This week the research centres on making large language model systems faster and more predictable. One key idea is parameter-efficient conversion of autoregressive models into diffusion language models alongside the trade-offs of speculative decoding for mixture-of-experts models. The other is that context in agent harnesses behaves as a pipeline where information degrades, a theme that connects to broader work on quantisation and experiment modelling.
Deduplicating experts decides when speculative decoding speeds mixture of experts
Speculative decoding promises fewer model passes by verifying a batch of drafted tokens at once. For mixture of experts models the cost of that verification depends on how experts are loaded. Consecutive tokens often pick the same experts, so a deduplicating engine can fetch each distinct expert only once per pass while a per slot engine fetches the same expert again for every token that selects it. The analysis frames a step as bytes requested by each kernel divided by its effective bandwidth plus fixed overhead, then assumes geometric acceptance to ask when drafting actually saves time. That framing gives builders a way to pick draft length from memory traffic and bandwidth rather than guesswork, and it clarifies when an inference stack should deduplicate experts to make speculation pay.
Two bit quantization should be judged on reasoning not perplexity
Reasoning language models often emit thousands of tokens per answer, and that long generation is bottlenecked by memory bandwidth. Cutting weight precision from bfloat16 to two bits could in theory make inference eight times cheaper. The problem is that standard checks for two bit quantization lean on perplexity and multiple choice question answering, which reveal little about harder step by step reasoning. ThinkBit responds with a strictly scalar two bit framework designed specifically to keep reasoning intact. For product teams the takeaway is to price ultra low bit models on reasoning benchmarks rather than perplexity alone.
The daily brief
Get the free daily brief with the day's AI news, new AI startups and funding rounds every morning.
Join 64,000 other subscribers.
World model predicts machine learning experiments without running code
AI research agents that write machine learning code quickly hit a bottleneck because each candidate program must be trained and evaluated before the next idea can be judged. That execution cost caps how many variants an agent can try in one loop. MLE-World offers a shortcut by learning to forecast what would happen without running anything. Fine tuned from Qwen3.5-4B, the world model takes code plus a data description and predicts the terminal output including error traces, training logs and validation scores, along with a confidence score. If reliable, that kind of preview lets agent harnesses filter broken or weak candidates early and spend compute only on promising runs.
A third of tool outputs are lost before reaching the model
Large language model agents are scored on context window size and final answer accuracy, which hides how much work disappears inside the harness. A study instrumented a tool using agent that diagnoses injected faults from Linux kernel traces. Across 720 runs the largest prompt filled only 5.75 percent of the context window, so window size was not the limit. Even so 3,827 of 11,189 tool results were cut before the model saw them. That attrition suggests builders should audit the pipeline that forwards tool output, not just the window.
Low rank adapters alone turn autoregressive models into diffusion models
Build Your Own DLM addresses a cost problem in diffusion language models. Diffusion language models allow parallel token updates, flexible generation and bidirectional attention, but current leading systems depend on expensive pretraining or full-model fine-tuning. The authors present BYOD, a generalizable recipe that uses only low rank adapters for parameter-efficient conversion from autoregressive to diffusion generation. Without architecture-specific tuning, BYOD induces diffusion generation in Gemma 2 9B, Llama 3.1 8B, Qwen2.5 7B and Ministral 8B. For builders this points to a cheaper path to diffusion-style inference without retraining large base models from scratch.
