Search: “gpu-inference”

Smart search ranks by meaning as well as keywords (one row per story, last 45 days).

Implementation of Machine Learning Workflows with NVIDIA cuML, RAPIDS, GPU Benchmarking, Explainability, Clustering, and Model Inference

Hands-on tutorial implements NVIDIA cuML and RAPIDS to GPU-accelerate scikit-learn-style ML workflows with benchmarking, clustering, and inference.

The tutorial demonstrates NVIDIA cuML as a GPU-accelerated machine learning framework, using cuml.accel to speed up unmodified scikit-learn scripts with zero code changes and the native cuML API for CuPy/cuDF interoperability. It benchmarks CPU versus GPU implementations of PCA, K-Means, nearest-neighbor search, logistic regression, random forests, and DBSCAN on datasets up to 200,000 samples with 64 features. It also builds GPU pipelines with UMAP, t-SNE, and HDBSCAN, validates GPU-generated SHAP explanations, uses the FIL library for forest inference, and covers model serialization and GPU/CPU portability.

MarkTechPost·Sep 13, 01:42 UTC · 3d agoAI tools & infra

14

SpliTEE: Improving LLM Inference on Trusted Hardware with Differentially Private GPU Outsourcing

SpliTEE splits LLM inference between Intel TDX trusted execution and untrusted GPUs, using differential privacy instead of encryption to protect intermediate representations.

SpliTEE extends split inference to LLMs, running inference partly inside an Intel TDX TEE while masking intermediate inputs sent to untrusted GPUs with differential privacy rather than encryption. The authors show a prompt-reconstruction attack recovers nearly 80% of prompts from unmasked intermediate representations, motivating the masking. A global sensitivity analysis bounds the required DP noise scale, avoiding quantization and keeping models in floating point. The implementation is nearly twice as fast as full CPU-based TDX inference and 5-15 seconds faster than encryption-based Slalom with higher accuracy, evaluated on Llama-3.2-3B and Qwen3-4B.

arXiv cs.CR·Sep 14, 04:55 UTC · 2d agoResearch

34

Towards Standardized Evaluation of GPU Memory Safety with GMSBench

GMSBench provides 149 CUDA tests covering spatial, temporal, and concurrency GPU memory errors, exposing detection gaps in Compute Sanitizer.

GMSBench is a GPU memory safety benchmark comprising 149 self-contained CUDA tests spanning spatial, temporal, and concurrency errors across different GPU memory spaces and execution scenarios. The authors evaluate NVIDIA's Compute Sanitizer across multiple GPU architectures using the suite, exposing gaps in its detection coverage. The benchmark offers a standardized foundation for comparative evaluation of GPU memory safety mechanisms.

arXiv cs.CR·Sep 8, 15:11 UTC · 7d agoResearch

28

Speculative Decoding in vLLM on AMD GPUs

vLLM benchmarks speculative decoding on AMD Instinct MI300X and MI355X GPUs across five drafting methods including EAGLE-3 and native MTP.

The vLLM project documents draft-and-verify speculative decoding support for AMD GPUs via ROCm, comparing native MTP, Gemma 4 MTP, EAGLE-3, DFlash, and DSpark drafting approaches. Output-token throughput effects varied with drafting method, proposal length, model family, draft checkpoint, workload, and acceptance behavior. The post also covers how to enable each method plus practical tuning and observability considerations.

Hacker News · AI·Sep 7, 09:26 UTC · 9d agoAI tools & infraHN 141↑ · 52 comments1

40

OPEN-1B: A Fully Auditable Training Run

Open-1B releases a 1B-parameter model with bitwise-reproducible training, letting independent auditors verify every step of the run on commodity hardware.

The paper introduces a 'fully auditable' tier of model transparency: every training operation is reproducible with bitwise certainty on heterogeneous commodity hardware by imposing definite ordering on GPU kernel reductions, data batch ordering, and collective communication. Because replaying a full run on one machine is infeasible, a collective verification scheme lets many independent auditors certify individual steps covering the whole run. The authors release Open-1B with its full pretraining dataset, every intermediate checkpoint, the training codebase, and an audit harness. This rules out undisclosed data, injected biases, or backdoors that proof-of-learning or proof-of-training-data techniques cannot exclude.

arXiv cs.AI / cs.LG / cs.CL·Sep 15, 16:23 UTC · 22h agoAI research

58

Inside NVIDIA’s cuDNN Graph API: Fusion, Autotuning, and Plan Reuse with cuDNN Frontend

MarkTechPost tutorial walks through NVIDIA's cuDNN Frontend graph API, covering kernel fusion, autotuning, plan reuse, and CUDA graph capture on Colab GPUs.

