Chronological feed of everything captured from Stephanie Zhan.
Despite strong performance on existing benchmarks, current AI agents fail to generalize to economically meaningful workflows — a gap the authors attribute primarily to flawed evaluation methodology. Agents' Last Exam (ALE) addresses this by benchmarking agents on 1,000+ long-horizon, verifiable, real-world tasks across 13 industry clusters, developed with 250+ domain experts using O*NET/SOC 2018 as an occupational reference frame. The hardest task tier remains nearly unsolved, with an average full pass rate of just 2.6% across mainstream agent configurations, signaling that current systems are far from production-ready for professional deployment. ALE is designed as a living benchmark that expands continuously, positioning it as an instrument for measuring GDP-relevant AI impact rather than leaderboard performance.
KV-Control introduces a parameter-efficient adapter for frozen text-to-motion transformers that injects trajectory constraints as key/value memories into every self-attention layer, leaving the query stream, text cross-attention, FFN, and all backbone weights untouched. This avoids the two dominant failure modes of existing approaches — duplicating large model components for layer-wise control access, or relying on expensive test-time optimization. The system is co-designed with PartVQ, an anatomy-aligned part codebook tokenizer, and T-Concat, which exposes per-frame-part tokens as attention-addressable sites, enabling precise geometric control while preserving text-conditioned motion style. The result is sub-centimeter tracking accuracy on root and multi-joint trajectory constraints with only lightweight injection parameters trained on top of a shared trajectory encoder.
This paper reframes ESG not as a source of unconditional return alpha, but as a multi-channel signal for tail-risk and fragility reduction during market stress. Using S&P 500 monthly data from 2014–2025, the authors construct a "cofragility" state capturing simultaneous adverse downside returns, volatility spikes, and illiquidity within the same firm-month. The core finding is that ESG's protective association is stress-amplified: benefits in return, volatility, and liquidity channels are concentrated during periods of aggregate market weakness, not normal conditions. Double Machine Learning confirms the ESG–cofragility relationship is robust to flexible controls for observable firm characteristics.
This paper demonstrates that a biaxial-tilt overactuated hexacopter (BTO) can passively maintain full 6-degree-of-freedom actuation following abrupt, unknown total rotor failures — a regime where conventional multirotors lose controllability entirely. Two fault-tolerant schemes are proposed: one at the controller layer using a High-Order Fully Actuated (HOFA) controller with a Linear Extended State Observer (LESO), and one at the allocator layer using model-reference adaptive control allocation with momentum-based wrench estimation. Critically, neither scheme requires explicit fault detection, isolation, or mode-switching, making the approach robust to a priori unknown failures. The BTO geometry is shown to offer strictly larger attainable wrench space recovery margins than uniaxial-tilt and coplanar hexacopter designs under up to three simultaneous failures.
LiAuto-GeoX is a compact 155M-parameter onboard transformer that distills a large-scale driving geometry model into a deployable, ego-centric 3D scene understanding system. It leverages sparse LiDAR priors for geometric grounding and uses a two-stage knowledge distillation framework—combining mask-guided depth-aware distillation and relative-pose relational distillation—to preserve metric fidelity and cross-view spatial consistency. The result is a model that runs at 220 FPS on KITTI while achieving competitive performance across trajectory, occupancy, and future-frame prediction tasks, positioning dense 3D reconstruction as a general-purpose geometric backbone for autonomy stacks.
This paper presents a first-principles analytical hydrodynamic model for bacterial transport in low-Reynolds-number flows, resolving a long-standing gap in active matter theory. The central result is that shear-induced bacterial depletion near walls and flow-alignment are not independent phenomena but dual manifestations of a single active phase-space topology. The model introduces a "hydrodynamic locking" mechanism with mean-drift invariance — framed as an active analogue to Taylor dispersion — and is validated against microfluidic experiments across multiple bacterial species and flow geometries from 1D to 3D.
Standard LoRA and other PEFT methods were designed for autoregressive models and apply static weight updates that are blind to noise level — a fundamental mismatch for diffusion LLMs, where input distributions and generation difficulty shift continuously along the denoising trajectory. NaRA addresses this by injecting a lightweight, globally shared hypernetwork that generates a low-rank core matrix conditioned on the current noise level, allowing update matrices to adapt dynamically across the diffusion process. The approach adds negligible parameter and latency overhead while yielding consistent benchmark gains over noise-agnostic baselines on commonsense reasoning, mathematical reasoning, and code generation tasks.
GR3D is a spatial vision-language model (VLM) that unifies three grounding modalities—explicit 2D, implicit 2D, and monocular 3D—within a single framework to improve spatial understanding. The implicit grounding mechanism inserts region tokens inline during text generation, enabling on-the-fly visual referencing during spatial chain-of-thought reasoning. Monocular 3D bounding box prediction is achieved via region-prompted queries with intrinsic-aware normalization and dense geometric supervision, allowing 3D inference to be anchored in grounded 2D perception. Crucially, the grounding capabilities generalize beyond grounding tasks themselves, improving performance on broader non-grounded spatial benchmarks.