LFM2.5-Encoders for Fast Long-Context Inference on CPU
LFM2.5-Encoders targets the biggest bottleneck for long-context LLM use on non-GPU hardware by rethinking how long sequences get encoded for faster inference. The result is a more deployment-friendly path for CPU-based long-context workloads, paired with an auditable setup intended to make performance claims easier to verify.
AI leaders sign statement asking the government to do something about automated AI
Employees from major frontier labs argue that rapid advances in automated AI research and capability iteration justify faster, coordinated governance – even if they also acknowledge the technology’s potential to improve the future. The statement presses the US government toward clearer action and faster oversight as AI systems become better at driving their own progress.
SCOPE and SCION: A Benchmark and an Auditable Reference Pipeline for Schema Induction and Fusion from Text
SCOPE introduces a train-text-only benchmark for building schema graphs from raw text, tackling the upstream bottleneck that many extraction pipelines assume away by starting with an existing schema. SCION then delivers an auditable reference pipeline that restricts naming, merges and validates candidates under strict JSON contracts, producing evidence-linked outputs rather than “black-box” schema claims.
AgentKVShift: Efficient KV Cache Reuse for Agentic Memory Systems
AgentKVShift tackles a real latency killer in long-horizon agents – every memory retrieval currently forces expensive re-encoding into KV cache states. By decomposing per-memory KV reuse residuals into a shared offset plus small token fluctuations, the method refreshes only 10-30% of the cache while approaching near-full recompute quality.
TILT: Improving Compositional Generation in Diffusion Models with a Model-Intrinsic Reward
TILT improves compositional text-to-image prompting by aligning the sampling trajectory at test time, not by retraining the diffusion model. It defines an intrinsic reward that prefers samples where all concepts appear jointly, then uses a KL-constrained objective with principled tilted guidance steps to reduce compositional failures.
Securing Multimodal AI through Internal Information Decomposition
FlowGuard reframes multimodal security as an internal consistency problem: benign inputs yield compatible fused behavior between text-only and vision-only reasoning, while adversarial attempts break that alignment. It uses FlowVectors inspired by Partial Information Decomposition to detect harmful inputs at inference time without relying on scalar confidence heuristics, cutting attack success rates dramatically with low utility loss.
FlowEvo: Self-Evolving Agents through the Co-Evolution of Workflows and Executable Skills
FlowEvo turns short-lived agent traces into persistent “skill records” that can be reused later, letting agents accumulate capability at inference time without parameter updates. It compiles workflows into executable artifacts, retrieves relevant skills for future runs, and suppresses harmful negative-transfer skills based on downstream utility.
Risk Is Not the Target: A Monotonic Framework for Evaluating Wildfire Operational Risk Signals
This work argues that standard ML metrics like F1 and IoU don’t measure what matters operationally – namely whether risk scores behave consistently as real-world workload rises. The proposed monotonic evaluation framework tests whether higher predicted risk scores reliably correspond to increases in observed intervention load, comparing expert indices, GRU predictors, and a hybrid multi-agent system.
From Frame-Level Recognition to Event-Level Confirmation: Repair Traces and Runtime Failure Analysis of Public-Space Gesture Interaction
This paper zooms out from “accuracy on frames” to what deployed gesture systems must actually deliver: stable interaction events that confirm user intent. Based on repair records from a real public kiosk deployment, it catalogs failure modes like temporal mismatch, lifecycle runtime issues, and feedback synchronization breakdowns, grounding future evaluation in event-level confirmation.
Transferable Latency Prediction for Fast LLM Screening on Heterogeneous Edge Devices
Accurate latency modeling is crucial when you need to screen which LLM will run well across wildly different edge hardware, runtimes, and thermals. This framework splits inference into prefill and decode, fuses static descriptors with dynamic telemetry via a gated predictor, and shows improved R-squared on Pixel and Jetson-like setups while reducing profiling cost.
Perplexity’s Personal Computer turns Windows PCs into AI agents
Perplexity expands its agentic “Personal Computer” tool to Windows, enabling locally running AI actions like creating documents and updating spreadsheets using files and installed apps. The move builds on earlier integrations for Microsoft 365 and Teams, aiming to close the usability gap beyond the Mac version.
Scientific computing in the age of agentic AI
OpenAI describes how agents can modernize scientific computing by combining reasoning, tool use, and code execution to speed up software development and accelerate discovery workflows. The emphasis is on practical agentic execution patterns that reduce friction between ideas and running experiments.
Samsung’s chip workers are jumping ship to rival SK Hynix
A new report details how employees in Samsung’s semiconductor division are leaving for SK Hynix, highlighting shifting labor dynamics in the race to scale advanced manufacturing. It adds a human, operational angle to the broader AI hardware story – where capacity and expertise matter as much as new silicon roadmaps.
The OlmoEarth Platform: Geospatial inference at planetary scale
OlmoEarth positions itself as infrastructure for running geospatial inference models at planetary scale, turning Earth-data workloads into something closer to on-demand AI computation. For practitioners, the key value is reducing friction between large-scale geographic datasets and deployable inference pipelines.
Samsung’s chip workers are jumping ship to rival SK Hynix
Graph neural networks can fail long-range tasks even when models look capable, because some graph topologies silently block information propagation. Spectral Flow Certificates provide a training-free single-scalar diagnostic based on graph Laplacian spectra that flags topology limits before expensive training.