Papers for
llm deployment teams
Papers whose findings have a practical use for this group, as judged from the abstract. Open a paper to read what it means in practice.
Microscaling cuts hardware costs in large language model acceleration
MIAO: Microscaling Intermediate Accumulation Optimization
Abstract: Microscaling (MX) quantization enables lower-precision large language model (LLM) inference with limited accuracy degradation, reducing memory requirement and multiplication cost. However, despite reduced operand precision, accumulation is often retained at high floating-point precisions, leaving a substantial hardware cost that operand quantization alone does not address. In particular, systolic arrays, as used in tensor accelerators such as Google TPU and AWS Trainium, distribute accumulation across distinct processing elements but conventionally use uniform accumulator precision throughout the array. This spatial organization offers an opportunity to reduce hardware cost by tailoring precision to individual accumulation stages. In this work, we present MIAO, a hardware-software co-design framework that jointly selects product precision and stage-dependent accumulator widths to minimize hardware area and energy cost under an end-to-end perplexity constraint. MIAO first calibrates MX block-scale shifts to reduce intermediate magnitudes while preserving operand quantization accuracy, then assigns narrower accumulators to earlier stages and progressively wider accumulators to later stages. To efficiently explore the resulting design space, MIAO combines hardware cost estimates derived from logic synthesis with a calibrated accuracy predictor to identify promising precision configurations, then validates the accuracy of promising candidates through bit-exact end-to-end LLM evaluation. Across MXFP8, MXFP6, and MXFP4, MIAO reduces processing element (PE) array area by up to 38.7% and energy by up to 36.1% relative to the all-BF16 accumulation baseline, with at most 1.60% additional perplexity degradation over operand-only MX quantization across six LLMs.
Circe uses model internals to check large language model answers
Circe: Training-Free LLM Self-Evaluation via Component Interaction Consistency
Abstract: As large language models (LLMs) are increasingly deployed in real-world applications, evaluating the correctness of their responses is crucial for reliable use. However, such evaluation is challenging because reference answers are rarely available in practice. Existing reference-free methods mainly rely on output signals, which may assign high scores to incorrect responses and thus do not necessarily reflect correctness. This motivates exploring the model's internal computation for complementary correctness signals. Recent studies exploit internal states, but they mainly analyze hidden-state trajectories across layers, which record only the combined updates of attention and feed-forward networks (FFNs) and overlook how these components interact. We therefore propose Circe, a training-free self-evaluation method that assesses a response using both internal and external signals, without reference answers or repeated sampling. For internal signals, we first introduce component interaction consistency (CIC), which measures the directional alignment between attention updates in earlier layers and attention-induced FFN output changes in later layers. We further incorporate representation geometry (RG) to characterize normalized displacement and angular changes along layerwise hidden-state trajectories. For the external signal, we convert the mean negative log-likelihood of generated tokens into a bounded likelihood confidence (LC) score. Finally, we fuse CIC, RG, and LC into a unified correctness score. Experiments on five LLMs and six benchmarks across four domains show that our method distinguishes correct from incorrect responses more effectively than existing baselines, and generalizes to multilingual, multimodal, and open-world settings.
Tri-level method improves AI preference learning under shifting conditions
TPO: Tri-level Distributionally Robust Learning for OOD Direct Preference Optimization
Abstract: Direct Preference Optimization (DPO) minimizes empirical risk over observed preference comparisons, implicitly assuming that the training preference distribution matches the deployment preference distribution. In practice, this i.i.d. assumption often fails, as preference distribution shift exhibits a natural hierarchical structure across three levels: group-level shift, prompt-level shift, and preference-level shift. Such multi-level distributional mismatch poses a critical challenge for safe LLM deployment, yet existing robust DPO methods typically either model uncertainty over the joint preference distribution or focus on individual sources of shift, leaving the hierarchical structure of OOD preference shifts underexplored. We propose TPO, a tri-level distributionally robust framework for OOD preference alignment, which jointly models all three levels of distributional uncertainty through a hierarchically structured rectangular ambiguity set. By instantiating it with Wasserstein distance, we derive a tractable and interpretable objective that combines a closed-form preference-level margin penalty, prompt-level adversarial embedding perturbation, and worst-case group reweighting, each governed by an independent radius. We establish finite-sample convergence guarantees and empirically validate the effectiveness of TPO over existing robust DPO methods.
