LIVE PULSE
4.0 Anthropic CEO Amodei calls for slower AI development and shared safety rules — 11 src2.2 Agility Robotics unveils Digit 5 humanoid for warehouses and factories — 2 src2.0 Apple ships rebuilt Siri with Google Gemini, but not in the EU — 2 src1.8 Siri AI in macOS 27 Golden Gate: FAQ, Germany availability, privacy questions — 2 src1.4 Sam Altman says OpenAI will not go public in 2026, citing AI safety concerns — 5 src1.1 OpenAI contractors review real ChatGPT conversations to rate responses, report says — 2 src1.1 Anthropic data retention policy prompts firms to limit Claude use for sensitive work — 1 src1.1 Study examines issue bias in LLMs used as writing assistants before Swedish 2026 election — 1 src1.1 Study Audits Misalignment in Multi-Modal World Models — 1 src1.1 Retrieval-Grounded Reasoning Approach Proposed for Universal Multimodal Embeddings — 1 src4.0 Anthropic CEO Amodei calls for slower AI development and shared safety rules — 11 src2.2 Agility Robotics unveils Digit 5 humanoid for warehouses and factories — 2 src2.0 Apple ships rebuilt Siri with Google Gemini, but not in the EU — 2 src1.8 Siri AI in macOS 27 Golden Gate: FAQ, Germany availability, privacy questions — 2 src1.4 Sam Altman says OpenAI will not go public in 2026, citing AI safety concerns — 5 src1.1 OpenAI contractors review real ChatGPT conversations to rate responses, report says — 2 src1.1 Anthropic data retention policy prompts firms to limit Claude use for sensitive work — 1 src1.1 Study examines issue bias in LLMs used as writing assistants before Swedish 2026 election — 1 src1.1 Study Audits Misalignment in Multi-Modal World Models — 1 src1.1 Retrieval-Grounded Reasoning Approach Proposed for Universal Multimodal Embeddings — 1 src
HEATPULSEAI MAGAZINES
FLIP · FOLLOW · SAVE
Log inSign up

speculative-decoding

topic6 events
papersTODAY 04:00 UTC

Study examines numerical precision limits in Orthrus lossless speculative decoding

A new arXiv paper investigates whether speculative decoding with the Orthrus architecture remains truly lossless when numerical precision is taken into account. Orthrus is a hybrid autoregressive-diffusion system that drafts several tokens at once and verifies them with a frozen autoregressive model, and its claimed exactness depends on the draft and verification steps producing identical results. The work analyzes how floating-point rounding in these computations can break that equivalence in practice.

speculative-decoding/research/inference/numerical-precision
papersTODAY 04:00 UTC

Paper Proposes Reusing Rejected Tokens to Speed Up Speculative Decoding

A new arXiv paper introduces a technique called carryover drafting that reuses the model states computed for rejected draft tokens during speculative decoding. Because verification already produces representations for both accepted and rejected tokens, the method puts the discarded ones to further use instead of throwing them away. The goal is to improve the efficiency of LLM inference, where speculative decoding lets one target forward pass accept multiple drafted tokens at once.

speculative-decoding/llm-inference/efficiency/arxiv
papersTODAY 04:00 UTC

AdaFlash: Adaptive Speculative Decoding with On-Policy Distilled Diffusion Drafters

A new arXiv paper proposes AdaFlash, a speculative decoding method that uses diffusion-based draft models distilled on-policy to speed up large language model inference. The approach adapts the drafting process rather than relying on a fixed draft model, aiming to improve acceptance rates during verification by the target model. It builds on prior work in this line, including DFlash, and appears as a revised submission.

speculative-decoding/diffusion/inference/research
papersSEP 10 04:00 UTC

Osprey: Target-Agnostic Pre-training Builds Stronger Draft Models for Speculative Decoding

Researchers introduce Osprey, a pre-training approach for draft models used in speculative decoding that is not tied to any specific target model. Draft models are typically tuned to a single target's output distribution, causing acceptance rates to fall when workloads change, and target-agnostic pre-training aims to keep them robust. The work focuses on achieving faster and more stable inference for large language models.

speculative-decoding/llm-inference/efficiency/research
papersSEP 10 04:00 UTC

X-CoSD: Cross-Vocabulary Collaborative Speculative Decoding for Efficient LLM Inference

Researchers introduce X-CoSD, a distributed inference framework in which a small on-device model drafts tokens and a larger server-side LLM verifies them. The method tackles the mismatch between the two models' vocabularies while cutting the communication cost between device and server. The work aims to make collaborative speculative decoding practical for latency-sensitive edge deployments.

inference/edge-ai/efficiency/distributed-systems
papersSEP 10 04:00 UTC

PELM paper proposes speculative decoding and DVFS for power-efficient on-device LLM inference

A new arXiv paper introduces PELM, a system aimed at making large language model inference more power-efficient on mobile devices. The approach combines speculative decoding with dynamic voltage and frequency scaling to reduce the energy demands of running LLMs at the edge, where privacy, personalization, and lower latency are key motivations.

on-device/edge-ai/efficiency/research