Efficient Language Model Inference

Follow this topic Ask about this topic

The field studies how to serve language models with less computation, memory, time, and energy while preserving useful outputs. Approaches include quantization and sparsity, faster decoding such as speculative and parallel generation, KV-cache and context compression, and smaller or specialized models routed to tasks they can handle. For reasoning models, early stopping and adaptive compute allocation aim to spend more tokens on difficult questions and fewer on easy ones. The central challenge is that savings depend on the model, workload, hardware, and serving setup, and reducing token counts or latency does not always reduce total cost or energy without harming accuracy or reliability.

Latest papers

321 papers64 papers in the last four weeks, up 68% on the four weeks before. 0.4% of all new papers.

Apr 9, 2026

VineLM: Trie-Based Fine-Grained Control for Agentic Workflows

Agentic workflows interleave configurable LLM stages with tool stages and often include retries or refinement loops. Existing workflow managers profile full workflow configurations offline and assign each request a static workflow-level plan that binds each configurable LLM stage to a single model, reuses that model across repeated loop iterations, and does not revisit those choices at runtime. We present VineLM, a workflow manager that enables fine-grained control by choosing the model for each stage invocation as execution unfolds under request-level objectives such as maximizing accuracy under cost or latency budgets. VineLM represents feasible executions as an annotated trie of model-choice prefixes and uses checkpointing and cascade profiling to estimate path accuracy, cost, and latency without exhaustively profiling every request on every path. At runtime, VineLM re-roots the trie after each stage invocation and replans over the remaining subtrie using the realized execution prefix and remaining latency budget. On NL2SQL and math reasoning workflows, VineLM improves the cost-latency-accuracy frontier over coarse workflow-level baselines, achieving up to 18% higher accuracy at the same per-request budget with its sparse profiling reducing offline profiling cost by 98-99.8% when compared to exhaustive profiling.

Nikos Pagonas, Matthew Lou, Tianyi Peng +2

Columbia University

LLM Agent Workflow Optimization Agentic Workflows LLM Routing Efficient Language Model Inference

Apr 9, 2026

Alloc-MoE: Budget-Aware Expert Activation Allocation for Efficient Mixture-of-Experts Inference

Mixture-of-Experts (MoE) has become a dominant architecture for scaling large language models due to their sparse activation mechanism. However, the substantial number of expert activations creates a critical latency bottleneck during inference, especially in resource-constrained deployment scenarios. Existing approaches that reduce expert activations potentially lead to severe model performance degradation. In this work, we introduce the concept of activation budget as a constraint on the number of expert activations and propose Alloc-MoE, a unified framework that optimizes budget allocation coordinately at both the layer and token levels to minimize performance degradation. At the layer level, we introduce Alloc-L, which leverages sensitivity profiling and dynamic programming to determine the optimal allocation of expert activations across layers. At the token level, we propose Alloc-T, which dynamically redistributes activations based on routing scores, optimizing budget allocation without increasing latency. Extensive experiments across multiple MoE models demonstrate that Alloc-MoE maintains model performance under a constrained activation budget. Especially, Alloc-MoE achieves

1.15\times

prefill and

1.34\times

decode speedups on DeepSeek-V2-Lite at half of the original budget.

Baihui Liu, Kaiyuan Tian, Wei Wang +3

National Key Laboratory of Parallel and Distributed Computing, College of Computer Science and Technology, National University of Defense Technology

Mixture-of-Experts Inference Mixture of Experts Efficient Language Model Inference

Apr 8, 2026

CROP: Token-Efficient Reasoning in Large Language Models via Regularized Prompt Optimization

Large Language Models utilizing reasoning techniques improve task performance but incur significant latency and token costs due to verbose generation. Existing automatic prompt optimization(APO) frameworks target task accuracy exclusively at the expense of generating long reasoning traces. We propose Cost-Regularized Optimization of Prompts (CROP), an APO method that introduces regularization on response length by generating textual feedback in addition to standard accuracy feedback. This forces the optimization process to produce prompts that elicit concise responses containing only critical information and reasoning. We evaluate our approach on complex reasoning datasets, specifically GSM8K, LogiQA and BIG-Bench Hard. We achieved an 80.6% reduction in token consumption while maintaining competitive accuracy, seeing only a nominal decline in performance. This presents a pragmatic solution for deploying token-efficient and cost-effective agentic AI systems in production pipelines.