The tutorial explains how to express GPU computations as operation graphs via the cuDNN Frontend graph API, running the five-step build pipeline of validate, build operation graph, create execution plans, check support, and build plans. It progresses from a single fused convolution with bias and ReLU to autotuning across engine configs, FP8-style epilogues, attention, plan serialization, dynamic shapes, and CUDA graph capture. Each kernel is benchmarked against a PyTorch reference on a single Colab GPU to verify correctness and measure cost. The piece also covers practical setup issues like making libcudnn.so visible to the frontend's dynamic loader.

MarkTechPost·Sep 15, 21:37 UTC · 17h agoAI tools & infra

18

GPU-CFR: 80x Faster Counterfactual Regret Minimization by Compiling the Game to Static Dataflow and CUDA Graph Replay

GPU-CFR compiles counterfactual regret minimization into static dataflow with CUDA Graph Replay, achieving 29.8-80.4x speedups over prior GPU solvers.

The paper presents a compiler and runtime that turns any fixed game's CFR iteration into a static dataflow graph of flat arrays and precomputed indices, cutting framework operations by up to 18.1x. Because shapes and buffer addresses never change, CUDA Graph Replay records the iteration once and replays it with a single launch. On one A100 across an eight-game suite, GPU-CFR runs 29.8-80.4x faster than the fastest prior GPU CFR and 14-258x faster than the CPU implementation LiteEFG on the four largest games, while reproducing reference iterates bitwise on CPU.

arXiv cs.AI / cs.LG / cs.CL·Sep 10, 17:58 UTC · 5d agoAI research3· 1 read

38

NVIDIA Details BioNeMo Inference Runtime (BioIR): 2.90x Higher Boltz-2 Folding Throughput and 58.5K Residues per GPU-Hour on 8xH100

NVIDIA released BioNeMo Inference Runtime (BioIR), an open-source PyTorch-compatible library delivering 2.90x higher Boltz-2 protein-folding throughput on 8xH100 GPUs.

NVIDIA detailed BioIR, a Python library that accelerates Boltz-2, OpenFold2, and OpenFold3 structure-prediction inference on NVIDIA GPUs while preserving standard PyTorch workflows. On a matched benchmark of 1,000 human dimers on 8xH100 80GB GPUs, BioIR delivered 58.5K folded residues per GPU-hour versus 20.2K for a torch.compile baseline, a 2.90x throughput gain. BioIR already powered the AlphaFold Database expansion, generating about 31 million candidate complexes across 4,777 proteomes, with 1.81 million released as high-confidence predictions. Extrapolated to 1 million targets, estimated folding energy drops from 35 MWh to 11 MWh at 8-GPU TDP equivalents.

MarkTechPost·Sep 10, 21:57 UTC · 5d agoAI tools & infra1

25

Introducing @huggingface/kernels: 200+ WebGPU Kernels for Local AI

Hugging Face released @huggingface/kernels, a library offering 200+ WebGPU compute kernels to accelerate AI inference locally in browsers.

Hugging Face introduced the @huggingface/kernels package, bundling more than 200 optimized WebGPU compute kernels for running AI workloads locally. The release targets browser-based and on-device inference, reducing reliance on server-side compute. No article body was available beyond the title, so benchmark results and supported models are not specified.

Hugging Face Blog·Sep 1, 00:00 UTC · 15d agoAI tools & infra

32

Online Draft Co-Training for Speculative Decoding in Large-Scale, Long-Context RL Post-Training

NVIDIA researchers detail an end-to-end system for online draft co-training that speeds speculative decoding in large-scale long-context RL post-training.

The paper tackles scaling online draft co-training for speculative decoding in RL post-training, where rollout generation dominates cost. It extends packed, load-balanced zigzag ring attention to merge rank-local branch attention with causal main-sequence attention for context parallelism, and introduces TapChannel to transport target features across pipeline-parallel stages without changing the schedule. Experiments show co-trained drafts tracking the policy baseline with substantial rollout and end-to-end speedups up to 122B parameters and strong scaling at 256K tokens.

Hugging Face daily papers·Sep 6, 20:00 UTC · 9d agoAI research

26

Scalability Analysis of Distributed Kolmogorov-Arnold Network Training on High-Performance Computing Systems

An empirical study shows distributed Kolmogorov-Arnold Network training reaches 74.7% parallel efficiency at 8 A100 GPUs, with overheads driven by All-Reduce choices.

