2026-07-31

TL;DR - Tycho is a coding-agent system that builds and selectively uses executable world models to solve interactive ARC-AGI-3 games efficiently. Its results suggest that deciding when to construct, repair, use, or bypass a model is as important as simulator accuracy.

• Tycho models games as parameterized rendered deterministic Moore machines and separates actionable states from animation and terminal frames.
• Actor-requested delegation to a model builder achieved the best tested orchestration result, averaging 88.49 Relative Human Action Efficiency across 25 public games.
• With that policy, GPT-5.6 Sol and Opus 5 completed all 183 levels at 100.00 RHAE; Opus 5 used 61% fewer scored actions than aggregate official human baselines.
• Automatic model repair improved transition reproduction but reached only 83.07 RHAE, showing that dynamics accuracy alone does not ensure effective planning or objective discovery.

TL;DR - ClawTrack evaluates autonomous agents at both outcome and reasoning-trace levels, exposing lucky successes and specific process failures. Its results identify inadequate result verification as a recurring weakness and show that process-based trajectory filtering improves post-training.

• Includes 320 tasks across 8 domains, 25+ deterministic mock services, and 12,541 task-specific rubric items.
• Scores each reasoning turn on goal alignment, efficiency, information use, and result verification.
• Evaluation of 21 models over 16,000+ trials found complementary process dimensions and robustness across judge LLMs.
• Trace-level scoring enables more precise failure attribution than outcome-only benchmarks.

TL;DR - KAISEN is a reproducible, five-phase framework for auditing subgroup fairness in clinical risk models. Synthetic stress tests show that mitigation, diagnostics, and drift-monitoring components can fail unpredictably or silently.

• Per-group threshold optimization reduced equalized-odds disparity in all 48 held-out runs.
• Group-wise Platt scaling improved calibration but had inconsistent, near-zero average effects on fairness.
• Mechanism diagnostics identified all controlled cases but missed every model-driven case under proxy misspecification without warning.
• CUSUM drift thresholds transferred poorly across cohort realizations; the synthetic results do not establish clinical validity.

TL;DR - READII-2-ROQC uses volume-preserving negative controls to test whether radiomics and imaging foundation-model features capture meaningful spatial signals rather than tumour volume or acquisition artifacts. Across three cancer imaging cohorts, it exposed confounding in models that remained predictive after image structure was destroyed.

• Generates voxel-perturbed controls for tumour, background, and whole-image regions while preserving volume.
• Evaluated PyRadiomics and foundation-model features across 3,552 tumour volumes and nine matched controls.
• Reproduced survival and HPV-status signatures, finding both confounded and genuinely perturbation-sensitive models.
• Provides an open-source quality-control framework for more interpretable and reproducible imaging biomarkers.

TL;DR - This paper formalizes derived-feature over-trust, where LLMs treat uncertain sensor-derived measurements as direct facts, and proposes metrics and reliability evidence to evaluate and mitigate it in physiological sensing.

• Tests over-trust using PPG-derived heart rhythms checked against privileged offline ECG references never shown to the LLM.
• Introduces five metrics covering conflicting evidence, context-induced errors, error repair, evidence specificity, and unnecessary verification.
• Evaluates privileged ECG-to-PPG distillation on 50,000 paired records and a protocol-locked 187-patient test set.
• The baseline improved four repair and specificity endpoints by 1.82–6.69 percentage points; verification-related harm rose by 0.67 points, with its confidence interval spanning zero.

TL;DR - This paper proposes the Bifurcation Connectedness Score (BCS) for coronary artery segmentation, measuring vessel-tree connectivity that overlap metrics can miss. Better BCS correlates with closer FFR-CT treatment-decision agreement, especially in severe disease.

• BCS evaluates connectivity at each ground-truth bifurcation and strongly detects localized vessel breaks.
• It remains largely stable under narrowing that preserves connectivity, distinguishing topology from volumetric overlap.
• Higher BCS was associated with better predicted-versus-ground-truth FFR-CT decision agreement in severe disease (OR 2.16, CI [1.23, 4.18]).
• Soft-BCS provides a differentiable training surrogate; results suggest branch recovery and branch connectivity should be reported separately.

TL;DR - PathVU is a large, vision-anchored benchmark testing whether multimodal models understand pathology images across local regions and whole-slide views. Results across 18 models reveal substantial limitations in fine-grained, multiscale visual reasoning.

• Includes 14 VQA tasks, 61,673 images, and 308,070 samples from 23 public datasets.
• Covers 28 organs using over 7.25 million human-supervised labels and spatial annotations.
• Tests localization, recognition, quantity estimation, spatial reasoning, and insufficient-context judgment.
• Uses deterministic targets for reproducible, programmatic scoring rather than evaluating only final diagnoses or reports.

TL;DR - Chimera is a hybrid diffusion transformer designed for compute-efficient, long-context image and video generation. Its linear attention, sparse MoE, and scaling recipe enable substantial efficiency gains and zero-shot video-length extrapolation.

• Combines O(N) Kimi Delta Attention, interleaved global latent attention, and modality-aware local convolutions.
• HeteroP transfers hyperparameters across heterogeneous modules to derive Chinchilla-style compute-optimal scaling laws.
• The 11B-parameter model activates 2B parameters and achieves up to 7.3× the compute efficiency of a matched full-attention baseline.
• It extrapolates from 5-second training clips to 30-second videos, with 6.5% FID degradation in the final five seconds.

TL;DR - EgoGenesis generates controllable egocentric robot-manipulation videos using geometry-aware memory and action conditioning. Its synthetic trajectories improve real-robot generalization, especially for dual-arm tasks.

• OAPM anchors generation to the first-frame 3D scene while refreshing recent state during long autoregressive rollouts.
• A3D-RoPE injects camera-aware end-effector motion into skeleton-to-video cross-attention for precise action control.
• Adding 400 synthetic trajectories to 400 real ones raises out-of-distribution success from 77% to 84% for single-arm tasks.
• Dual-arm success improves from 53% to 70% with the same augmentation strategy.

TL;DR - Chem World is a standardized chemical property prediction benchmark spanning 17 datasets and more than 800,000 molecules. It also introduces Mixture-PINN, which integrates chemical priors to improve prediction accuracy, robustness, and reliability.

• Covers properties including density, electrical conductivity, solubility, and other molecular characteristics.
• Unifies diverse datasets and evaluation protocols for systematic model comparison.
• Mixture-PINN combines physics-informed constraints with data-driven neural learning.
• Experiments reportedly outperform existing methods, though the abstract provides no quantitative results.

TL;DR - UNICON is a foundation model that learns predictive relationships from graph-based numerical examples and generalizes across scientific and social systems. It approaches specialist performance without retraining, including in disciplines absent from training.

• Infers shared predictive operators from in-context examples and applies them to new queries.
• Uses one model across multiple disciplines rather than training task-specific specialists.
• Pairing UNICON with language-model agents reportedly surpasses state-of-the-art specialists in an unseen discipline.
• Greater training-corpus diversity improves cross-disciplinary generalization.