Your daily OSS contribution digest

This paper demonstrates that high-quality text-to-image generation models can be trained from scratch without relying on large-scale language model pretraining, significantly reducing computational resources.

• TokenGate: Adapting LLMs for Fast and Memory-Efficient Prefill with Learned Token Pruning

  This paper introduces TokenGate, a lightweight method that speeds up and reduces memory use in large language models by learning to skip unimportant tokens during processing, without losing accuracy.

  Ai Ml · relevance 9/10
• Parallel Language Models via Variational Decoding

  This paper introduces a new method to generate multiple tokens simultaneously in language models, significantly speeding up inference without sacrificing quality, especially for large models.

  Ai Ml · relevance 9/10

• Tight Inapproximability for Generalized Circuits

  This paper precisely determines the boundary between hard and easy cases for finding approximate solutions to generalized circuits, a key problem underlying many equilibrium computations, and uses these results to improve approximation algorithms in game theory and network models.

  Algorithms Theory · relevance 7/10
• FPT Isomorphism Test for $F$-Free Tournaments

  The paper presents a fixed-parameter tractable algorithm to test isomorphism between tournaments that exclude a fixed substructure, improving the efficiency of such tests for a broad class of directed graphs.

  Algorithms Theory · relevance 7/10
• The Complexity of Multiplayer Colonel Blotto Games with Player-Specific Values

  This paper shows that finding approximate Nash equilibria in multiplayer Colonel Blotto games with player-specific battlefield values is computationally hard (PPAD-hard), except in some simple cases, highlighting complexity barriers in strategic resource allocation problems.

  Algorithms Theory · relevance 7/10

• A Type System for Communication-Safe Hardware Design

  This paper introduces Pact, a type system that ensures hardware components communicate correctly and safely by verifying their interaction protocols before integration, helping catch errors early and supporting incremental hardware design.

  Compiler · relevance 8/10

• SARP: An On-SSD Index for KVCache with Fast Prefix Matching and Recovery

  This paper introduces SARP, a new on-SSD indexing method that enables fast prefix matching and rapid recovery for key-value caches in distributed large language model serving, significantly improving throughput and reducing downtime after failures.

  Database · relevance 8/10

• SafeWall: A Risk-Aware Job Walltime Recommendation for HPC Scheduling

  This paper introduces SafeWall, a method that improves HPC job runtime estimates by adding a risk-aware safety buffer to reduce premature job termination while maintaining prediction accuracy.

  Distributed Systems · relevance 8/10
• FaaSpot: Fault-Tolerant Serverless Workflow Execution on Spot Instances

  FaaSpot enables running serverless workflows entirely on cheaper spot instances by identifying critical tasks prone to failure and applying targeted fault-tolerance techniques, significantly reducing cost and latency without relying on expensive on-demand resources.

  Distributed Systems · relevance 8/10
• GeoSplit: Selective Edge--Server Map Acquisition for Cross-User Travel-Mode Recognition

  This paper introduces GeoSplit, a system that improves travel-mode recognition by selectively querying map data from servers, reducing communication costs while maintaining high accuracy.

  Distributed Systems · relevance 7/10

• Tight lower bounds for Lipschitz geodesically convex optimization on hyperbolic space

  This paper establishes the exact number of gradient queries needed to minimize Lipschitz geodesically convex functions on hyperbolic spaces, matching known upper bounds and introducing a novel proof technique that accounts for curvature effects.

  Math · relevance 7/10
• A boundary integral method for computing electron repulsion integrals in heterogeneous dielectric environments

  This paper introduces an efficient computational method to accurately calculate electron repulsion integrals in complex layered dielectric environments, which is important for modeling electronic properties of advanced materials like 2D graphene devices.

  Math · relevance 7/10
• Residual-Based Adaptive Refinement in Hybrid Variational-Collocation PINNs for Singular Elliptic PDEs with Mixed Dirichlet-Neumann Boundary Conditions

  This paper introduces a method to improve neural network solutions for challenging elliptic PDEs with singularities by adaptively refining training points based on residual errors, significantly boosting accuracy over uniform sampling.

  Math · relevance 7/10

• Connection Escrow: Holding AI-Agent Pod Egress in the Kernel Until an Isolated Verifier Rules

  This paper introduces a kernel-level method to intercept and hold outbound network connections from AI-agent containers until a separate verifier approves them, enhancing security without modifying the agents or breaking TLS encryption.

  Os · relevance 8/10

• Fake News, Real Crashes: Information Injection Attacks on LLM Financial Trading Agents

  This paper demonstrates that fabricated news can reliably manipulate large language model-based financial trading agents to make harmful trades, revealing a critical vulnerability in automated trading systems that rely on news feeds.

  Security · relevance 9/10
• Proof-of-Machine-Learning-Inference: Mining Knowledge, Not Hashes

  This paper proposes a blockchain consensus mechanism where miners perform useful machine learning inference tasks instead of wasteful hash computations, potentially making mining more efficient and beneficial.

  Security · relevance 9/10
• BrushOff: Stroke-Based Re-rendering as a Watermark Removal Attack on Painting

  This paper introduces a simple, training-free method to remove invisible watermarks from digital paintings by redrawing them with brush strokes, effectively defeating state-of-the-art watermarking techniques while preserving image quality.

  Security · relevance 8/10

• A 105 MS/s 4.18 nJ/inference 0.765 $\text{mm}^\text{2}$ Mixed-Signal Streaming ML Engine for Wireless Baseband in 12LP+ FinFET

  This paper introduces SIXPPAC, a compact and energy-efficient mixed-signal machine learning chip that processes wireless baseband signals in real time, enabling integrated sensing and communication with very low power consumption.

  Systems Architecture · relevance 8/10
• E-Layout: Layout-Stage Logic Restructuring via RL-Optimized Equality-Saturated Graph Retrieval-Augmented Generation

  This paper introduces E-Layout, a machine learning framework that improves integrated circuit layout by restructuring logic after placement, combining logic transformations with physical design feedback to enhance timing and reduce area.

  Systems Architecture · relevance 8/10