Distributed subliminal learning shares knowledge without sending model updates
Distributed Subliminal Learning: Replacing Model Updates with Random-Carrier Outputs
Abstract: Collaborative learning typically exchanges model parameters: federated clients communicate updates, while independently adapted foundation models are combined by exchanging adapters or checkpoints. This makes communication scale with model size and requires local specializations to be reconciled in weight space, where interference is common. We ask whether knowledge can instead be shared through model behavior on task-unrelated inputs. We introduce Distributed Subliminal Learning (DSL), a collaborative learning primitive in which participants adapt a common model locally, probe it with task-unrelated inputs, and transmit only the resulting carrier outputs. A coordinator pools these outputs and distills them into a shared model. The primitive supports one-shot foundation-model composition through carrier completions and iterative federated learning through carrier logits, without transmitting model updates or requiring task-related proxy data. In LLM composition, compared with LoRA averaging, DSL achieves higher preference retention (94.56% vs. 87.76%) and a larger GSM8K gain over the base model (22.0 vs. 0.6 points), while reducing upload by 30.6-49.0$\times$. In federated classification, DSL reaches 96.83% on MNIST with 8.9$\times$ less uplink than FedAvg and provides lower-communication operating points on CIFAR-10 and Tiny ImageNet. These results establish random-carrier outputs as a practical communication primitive for knowledge sharing across distinct collaborative learning paradigms.
Adversarial requests trick length predictions to jump scheduling queue
Jumping the Line: Exploiting Length Predictions in LLM Scheduling
Abstract: Efficient request scheduling is increasingly important for reducing completion time in large language model (LLM) serving. Size-based policies such as Shortest Job First prioritize shorter requests, but output lengths are unknown before generation, so practical schedulers rely on predicted lengths. We introduce JIL, an attack on prediction-based LLM schedulers that manipulates the scheduling signal to obtain higher priority and reduce completion time. Using TRAIL as a case study, JIL optimizes an adversarial suffix that causes a lightweight output-length probe to underestimate a request's length. We evaluate JIL on two datasets and four LLMs across varied request profiles and deployment configurations. JIL reduces predicted output lengths by up to 83.4 percent, and adversarial requests complete up to 1.53 times faster on average in end-to-end serving experiments. The reduction in predicted length is substantially larger than the change in actual output length, revealing a mismatch between the scheduler's estimate and the request's realized size. Response utility varies across models and tasks, exposing a trade-off between scheduling advantage and response quality. We also evaluate scheduler-side defenses and find that grouping length predictions into coarse intervals reduces JIL's scheduling advantage and mitigates delays to benign requests.
Agent evaluations depend on tasks and scaffolds for reliability
Agent Evaluation Reliability: More Tasks Won't (Always) Fix An Agent Leaderboard
Abstract: Agent evaluations are increasingly used to compare LLMs and inform deployment decisions, yet ranks can reflect not only the model but also the effects of the evaluation conditions such as the scaffolds or tasks. This makes reliability claim-dependent: an evaluation that reliably ranks deployed systems may not reliably rank underlying models. We ask which conclusions current agent evaluations reliably support and what additional evaluation would improve them. We develop a Bayesian variance-decomposition framework for sparse, imbalanced agent leaderboards and apply it to 22 benchmarks from the Holistic Agent Leaderboard and Harbor Index. The framework separates signal, performance differences relevant to the intended claim, from noise, irrelevant variation that can still change rankings. We find: (1) Reliability depends on the measurement goal. Fixed model-scaffold systems are ranked reliably (0.935-0.994), while underlying-model reliability is substantially lower (0.148-0.841). (2) Scaffold choice can change conclusions. Inter-scaffold reliability measures whether scaffolds preserve model rankings, showing that scaffold effects vary substantially across evaluations. (3) More tasks cannot resolve all uncertainty. Even infinitely many similarly constructed tasks improve model-ranking reliability of a benchmark by at most 0.097 when uncertainty is dominated by limited scaffold coverage. (4) Pooling diverse benchmarks can improve cross-task rankings at lower cost. For rankings across diverse agentic tasks, pooling benchmarks raises projected reliability from 0.44 to 0.75 at the same task budget and can reduce projected cost by up to 83\%. Evaluation design should follow the intended claim: identify what a score or ranking should mean, diagnose what limits its reliability, and spend evaluation budget on the sources of uncertainty that matter.