The study evaluates data-parallel Kolmogorov-Arnold Network (KAN) training on the FinisTerrae III supercomputer using up to 8 NVIDIA A100 GPUs across 4 nodes with PyTorch Distributed Data Parallel. Strong scaling yields 5.97x speedup and 74.7% parallel efficiency at 8 GPUs, comparable to conventional deep learning workloads, while communication overhead ranges from 1.3% to 6.1%, driven mainly by All-Reduce algorithm selection and inter-node latency rather than KAN's edge-wise gradient structure. Weak scaling shows an initial single-to-multi-GPU throughput drop followed by stability, and the parameter-to-memory ratio improves with model size even as training time scales unfavorably. The authors provide GPU topology and model-size deployment guidelines for KAN training.

arXiv cs.AI / cs.LG / cs.CL·Sep 7, 16:45 UTC · 8d agoAI research1

26

Perplexity Details Its GPU Embedding Stack: How Ivy, Tulip and ROSE Serve pplx-embed

Perplexity details its GPU embedding serving stack (Ivy, Tulip, ROSE), which reuses LLM prefill/decode kernels, CUDA graphs, and LazyTensors to cut launch overhead.

Perplexity engineers published a deep dive on the serving infrastructure behind pplx-embed, used across Perplexity Search and its API platform. The stack comprises Ivy (Rust HTTP gateway), Tulip (gRPC scheduling and batching), and ROSE (Runtime-Optimized Serving Engine), which reuses LLM prefill and decode kernels rather than running a separate embedding engine. Optimizations include whole-model CUDA graphs with lazy capture and a LazyTensor abstraction that overlaps CPU batch preparation with in-flight GPU work. Benchmarks are reported against vLLM v0.22.0 in BF16, with FlashAttention 4 generally fastest but FlashInfer 3 winning on Qwen-based models at very long sequence lengths.

MarkTechPost·Sep 6, 03:20 UTC · 10d agoAI tools & infra1

33

GPUThor: Amplifying Rowhammer Attacks via Non-Uniform Patterns to Exploit ECC-Protected GPUs

GPUThor uses non-uniform hammering to amplify Rowhammer on NVIDIA GPUs, achieving 500X-23,500X more bit flips and first exploits of ECC-protected GPUs.

GPUThor is a Rowhammer attack on NVIDIA GPUs that reverse-engineers memory-access coalescing behavior to enable non-uniform hammering patterns activating aggressor rows more intensely than decoy rows. By identifying refresh instances where in-DRAM mitigations apply, it constructs longer patterns that escape mitigation across refresh intervals. It yields 500X to 23,500X more bit flips than prior GPU Rowhammer attacks across NVIDIA A4000, A4500, A5000, and A6000 GPUs, and enables the first Rowhammer exploits on ECC-protected GPUs via uncorrectable double and triple bit flips, making denial-of-service and privilege-escalation attacks practical.

arXiv cs.CR·Sep 15, 02:46 UTC · 1d agoResearch

66

With Groq 3 LPX in Full Production, NVIDIA Extends Vera Rubin Inference for Agents

NVIDIA puts Groq 3 LPX into full production and extends Vera Rubin NVL72 rack-scale systems for fast token generation in agentic AI inference.

NVIDIA announced that Groq 3 LPX is in full production as part of an extension of the Vera Rubin NVL72 rack-scale platform aimed at agentic AI inference. The announcement frames the next era of inference as full-stack AI factory co-design across chips, networking, and systems rather than a single component breakthrough. The focus is improving token generation speed for agent workloads.

NVIDIA Blog·Aug 24, 15:00 UTC · 22d agoAI industry

62

Miles v0.1: Production-Level Post-Training

Radix Ark open-sources Miles v0.1, a full-stack RL post-training framework demonstrated with asynchronous agentic RL on GLM-5.2 744B-A40B across 64 GB300 GPUs.

Miles v0.1 is a full-stack, open-source system for frontier-scale reinforcement-learning post-training, built on slime with rollout engines on SGLang and trainers supporting NVIDIA Megatron-LM and PyTorch FSDP backends plus three weight-synchronization transports. It supports full-parameter RL, LoRA RL, on-policy distillation, supervised fine-tuning, true-on-policy rollout-training alignment, and extends to diffusion models. The end-to-end case study ran fully asynchronous agentic RL on GLM-5.2 744B-A40B for terminal-use coding tasks on 64 NVIDIA GB300 GPUs with a median step time of 263 seconds over the first 30 measured steps. The code is open-sourced on GitHub.

