The paper demonstrates that greedy decoding from large language models is not precision‑invariant: the same model, prompt, and decoding algorithm can produce different outputs when run in BF16 versus FP16 on identical hardware. Across six models (1.1B–7B parameters, four families, and 12B) and three benchmarks, 49–100 % of prompts diverge, with a single token flip often cascading into trajectory‑level divergence. The authors develop an empirical error‑propagation analysis that identifies the top‑two logit margin at the LM head as the key factor, and they propose a low‑overhead intervention—selective FP32 LM head recomputation—that improves exact agreement by 22–36 percentage points with less than 4 % latency overhead.
"whyItMatters":"The findings reveal that precision choices can fundamentally alter model outputs, challenging the assumption of deterministic greedy decoding and highlighting the need for precision‑aware inference strategies."
By Gaoyuan Du, Anam Nawaz Khan, Rex Zhou, Xiaoyang Liu, Deepayan Chakrabarti, Fnu Suya, Xueping Li
Prefill‑only decision models evaluate every candidate in a menu in a single forward pass, avoiding decoding and reducing cost by one to two orders of magnitude compared to generative language models. The paper demonstrates that when only the candidate menu changes, the model’s post‑intervention accuracy can be predicted solely from the cached first‑pass distribution using a simple estimator that renormalizes and selects the argmax, without any labels or second pass. Across seven model families, ten datasets, and two task types, this menu‑only intervention prediction is within 4.2 points of actual accuracy, and in one family it is exact, whereas a probability‑level variant fails by 21 points, indicating the property resides in ranking rather than calibrated probabilities.
By Ran Li, Lei Chen
arXiv:2608. 12385v2 Announce Type: replace Abstract: As large language models serve ever more requests, cumulative inference cost is growing relative to the one-time cost of training.
By Liming Liu, Mingze Wang, Tuo Zhao
The paper introduces PILL, a new infilling technique for diffusion language models that eliminates the need for a preset initial length and reduces inference overhead. PILL uses probing-based length-free decoding, cutting down on extra forward passes and speeding up generation. Experiments across five diffusion models and eight benchmarks show PILL outperforms the strongest baseline with higher pass rates and BLEU-2 scores while running 1.82× faster.
By Haobo Xu, Sirui Chen, Yuanchen Bei, Lingjie Chen, Yuchen Yan, Dongqi Fu, Jingrui He, Hanghang Tong
arXiv:2602. 15327v2 Announce Type: replace-cross Abstract: Machine learning model performance improvements tend to arise from competition and application.
By Hanlin Zhang, Jikai Jin, Vasilis Syrgkanis, Sham Kakade
The paper introduces PACE-dLLM, an acceleration method for diffusion language models (dLLMs) that uses the model’s own per‑step confidence to estimate a ‘confidence cliff’ and determine the optimal look‑ahead horizon for block decoding. By fitting this cliff in closed form at each step, PACE-dLLM sets the horizon to its saturation point and applies an independent confidence threshold for token commitment, thereby avoiding the trade‑offs inherent in fixed‑size block decoding. Experiments on reasoning and code benchmarks show that PACE-dLLM achieves the best average accuracy on open‑source dLLM backbones while delivering significant wall‑clock speedups—up to 5.23× on LLaDA and 3.06× on Dream—improving the quality‑throughput Pareto frontier.
By Xiaocheng Lu, Shuhan Guo, Ziyue Ma, Jie Zhang, Jian Liu, Jingcai Guo, Haoxuan Che, Song Guo
arXiv:2606. 11552v1 Announce Type: cross Abstract: Large language models (LLMs) achieve remarkable performance across a wide range of tasks, but their autoregressive decoding process incurs substantial inference costs due to inherently sequential token generation.
By Lexington Whalen, Yuki Ito, Ryo Sakamoto
arXiv:2609.38806v1 Announce Type: cross
Abstract: Next-token prediction has driven remarkable progress in large language models, yet a growing body of evidence suggests that they can struggle on prob...
By Woosang Jeon, Jaeyeon Kim, Sham Kakade, Yilun Du, Amrit Singh Bedi, Arun Kumar Chithanar, Chul Lee, Taehyeong Kim, Sitan Chen
ResiSpec is a framework that improves speculative decoding for large language models by reshaping the residual distribution during verification. It addresses the problem of residual drift, where rejected candidates cause the target distribution to diverge from the draft model’s predictions, rendering later candidates ineffective. By aligning the verification process with the draft model’s high‑confidence regions, ResiSpec prevents candidate obsolescence and achieves up to 1.92× speedup over existing multi‑candidate methods.
By Zhi-Kai Chen, Jun-Jie Tao, Wei-Xiang Mao, De-Chuan Zhan, Han-Jia Ye
arXiv:2605. 04215v3 Announce Type: replace-cross Abstract: Diffusion-based Large Language Models (D-LLMs) represent a promising frontier in generative AI, offering fully parallel token generation that can lead to significant throughput advantages and superior GPU utilization over the traditional autoregressive paradigm.
By Michael Rottoli, Subhankar Roy, Stefano Paraboschi
The paper introduces the Expected Decoding Rounds (EDR) objective, a theoretical framework that models speculative decoding as a Markov reward process and directly optimizes the number of decoding rounds. EDR weights local rejection costs by state occupancies, eliminating auxiliary hyperparameters and enabling unbiased stochastic optimization via temporal‑difference gradients. Applying EDR to fine‑tune state‑of‑the‑art draft models (DSpark and DFly) improves mean accepted length and outperforms prior training objectives across nine benchmarks in math reasoning, code generation, and chat.
By Yunxiao Zhao, Changxiao Cai
The paper introduces HARP, a hard-budget allocation method for evaluating large language models (LLMs) in multi-turn interactions where the time-to-event is partially observed due to resource limits. HARP guarantees that the total computational budget is never exceeded, reallocates unused budget, and provides lower predictive bounds (LPBs) with finite-sample coverage and unbiased metric estimates. Experiments on tasks such as jailbreaks, toxic content, and hallucinations demonstrate that HARP achieves near-nominal coverage with low variance while respecting the fixed budget.
By Shai Feldman, Yaniv Romano