Galahad cuts repeated reading costs for large language models
Working Around the Compute Ceiling: Byte-Exact Memory in Galahad Makes LLM Reading a One-Time Cost LLM Reading a One-Time Cost
Abstract: A transformer language model performs a bounded amount of computation per token, and recent work by Vishal Sikka, former CEO of Infosys, argues that this bound limits which tasks a model can carry out or verify (arXiv:2507.07505). We ask how much of the budget beneath that ceiling is spent on work the model has already done. Serving is stateless across requests: a model that answers a second question about a document recomputes the document's attention state from the first token. On seven real-world datasets, 98.7% of prompt tokens were text the model had already read. We present Galahad, a memory layer for vLLM, SGLang and llama.cpp that makes this reading a one-time cost. Taliesin saves the model's key-value (KV) state for a block of text and loads it on the next request that contains the same bytes, instead of recomputing it. Blaise keeps the documents themselves and passes the model only the section a question needs. On a recall test with 100 facts hidden in a 97,000-token corpus (Gemma 4 31B), Taliesin alone let the model attend to the whole corpus and answered 98 of 100 on llama.cpp at 3.0 s and 572 J per question, against 10 of 100, 9.3 s and 2,754 J for the same model without Galahad, which could hold only the last 12,000 tokens. With Blaise added, the model read about 668 tokens per question and answered 100 of 100 on all three runtimes at 0.59-0.64 s and 200-213 J; a tuned RAGFlow pipeline answered 77. Storing the corpus is a one-time cost of about 100 s and 28 kJ, whose energy is recovered after 13 questions. Restored state is bit-identical: all 262,144 output logits matched after restart, rehydration and hot-load. Galahad worked with all 30 models we tested under vLLM, and it fails closed: any load that does not pass its checks is recomputed. Together these results move LLM serving from stateless to stateful inference.
Language models recover meaning from corrupted inputs spontaneously
Spontaneous Context Restoration: How Language Models Recover from Corrupted Inputs
Abstract: Language models sometimes produce correct outputs even when their inputs are corrupted by deletion, replacement, or misspelling. We study the internal processes accompanying this behavior, which we call context restoration, in controlled attention-only transformers and five pretrained LLMs (1B-32B parameters) across arithmetic, reading comprehension, and multiple-choice reasoning tasks. In the attention-only transformers, restoration emerges spontaneously despite training exclusively on clean sequences, without corruption training or an explicit denoising objective. We find that context restoration follows a two-phase process: early layers localize effects associated with repair at corrupted positions, while later layers accumulate these effects at uncorrupted positions through the residual stream and ultimately concentrate them at the output position. Repair outcome is predictable from hidden states: cosine alignment with the clean state is highly predictive in attention-only models, while linear probes recover additional information in pretrained LLMs. A linear probe using only the corrupted prompt's first-block hidden state predicts failure with mean ROC-AUC 0.78. This enables failure triage under matched or even partially shifted deployment conditions and may reduce unnecessary verification or computation. Failed examples also show substantially greater nonlinearity along corruption directions. Moderate-corruption finetuning increases corruption tolerance while simultaneously reducing displacement-normalized linearization error, associating improved robustness with a more nearly linear response to corruption.