Deep Shah, Sanket Badhe, Nehal Kathrotia +1

Google LLC · Purdue University

Efficient Language Model Reasoning Automatic Prompt Optimization Efficient Language Model Inference

Apr 8, 2026

SHIELD: A Segmented Hierarchical Memory Architecture for Energy-Efficient LLM Inference on Edge NPUs

Large Language Model (LLM) inference on edge Neural Processing Units (NPUs) is fundamentally constrained by limited on-chip memory capacity. Although high-density embedded DRAM (eDRAM) is attractive for storing activation workspaces, its periodic refresh consumes substantial energy. Prior work has primarily focused on reducing off-chip traffic or optimizing refresh for persistent Key-Value (KV) caches, while transient and error-resilient Query and Attention Output (QO) activations are largely overlooked. We propose SHIELD, a lifecycle-aware segmented eDRAM architecture that jointly exploits temporal residency and bit-level sensitivity in bfloat16 (BF16) activations. SHIELD isolates the sign and exponent fields from the mantissa, disables refresh for transient QO mantissas, and applies relaxed refresh to persistent KV mantissas. Across multiple LLMs and inference scenarios, SHIELD reduces eDRAM refresh energy by 35% relative to a standard-refresh baseline while preserving accuracy on WikiText-2, PIQA, and ARC-Easy.

Jintao Zhang, Xuanyao Fong

Department of Electrical and Computer Engineering, National University of Singapore, Singapore

On-Device Language Model Inference Efficient Language Model Inference Memory-Efficient Inference Memory-Efficient Optimization

Apr 8, 2026

The Detection-Extraction Gap: Models Know the Answer Before They Can Say It

Modern reasoning models continue generating long after the answer is already determined. Across five model configurations, two families, and three benchmarks, we find that 52--88% of chain-of-thought tokens are produced after the answer is recoverable from a partial prefix. This post-commitment generation reveals a structural phenomenon: the detection-extraction gap. Free continuations from early prefixes recover the correct answer even at 10% of the trace, while forced extraction fails on 42% of these cases. The answer is recoverable from the model state, yet prompt-conditioned decoding fails to extract it. We formalize this mismatch via a total-variation bound between free and forced continuation distributions, yielding quantitative estimates of suffix-induced shift. Exploiting this asymmetry, we propose Black-box Adaptive Early Exit (BAEE), which uses free continuations for both detection and extraction, truncating 70--78% of serial generation while improving accuracy by 1--5pp across all models. For thinking-mode models, early exit prevents post-commitment overwriting, yielding gains of up to 5.8pp; a cost-optimized variant achieves 68--73% reduction at a median of 9 API calls. Code is available at https://github.com/EdWangLoDaSc/know2say.

Hanyang Wang, Mingxuan Zhu

The University of Chicago · Imperial College London

Efficient Language Model Reasoning Efficient Language Model Inference Overthinking in Language Models