Hugging Face daily papers·Sep 7, 20:00 UTC · 8d agoAI tools & infra

45

Why Scaling AI Compute Performance Requires a New Power Architecture

NVIDIA argues AI factories need 800 VDC power distribution as dense GPU racks outgrow traditional AC-based delivery.

NVIDIA's blog contends each generation of accelerated computing demands higher rack density and more efficient, scalable power distribution. It frames the bottleneck as how power moves from the grid to the GPU rather than raw wattage, and describes limitations of traditional AC power delivery. NVIDIA advocates a new 800 VDC power architecture for AI factories.

NVIDIA Blog·Aug 11, 15:00 UTC · Aug 11, 2026AI industry

18

An Open-Source End-to-End FHE Implementation for Privacy-Preserving Llama 3 8B Inference

Odin runs Llama-3-8B fully homomorphic encrypted inference on a single H100 in 366 seconds, a 4.51x speedup over THOR.

Odin is an open-source end-to-end GPU CKKS implementation for privacy-preserving Llama-3-8B inference that co-designs ciphertext packing with model execution. A feature-major cross-layer layout unifies residual connections and layer interfaces, while transient intra-operator layouts serve linear projections and attention, avoiding intermediate repacking of QK^T softmax outputs. Minimax polynomial approximation with input-range control reduces polynomial degree and multiplicative depth for nonlinear ops. With 128-token input, Odin evaluates all 32 Transformer layers on one NVIDIA H100 80 GB in 366.4 s using 58.9 GiB peak memory, versus 1651.9 s for the THOR baseline, a 4.51x speedup.

arXiv cs.CR·Sep 11, 02:51 UTC · 5d agoResearch

48

Unfold The World: Factorize 4D Properties in Reinforcing Spatial Reasoning

FactoSR factorizes 4D spatial reasoning into XY, Z, and T reinforcement-learning sub-objectives, boosting VLM performance on VSI-Bench by 5.9% and All-Angles-Bench by 4.5%.

Researchers present FactoSR, a factorized reinforcement learning framework that decomposes world-consistent reasoning into planar correspondence, depth consistency, and temporal reversibility sub-objectives. Optimizing these verifiable constraints turns the ill-posed projection recovery problem into tangible reasoning steps. Evaluations show gains of 5.9% on VSI-Bench and 4.5% on All-Angles-Bench for 3D and 4D reasoning, arguing VLMs' spatial bottleneck stems from training on 2D projections versus latent 3D geometry and temporal continuity.

Hugging Face daily papers·Sep 2, 20:00 UTC · 13d agoAI research

26

BeaconKV: Key-Value Cache Compression Guided by Beacon Queries for Efficient Large Reasoning Model Inference

BeaconKV introduces training-free KV cache compression using beacon queries, cutting long-reasoning inference memory up to 5.8x while preserving accuracy.

The paper shows recency-based KV cache compression assumptions fail in long-horizon reasoning because Thought Revisiting Tokens (TRT) re-attend to distant context such as early task-solving plans. TRT queries cluster into a small number of similarity groups, which BeaconKV exploits by maintaining compact beacon query representatives to anticipate revisited KV pairs without storing full query history. The training-free method achieves up to 5.8x memory reduction and over 4.3x throughput improvement across four open-source large reasoning models while nearly preserving full cache accuracy.

Hugging Face daily papers·Sep 3, 20:00 UTC · 12d agoAI research1

33

Rethinking Heterogeneous System Disaggregation for Subquadratic Attention

SQD disaggregates LLM inference by quadratic versus subquadratic attention layers, improving energy efficiency up to 56% on heterogeneous systems versus GPU-only baselines.

SQD (SubQuadratic Disaggregation) splits decode not by operator type but by quadratic versus subquadratic attention, matching their distinct arithmetic intensity and memory footprints. For sparse attention LLMs it separates top-k selection (requiring full KV indexing) from top-k attention plus FFN; for linear and sliding-window models it separates dense attention layers from subquadratic layers plus FFN. On an adjusted 8xB200 heterogeneous proxy, tokens-per-joule improves 53% on GLM 5.2, 31% on Nemotron 3 Ultra, and 56% on Gemma 4 31B. A Rubin plus LPX analytical model shows 1.2x-1.5x tighter achievable latencies and up to 3.6x higher throughput versus attention-FFN disaggregation.

arXiv cs.AI / cs.LG / cs.CL·Sep 11, 17:55 UTC · 4d agoAI research

28

Arm Mali G2-Ultra NX GPU: desktop-class mobile gameplay with AI-native graphics

Arm unveiled Mali G2-Ultra NX, its first AI-native mobile GPU with in-shader neural acceleration, third-gen ray tracing, and up to 24% higher benchmark performance.

