evidence and provides dedicated tools to verify and integrate textual and visual information.
• Its training pipeline produces 23K supervised trajectories and 14.7K reinforcement-learning tasks with three-layer verifiable checklists.
• WeBench-MMGenEdit evaluates bilingual, knowledge-intensive image generation and multi-image editing.
• Two-sided SFT and RL post-training improves both the agent policy and the image-generation backend.

TL;DR - This paper benchmarks view-dependent appearance models for 3D Gaussian Splatting and introduces compact per-primitive latent codes decoded by a shared MLP. The neural representation cuts appearance storage from 192 to 28 bytes per primitive while improving reconstruction quality and accelerating optimization by 1.3× over third-degree spherical harmonics.

• Compares spherical harmonics, newer spherical models, and the proposed neural representation in one optimized CUDA rasterization pipeline.
• Finds recent spherical models offer the best overall balance between reconstruction quality and efficiency.
• Provides a portable WebGL viewer targeting laptop and mobile GPUs.
• Shows that appearance parameterization affects recovered geometry and how expressive models absorb non-static scene content.

TL;DR - This paper proposes behavioral correctness assumptions for diagnosing how reference-based natural language generation evaluators respond to controlled, correctness-preserving or correctness-altering transformations. It reveals behavioral differences and trade-offs that aggregate agreement scores can obscure.

• The framework defines a taxonomy of expected evaluator behaviors and tests them using controlled response transformations.
• It covers lexical, character-level, semantic, LLM-based, and hybrid evaluation methods.
• Evaluators are analyzed for stability, sensitivity, repeat-run variability, configuration sensitivity, and reproducibility.
• No evaluator satisfies every proposed assumption, and evaluators with similar aggregate performance can have substantially different behavioral profiles.

TL;DR - RoboSPA is a large-scale benchmark for evaluating spatial reasoning and long-horizon procedural planning in Vision-Language-Action models. It reveals persistent weaknesses in complex spatial understanding, precise execution, and memory-intensive planning.

• Covers 10 task categories and 56 base tasks, expanded across five difficulty levels into 280 variants.
• Includes 527K manipulation trajectories spanning multiple robot embodiments and diverse scenes.
• Adds diagnostic metrics beyond binary task-success rates to identify specific reasoning and execution failures.
• Evaluations of representative VLA models show substantial difficulty as spatial ambiguity and procedural complexity increase.