Apr 6, 2026

Just Pass Twice: Efficient Token Classification with LLMs for Zero-Shot NER

Large language models encode extensive world knowledge valuable for zero-shot named entity recognition. However, their causal attention mechanism, where tokens attend only to preceding context, prevents effective token classification when disambiguation requires future context. Existing approaches use LLMs generatively, prompting them to list entities or produce structured outputs, but suffer from slow autoregressive decoding, hallucinated entities, and formatting errors. We propose Just Pass Twice (JPT), a simple yet effective method that enables causal LLMs to perform discriminative token classification with full bidirectional context. Our key insight is that concatenating the input to itself lets each token in the second pass attend to the complete sentence, requiring no architectural modifications. We combine these representations with definition-guided entity embeddings for flexible zero-shot generalization. Our approach achieves state-of-the-art results on zero-shot NER benchmarks, surpassing the previous best method by +7.9 F1 on average across CrossNER and MIT benchmarks, being over 20x faster than comparable generative methods. Interactive demo and access to the model are available at https://witness-ai-jpt-ner.hf.space/

Ahmed Ewais, Ahmed Hashish, Amr Ali

WitnessAI

Named Entity Recognition Decoder-Only Language Models Efficient Language Model Inference

Apr 5, 2026

Adaptive Cost-Efficient Evaluation for Reliable Patent Claim Generation

Automated patent claim validation demands low error tolerance. However, existing approaches face a rigidity-resource dilemma: lightweight encoders cannot track long-range legal dependencies, while exhaustive LLM verification incurs 4-5X higher overhead at million-claim scale. A naive confidence-based cascade cannot resolve this because binary validity scores fail to distinguish structurally distinct error types which require different reasoning depths. We propose a two-stage framework: Adaptive Cost-efficient Evaluation (ACE), which exploits the categorical structure of patent errors for uncertainty-aware routing. In the first stage, a fine-tuned encoder projects claims into a K+1 distribution over legal error types, whose predictive entropy serves as the routing signal. Claims exceeding an entropy threshold are escalated to the second stage, where an expert LLM executes a schema-constrained Chain-of-Patent-Thought (CoPT) protocol to map claim elements against 35 U.S.C. standards whose schema constraint reduces per-claim latency by 42% while producing legally grounded verdicts. We further present a 40,000-claim dataset ACE-40k with MPEP-grounded annotations, where ACE surpasses competitive baselines including a supervised 70B-parameter LLM while reducing costs by 78%. On real USPTO rejection data, the routing mechanism transfers without re-calibration, reducing inference time by 60% while maintaining competitive recall.

Yongmin Yoo, Qiongkai Xu, Longbing Cao

Frontier AI Research Centre, Macquarie University · School of Computing, FSE, Macquarie University

Claim Verification Automated Evaluation Adaptive Inference Efficient Language Model Inference

Apr 2, 2026

Batched Contextual Reinforcement: A Task-Scaling Law for Efficient Reasoning

Large Language Models employing Chain-of-Thought reasoning achieve strong performance but suffer from excessive token consumption that inflates inference costs. Existing efficiency methods such as explicit length penalties, difficulty estimators, or multi-stage curricula either degrade reasoning quality or require complex training pipelines. We introduce Batched Contextual Reinforcement, a minimalist, single-stage training paradigm that unlocks efficient reasoning through a simple structural modification: training the model to solve N problems simultaneously within a shared context window, rewarded purely by per-instance accuracy. This formulation creates an implicit token budget that yields several key findings: (1) We identify a novel task-scaling law: as the number of concurrent problems N increases during inference, per-problem token usage decreases monotonically while accuracy degrades far more gracefully than baselines, establishing N as a controllable throughput dimension. (2) BCR challenges the traditional accuracy-efficiency trade-off by demonstrating a "free lunch" phenomenon at standard single-problem inference. Across both 1.5B and 4B model families, BCR reduces token usage by 15.8% to 62.6% while consistently maintaining or improving accuracy across five major mathematical benchmarks. (3) Qualitative analyses reveal emergent self-regulated efficiency, where models autonomously eliminate redundant metacognitive loops without explicit length supervision. (4) Crucially, we empirically demonstrate that implicit budget constraints successfully circumvent the adversarial gradients and catastrophic optimization collapse inherent to explicit length penalties, offering a highly stable, constraint-based alternative for length control. These results prove BCR practical, showing simple structural incentives unlock latent high-density reasoning in LLMs.