Arm announced the Mali G2-Ultra NX, the first AI-native Mali GPU, integrating neural accelerators directly into shader cores alongside a new execution engine and third-generation hardware ray tracing. It introduces Neural Super Sampling (NSS), Neural Frame Rate Upscaling (NFRU), and Neural Super Sampling and Denoising (NSSD); the Neural Dawn demo with Sumo Digital showed up to 4x performance efficiency and 70% lower external memory traffic versus native rendering. Arm claims up to 24% higher benchmark performance, 13% lower DRAM traffic on ray tracing benchmarks, and up to 120 FPS with NFRU. Over 14 billion Mali GPUs have shipped to date.

Hacker News · AI·Sep 8, 04:06 UTC · 8d agoAI industryHN 71↑ · 53 comments1

45

ConvMem: Convolutional Memory for Long-Context Reasoning

Researchers propose ConvMem, a training-free framework treating LLMs as convolutional kernels for parallelizable long-context reasoning beyond fixed context windows.

ConvMem reformulates long-context reasoning as a hierarchical convolution in which the LLM summarizes text segments hierarchically, shortening the reasoning path from a linear chain to a logarithmic tree. It uses configurable strides, skip connections, and multi-kernel convolution to capture evidence, decompose queries, and enable massive parallelization across segments and reasoning threads. On RULER-HotpotQA and RULER-2WikiMultiHopQA it outperforms training-free baselines and avoids the out-of-distribution overfitting seen in RL-trained approaches like MemAgent.

arXiv cs.AI / cs.LG / cs.CL·Sep 9, 16:53 UTC · 6d agoAI research

38

OracleZoom: On-Policy Self-Distillation Inspired Reference-Constrained Recursive Image Super Resolution

OracleZoom enables recursive extreme-scale image super-resolution via reference-constrained on-policy distillation, reducing hallucinations at deep zoom scales.

OracleZoom tackles recursive super-resolution, where repeated feeding of predictions back into the same model leaves deeper-scale outputs unsupervised as required source resolution grows geometrically. The framework trains on its own trajectory while carrying the last ground-truth evidence beyond the supervision boundary, combining direct and cross-scale supervision, a no-reference quality objective, a KL-constrained pretrained latent prior, and EMA consistency. Across seven datasets it achieves state-of-the-art SR quality across zoom scales, averaging 0.713 CLIPIQA with larger gains at deeper scales and significantly reduced hallucinations. Code, data, and models are publicly released.

Hugging Face daily papers·Sep 5, 20:00 UTC · 10d agoAI research

20

Hoss: Fast Oblivious Semantic Search with Heterogeneous GPU-CPU-TEE Architecture

Hoss uses heterogeneous GPU-CPU TEEs for oblivious semantic search, achieving up to 67x speedup over the Compass ORAM-based system.

Hoss is a first-of-its-kind oblivious semantic search system built on a heterogeneous CPU-GPU TEE architecture, hosting hot-path HNSW graph traversal in large GPU TEE private memory and offloading lower graph layers to CPU TEEs. It invokes oblivious primitives only for lower-layer accesses and adds host-access ORAM and data-dependent optimizations. Benchmarked against prior state-of-the-art Compass, the prototype achieves up to 67x speedup while maintaining high recall, with larger gains at scale.

arXiv cs.CR·Sep 3, 22:21 UTC · 12d agoResearch

25

Why Is Video Still So Expensive? A Survey of Inference-Efficiency Mechanisms in Video and Audiovisual LLMs

A survey catalogs inference-efficiency techniques for video and audiovisual LLMs, mapping bottlenecks in sampling, encoding, token reduction, and LLM decoding.

This survey covers inference-efficiency mechanisms for visual and audiovisual video LLMs, reporting reductions in parameters, FLOPs, latency, memory, and token counts. It organizes methods by pipeline stage, covering frame sampling, modality encoding, connector-level token reduction, and LLM prefilling and decoding for systems built since late 2022. The authors compile accuracy-cost comparisons under shared host models and input protocols, identify gaps in audiovisual efficiency and standardized evaluation, and maintain a public repository.

