arXiv:2602.06074v1 Announce Type: new Abstract: Performance in web applications is a key aspect of user experience and system scalability. Among the different techniques used to improve web application performance, caching has been widely used. While caching has been widely explored in web performance optimization literature, there is a lack of experimental work that explores the effect of simple inmemory caching in small-scale web applications. This paper fills this research gap by experimentally comparing the performance of two server-side […]

Ver mais

Modern production systems generate more data than most developers can realistically process. Every request emits logs. Every service exports metrics. Every dependency introduces another layer of signals. In theory, this should make systems easier to understand. In practice, it does the opposite. Dashboards become dense, alerts become noisy, and when something breaks, the same questions still come up. What is actually wrong? Who is affected? Where do you even start? The problem is not observability. It is interpretation. […]

Ver mais

I enjoyed my conversations with Timnit when she was in the MSR-NYC lab, so her situation has been on my mind throughout NeurIPS. Piecing together what happened second-hand is always tricky, but Jeff Dean’s account and Timnit’s agree on a basic outline. Timnit and others wrote a paper for FAccT which was approved for submission by the normal internal review process, then later unapproved. Timnit threatened to leave unless various details about this unapproval were clarified. Google then […]

Ver mais

arXiv:2603.06798v1 Announce Type: cross Abstract: The growing scale of deep learning demands distributed training frameworks that jointly reason about parallelism, memory, and network topology. Prior works often rely on heuristic or topology-agnostic search, handling communication and memory separately. Without per-device memory awareness, these methods typically ensure feasibility post hoc by sharding parameters and activations across many devices, increasing synchronization, inflating communication, and underutilizing compute-limiting scalability and efficiency on real datacenter networks. We present NEST, a network-, compute-, and […]

Ver mais

arXiv:2603.23536v1 Announce Type: new Abstract: Atomistic structural data are central to materials science, condensed matter physics, and chemistry, and are increasingly digitised across diverse repositories and databases. Interoperable access to these heterogeneous data sources enables reusable clients and tools, and is essential for cross-database analyses and data-driven materials discovery. Toward this aim, the OPTIMADE (Open Databases Integration for Materials Design) specification defines a standard REST API for atomistic structures and related properties. However, deploying and maintaining compliant services […]

Ver mais

arXiv:2601.01480v2 Announce Type: replace Abstract: Large-scale traffic forecasting relies on fixed sensor networks that often exhibit blackouts: contiguous intervals of missing measurements caused by detector or communication failures. These outages are typically handled under a Missing At Random (MAR) assumption, even though blackout events may correlate with unobserved traffic conditions (e.g., congestion or anomalous flow), motivating a Missing Not At Random (MNAR) treatment. We propose a latent state-space framework that jointly models (i) traffic dynamics via a linear […]

Ver mais

World models learn to simulate environment dynamics from experience, enabling sample-efficient reinforcement learning. But what do these models actually represent internally? We apply interpretability techniques–including linear and nonlinear probing, causal interventions, and attention analysis–to two architecturally distinct world models: IRIS (discrete token transformer) and DIAMOND (continuous diffusion UNet), trained on Atari Breakout and Pong. Using linear probes, we find that both models develop linearly decodable representations of game state variables (object positions, scores), with MLP probes yielding only […]

Ver mais

arXiv:2602.05230v1 Announce Type: cross Abstract: Linear attention methods offer Transformers $O(N)$ complexity but typically underperform standard softmax attention. We identify two fundamental limitations affecting these approaches: the restriction to convex combinations that only permits additive information blending, and uniform accumulated weight bias that dilutes attention in long contexts. We propose Zero-Sum Linear Attention (ZeroS), which addresses these limitations by removing the constant zero-order term $1/t$ and reweighting the remaining zero-sum softmax residuals. This modification creates mathematically stable weights, […]

Ver mais

DPO (Rafailov et al., NeurIPS 2023) is supposed to be the clean alternative to PPO. No reward model in the training loop, no value function, no rollout collection. Just a binary cross-entropy loss over preference pairs. And the math is elegant the partition function Z(x) cancels out when you substitute the log-ratio reparameterisation into the Bradley-Terry model. I implemented it from scratch as part of a multi-stage RLHF project (same model, same tokenizer, same evaluation suite as my […]

Ver mais