Bangji Yang, Hongbo Ma, Jiajun Fan +1

University of Illinois Urbana-Champaign · Tsinghua University

Efficient Language Model Reasoning Efficient Language Model Inference LLM Reasoning RL for Language Model Reasoning

Apr 1, 2026

Universal YOCO for Efficient Depth Scaling

The rise of test-time scaling has remarkably boosted the reasoning and agentic proficiency of Large Language Models (LLMs). Yet, standard Transformers struggle to scale inference-time compute efficiently, as conventional looping strategies suffer from high computational overhead and a KV cache that inflates alongside model depth. We present Universal YOCO (YOCO-U), which combines the YOCO decoder-decoder architecture with recursive computation to achieve a synergistic effect greater than either alone. Built on the YOCO framework, YOCO-U implements a Universal Self-Decoder that performs multiple iterations via parameter sharing, while confining the iterative process to shallow, efficient-attention layers. This combination yields a favorable capability-efficiency tradeoff that neither YOCO nor recursion achieves independently. The YOCO architecture provides a constant global KV cache and linear pre-filling, while partial recursion enhances representational depth with limited overhead. Together, YOCO-U improves token utility and scaling behavior while maintaining efficient inference. Empirical results confirm that YOCO-U remains highly competitive in general and long-context benchmarks, demonstrating that the integration of efficient-attention architectures and recursive computation is a promising direction for scalable LLMs.

Yutao Sun, Li Dong, Tianzhu Ye +3

Microsoft Research · Tsinghua University

Efficient Language Model Inference Inference-Time Scaling Transformer

Mar 29, 2026

Do Small Language Models Know When They're Wrong? Confidence-Based Cascade Scoring for Educational Assessment

Automated scoring of student work at scale requires balancing accuracy against cost and latency. In "cascade" systems, small language models (LMs) handle easier scoring tasks while escalating harder ones to larger LMs -- but the challenge is determining which cases to escalate. We explore verbalized confidence -- asking the LM to state a numerical confidence alongside its prediction -- as a routing signal. Using 2,100 expert-scored decisions from student-AI math conversations, we evaluate cascade systems built from GPT-5.4, Claude 4.5+, and Gemini 3.1 model pairs. We find that: (1) confidence discrimination varies widely across small LMs, with the best achieving AUROC 0.857 and the worst producing a near-degenerate confidence distribution; (2) confidence tracks human scoring difficulty, with lower LM confidence where annotators disagreed and took longer to score; (3) the best cascade approached large-LM accuracy (kappa 0.802 vs. 0.819) at 76% lower cost and 61% lower latency. Confidence discrimination is the bottleneck: the two small LMs with meaningful confidence variance yielded cascades with no statistically detectable kappa loss, while the third -- whose confidence was near-degenerate -- could not close the accuracy gap regardless of threshold. Small LMs with strong discrimination let practitioners trade cost for accuracy along the frontier; those without it do not.

Tyler Burleigh

Khan Academy

Language Model Cascading Automated Grading Language Model Self-Assessment Educational Assessment Confidence Estimation in Language Models Efficient Language Model Inference

Mar 25, 2026

Schema on the Inside: A Two-Phase Fine-Tuning Method for High-Efficiency Text-to-SQL at Scale

Applying large, proprietary API-based language models to text-to-SQL tasks poses a significant industry challenge: reliance on massive, schema-heavy prompts results in prohibitive per-token API costs and high latency, hindering scalable production deployment. We present a specialized, self-hosted 8B-parameter model designed for a conversational bot in CriQ, a sister app to Dream11, India's largest fantasy sports platform with over 250 million users, that answers user queries about cricket statistics. Our novel two-phase supervised fine-tuning approach enables the model to internalize the entire database schema, eliminating the need for long-context prompts. This reduces input tokens by over 99%, from a 17k-token baseline to fewer than 100, and replaces costly external API calls with efficient local inference. The resulting system achieves 98.4% execution success and 92.5% semantic accuracy, substantially outperforming a prompt-engineered baseline using Google's Gemini Flash 2.0 (95.6% execution, 89.4% semantic accuracy). These results demonstrate a practical path toward high-precision, low-latency text-to-SQL applications using domain-specialized, self-hosted language models in large-scale production environments.