Hugging Face daily papers·Sep 8, 20:00 UTC · 7d agoAI research

40

Foundation Models for Generalizable Semantic and Goal-Oriented Communication

FMSGOC uses vision-language foundation model priors plus diffusion reconstruction to enable generalizable semantic communication at 0.039 bits per pixel for 6G.

FMSGOC targets generalization failures in semantic and goal-oriented communication for 6G by leveraging broad visual-linguistic foundation model priors. A vision-language model selects sparse, goal-aligned semantic anchors while a fine-tuned diffusion model performs masked completion to reconstruct images at the receiver, decoupling what to send from how to reconstruct. On CIFAR-10 it reaches 0.039 bits per pixel with cosine similarity 0.87-0.90 and 0.83-0.86 on unseen ImageNet inputs, outperforming end-to-end baselines at lower bit rates.

arXiv cs.AI / cs.LG / cs.CL·Sep 7, 18:12 UTC · 8d agoAI research

22

Understanding the Impact of Model Pruning on Long-Tail Forgetting and Explanation Reliability in Medical Imaging

Systematic study finds model pruning causes frequency-dependent long-tail forgetting in medical imaging and that gradient-informed methods best preserve explanations.

Across two long-tailed medical imaging datasets, two CNN architectures, four pruning methods, and sparsity up to 95%, the study measures predictive performance, explanation stability, and faithfulness. Rare classes degrade earlier and more severely than frequent ones, while explanation reliability depends mainly on the pruning strategy, with gradient-informed methods degrading least. Mechanistic analysis ties explanation collapse to loss of class-discriminative gradients rather than vanishing feature activations, recommending class- and explanation-aware evaluation of compression.

arXiv cs.AI / cs.LG / cs.CL·Sep 7, 17:44 UTC · 8d agoAI research

27

FlashVector: Agent for Hierarchical Model Serving Stack Optimization

FlashVector agent optimizes all layers of Unity's ad-serving stack, delivering up to 2x model-server throughput and 1.98x latency speedup in production.

FlashVector is an agentic system that optimizes performance across GPU kernels, ML framework computation graphs, model servers, and on-demand feature processing. Deployed in Unity's Vector advertising platform, it achieved up to 2x model-server throughput increase, 1.98x latency speedup, and 1.6x feature-store throughput gain. Optimizations spanned NVIDIA Triton's C++ codebase and the Python feature transformation service, demonstrating extensibility beyond single-kernel tuning.

arXiv cs.AI / cs.LG / cs.CL·Sep 15, 16:27 UTC · 22h agoAI research

30

OptiPrime: Optimizing Private Inference through Protocol-Hardware Co-design

OptiPrime co-designs HE-MPC protocols with hardware acceleration to remove network communication bottlenecks in private DNN inference, beating Cheetah by up to 5.7x.

OptiPrime is a protocol-hardware co-optimization framework for private deep neural network inference based on hybrid homomorphic encryption (HE) and multi-party computation (MPC). It introduces a novel HE protocol for convolutions that reduces the number of transmitted output ciphertexts, addressing the network bottleneck that limits gains from commercial HE accelerators. A lightweight compression system reduces weight plaintext memory traffic by 10x, while a specialized dataflow maximizes on-chip reuse of intermediate ciphertexts. Experiments show up to 5.7x speedup over the Cheetah baseline on CPUs and 4.2x with an accelerator.

arXiv cs.CR·Sep 15, 09:29 UTC · 1d agoResearch

28

SpecGuard: Inference-Time Backdoor Detection For Free

SpecGuard detects backdoored LLM behavior at inference time using speculative decoding acceptance rates, adding no extra model computation.

Researchers propose SpecGuard, an inference-time backdoor detector that repurposes draft-token acceptance rates from speculative decoding as a detection signal at zero added model-computation cost. When a trigger shifts the target model toward attacker-controlled behavior, the clean draft model's acceptance rate changes, exposing the backdoor; the paper formalizes when this signal appears and shows suppressing it weakens the backdoor. Experiments across diverse backdoor types and model families show reliable detection, including stealthy cases invisible to input-level filters. Speculative decoding is positioned as a free, always-on monitor for frequently updated deployed models.

arXiv cs.CR·Sep 10, 16:51 UTC · 5d agoAI safety & security 2 sources2

45