Chinmay Soni, Shivam Chourasia, Gaurav Kumar +1

Sporta Technologies Private Limited, Mumbai, Maharashtra, India

Sports Analytics Text-to-SQL LLM Fine-Tuning Efficient Language Model Inference LLM Grounding

Mar 22, 2026

Task-Specific Efficiency Analysis: When Small Language Models Outperform Large Language Models

Large Language Models achieve remarkable performance but incur substantial computational costs unsuitable for resource-constrained deployments. This paper presents the first comprehensive task-specific efficiency analysis comparing 16 language models across five diverse NLP tasks. We introduce the Performance-Efficiency Ratio (PER), a novel metric integrating accuracy, throughput, memory, and latency through geometric mean normalization. Our systematic evaluation reveals that small models (0.5--3B parameters) achieve superior PER scores across all given tasks. These findings establish quantitative foundations for deploying small models in production environments prioritizing inference efficiency over marginal accuracy gains.

Jinghan Cao, Yu Ma, Xinjin Li +2

San Francisco State University - Department of Computer Science 1600 Holloway Ave, San Francisco - USA · Carnegie Mellon University - Department of Computer Science 5000 Forbes Ave, Pittsburgh - USA · Columbia University - Department of Computer Science 116th and Broadway, New York - USA +2

Small Language Models Efficient Language Model Inference LLM Evaluation

Mar 22, 2026

A transformer architecture alteration to incentivise externalised reasoning

We propose a new architectural change, and post-training pipeline, for making LLMs more verbose reasoners by teaching a model to truncate forward passes early. We augment an existing transformer architecture with an early-exit mechanism at intermediate layers and train the model to exit at shallower layers when the next token can be predicted without deep computation. After a calibration stage, we incentivise the model to exit as early as possible while maintaining task performance using reinforcement learning. We provide preliminary results to this effect for small reasoning models, showing that they learn to adaptively reduce computations across tokens. We predict that, applied at the right scale, our approach can minimise the amount of excess computation that reasoning models have at their disposal to perform non-myopic planning using their internal activations, reserving this only for difficult-to-predict tokens.

Elizabeth Pavlova, Mariia Koroliuk, Karthik Viswanathan +3

MARS · Geodesic Research, equal contribution · University of Amsterdam

Efficient Language Model Inference Large Reasoning Models Efficient Transformer Inference

Mar 18, 2026

Adaptive Domain Models: Bayesian Evolution, Warm Rotation, and Principled Training for Geometric and Neuromorphic AI

Transformer-based large language models have become a prominent focus of machine learning. We develop adaptive domain models to supply domain-specific predictions and checks within language-model pipelines, as well as to serve direct machine automation efficiently. Their predictions can inform generation, while declared constraints and error bounds support checks on the calculations underlying proposed responses and actions. These feedforward and feedback paths aim to reduce domain errors while supporting use cases such as constraint solving for machine processes as well as preserving accuracy in a language model's conversational setting. Physical dimensions and geometric support restrict admissible computation, while a probability model expresses uncertainty. Our type and program hypergraph work connects these relationships to numerical and deployment obligations. We derive a natural-gradient calculation over forward tangent directions using a declared program metric, with a calibration example and a bound separating metric, derivative and solve errors. Forward differentiation and posit accumulation expose complementary resource and error tradeoffs. Our worked linear-Gaussian update connects posterior computation to error and a distribution-shift decision. Recent results from Urschel bound elimination growth probabilistically under distinct Gaussian-input and random-preconditioning hypotheses. Our unique approach to warm rotation binds candidate parameters to their numerical evidence and destination while each request observes a coherent model and state. We propose Bayesian distillation as a source of prior information. Our architecture offers a route to domain computations with explicit guarantees that can guide automated decisions and model responses. We specify experiments to evaluate prediction quality, domain errors across conversational turns, and total resource cost.

Houston Haynes

Braidpoint, Asheville, NC

Efficient Language Model Inference Bayesian Inference

Mar 17, 2026

EngGPT2: Sovereign, Efficient and Open Intelligence

EngGPT2-16B-A3B is the latest iteration of Engineering Group's Italian LLM and it's built to be a Sovereign, Efficient and Open model. EngGPT2 is trained on 2.5 trillion tokens - less than Qwen3's 36T or Llama3's 15T - and delivers performance on key benchmarks, including MMLU-Pro, GSM8K, IFEval and HumanEval, comparable to dense models in the 8B-16B range, while requiring one-fifth to half of the inference power, and between one-tenth to one-sixth of the training data and consequent needed training power. Designed as a trained-from-scratch Mixture-of-Experts (MoE) architecture, EngGPT2 features 16 billion parameters with 3 billion active per inference, with expert sizes positioned between those used in GPT-OSS and Qwen3. Approximately 25% of its training corpus consists of Italian-language data, to deliver strong capabilities for European and Italian NLP tasks among models of similar scale. This efficiency aims to position EngGPT2 as a key contributor to the growing portfolio of open-weight European models, combining performance and efficiency with full alignment to the EU AI Act. EngGPT2 is also a single model capable of multiple reasoning modes: non-reasoning, reasoning in Italian or English, and turbo-reasoning (a concise, bullet-point style reasoning available in both languages designed for real-time reasoning use cases). EngGPT2 aims to set a new standard for resource-conscious, high-performance LLMs tailored to European and Italian contexts.

G. Ciarfaglia, A. Rosanova, S. Cipolla +17

Language Modeling Open-Weight Language Models Efficient Language Model Inference Efficient Language Model Training Mixture-of-Experts Language Models

Mar 17, 2026

SpecSteer: Synergizing Local Context and Global Reasoning for Efficient Personalized Generation

Realizing personalized intelligence faces a core dilemma: sending user history to centralized large language models raises privacy concerns, while on-device small language models lack the reasoning capacity required for high-quality generation. Our pilot study shows that purely local enhancements remain insufficient to reliably bridge this gap. We therefore propose SpecSteer, an asymmetric collaborative inference framework that synergizes private on-device context with cloud-scale reasoning. SpecSteer casts collaboration as Bayesian knowledge fusion and repurposes speculative decoding as a distributed alignment protocol, yielding a Draft--Verify--Recover pipeline: the on-device model drafts personalized sequences; the cloud validates via a ratio-based mechanism that decouples reasoning verification from private context, filtering logical flaws without accessing raw user context; upon rejection, a steering recovery injects local intent during correction. Experiments demonstrate that SpecSteer successfully closes the reasoning gap and achieves superior personalized generation performance, while delivering a 2.36x speedup over standard baselines.

Hang Lv, Sheng Liang, Hao Wang +6

University of Science and Technology of China · Huawei Technologies Co., Ltd.

Personalized Language Models On-Device Language Model Inference Privacy-Preserving Language Models Efficient Language Model Inference Language Model Steering

Mar 17, 2026

NeuronSpark: Selective Spiking State Space Models with Adaptive Computation for Language Modeling

Spiking neural networks promise brain-inspired neural dynamics, yet native large-scale spiking language-model pretraining remains largely unexplored relative to conversion, distillation, and backbone-adaptation approaches. We diagnose a central obstacle: standard spiking neurons apply the same firing sensitivity to every token, creating dense- or silent-firing regimes that are difficult to resolve with globally tuned parameters alone. This problem is particularly acute in language modeling, where a single sequence contains tokens of vastly different informational roles, making a uniform firing regime poorly matched to the data. We address this with NeuronSpark, a self-attention-free language model built around input-dependent spiking membrane dynamics, where each neuron's decay rate, gain, and firing threshold adapt to the current token content. Together with gated associative memory and adaptive temporal aggregation, this selectivity enables billion-scale pretraining from random initialization without backbone transfer, conversion, or distillation. We relate the neuron to fixed-parameter spiking models and the memory to linear attention. At 1.1B parameters and 2.1B tokens, NeuronSpark reaches a 40.0 eight-task zero-shot average, exceeds comparable-budget Pythia/Mamba checkpoints, retains 95.2%/93.2% of their fully trained scores, and spontaneously develops a fast--slow hierarchy. At 2,048 tokens, parameter-matched VHDL points reduce operation-energy proxy by 90.1%/67.0% and cycles by 94.1%/80.3% versus Transformer/Mamba; a physical FPGA path reaches 24.000s TTFT and 0.082440 TPS.

Zhengzheng Tang

Boston University

Spiking Language Models Spiking Neural Networks Language Model Pretraining Language Modeling Efficient Language Model Inference State Space Models

Mar 16, 2026

Mamba-3: Improved Sequence Modeling using State Space Principles

Scaling inference-time compute has emerged as an important driver of LLM performance, making inference efficiency a central focus of model design alongside model quality. While the current Transformer-based models deliver strong model quality, their quadratic compute and linear memory make inference expensive. This has spurred the development of sub-quadratic models with reduced linear compute and constant memory requirements. However, many recent linear models trade off model quality and capability for algorithmic efficiency, failing on tasks such as state tracking. Moreover, their theoretically linear inference remains hardware-inefficient in practice. Guided by an inference-first perspective, we introduce three core methodological improvements inspired by the state space model (SSM) viewpoint of linear models. We combine: (1) a more expressive recurrence derived from SSM discretization, (2) a complex-valued state update rule that enables richer state tracking, and (3) a multi-input, multi-output (MIMO) formulation for better model performance without increasing decode latency. Together with architectural refinements, our Mamba-3 model achieves significant gains across retrieval, state-tracking, and downstream language modeling tasks. At the 1.5B scale, Mamba-3 improves average downstream accuracy by 0.6 percentage points compared to the next best model (Gated DeltaNet), with Mamba-3's MIMO variant further improving accuracy by another 1.2 points for a total 1.8 point gain. Across state-size experiments, Mamba-3 achieves comparable perplexity to Mamba-2 despite using half of its predecessor's state size. Our evaluations demonstrate Mamba-3's ability to advance the performance-efficiency Pareto frontier.

Aakash Lahoti, Kevin Y. Li, Berlin Chen +5

Carnegie Mellon University · Princeton University · Together AI +1

State Space Language Models Sequence Modeling Efficient Language Model Inference State Space Models State Tracking

Mar 15, 2026

FlashHead: Efficient Drop-In Replacement for the Classification Head in Language Model Inference

Language models are increasingly adopting smaller architectures optimized for consumer devices. In this setting, inference efficiency is the primary constraint. Meanwhile, vocabulary sizes continue to grow rapidly, making the classification head a critical bottleneck that accounts for up to 60% of model parameters, and 50% of inference compute. We introduce FlashHead, the first efficient drop-in replacement for the dense classification head that is training-free and hardware-friendly. FlashHead builds on principles from information retrieval, reframing that computation at the output head as a retrieval problem rather than a dense classification over the full vocabulary. FlashHead introduces four key innovations: (1) a balanced clustering scheme that structures vocabulary partitions into compact hardware-efficient tensors, (2) extending multiprobe retrieval to language model heads, enabling thousands of clusters to be scored in parallel, (3) a novel inference-time sampling mechanism that extends retrieval beyond top tokens, enabling probabilistic sampling across the full vocabulary, and (4) selective quantization, enabling effective low-bit computation in the head. Experiments on Llama-3.2, Gemma-3, and Qwen-3 show that FlashHead delivers model-level inference speedups of up to 1.75x which maintaining output accuracy compared to the original head. By overcoming the classification head bottleneck, FlashHead establishes a new benchmark for efficient inference and removes a key barrier to developing smaller, capable models for consumer hardware.

Wilhelm Tranheden, Shahnawaz Ahmed, Devdatt Dubhashi +2

Embedl Gamla Almedalsvägen 21, 412 63 Gothenburg, Sweden

LLM Inference Acceleration Efficient Language Model Inference

Mar 14, 2026

An Alternative Trajectory for Generative AI

The generative artificial intelligence (AI) ecosystem is undergoing rapid transformations that threaten its sustainability. As models transition from research prototypes to high-traffic products, the energetic burden has shifted from one-time training to recurring, unbounded inference. This is exacerbated by reasoning models that inflate compute costs by orders of magnitude per query. The prevailing pursuit of artificial general intelligence through scaling of monolithic models is colliding with hard physical constraints: grid failures, water consumption, and diminishing returns on data scaling. This trajectory yields models with impressive factual recall but struggles in domains requiring in-depth reasoning, possibly due to insufficient abstractions in training data. Current large language models (LLMs) exhibit genuine reasoning depth only in domains like mathematics and coding, where rigorous, pre-existing abstractions provide structural grounding. In other fields, the current approach fails to generalize well. We propose an alternative trajectory based on domain-specific superintelligence (DSS). We argue for first constructing explicit symbolic abstractions (knowledge graphs, ontologies, and formal logic) to underpin synthetic curricula enabling small language models to master domain-specific reasoning without the model collapse problem typical of LLM-based synthetic data methods. Rather than a single generalist giant model, we envision "societies of DSS models": dynamic ecosystems where orchestration agents route tasks to distinct DSS back-ends. This paradigm shift decouples capability from size, enabling intelligence to migrate from energy-intensive data centers to secure, on-device experts. By aligning algorithmic progress with physical constraints, DSS societies move generative AI from an environmental liability to a sustainable force for economic empowerment.

Margarita Belova, Yuval Kansal, Yihao Liang +2

Princeton University, USA

Language Modeling Neuro-Symbolic AI Efficient Language Model Inference

Mar 14, 2026

Retrieval-Feedback-Driven Distillation and Preference Alignment for Efficient LLM-based Query Expansion

Large language models have recently enabled a generative paradigm for query expansion, but their high inference cost makes direct deployment difficult in practical retrieval systems. To address this issue, a retrieval-feedback-driven distillation and preference-alignment framework is proposed to transfer retrieval-friendly expansion behavior from a strong teacher model to a compact student model. Rather than relying on few-shot exemplars at inference time, the framework first leverages two complementary types of teacher-generated expansions, produced under zero-shot and few-shot prompting conditions, as supervision signals for distillation and as candidate pools for preference construction. A retrieval-metric-driven strategy is then introduced to automatically form chosen/rejected expansion pairs according to nDCG@10 differences, and Direct Preference Optimization is applied to explicitly align generation preferences with retrieval objectives. Experiments on TREC DL19/20/21 and MIRACL-zh show that the proposed approach preserves strong retrieval effectiveness while substantially reducing inference cost. In particular, the distilled Qwen3-4B model reaches about 97% of the teacher (DeepSeek-685B) model's nDCG@10 performance on DL19, and remains effective on the Chinese MIRACL-zh benchmark, demonstrating strong practicality across both English and Chinese retrieval settings.

Minghan Li, Guodong Zhou

School of Computer Science and Technology, Soochow University, China

Language Model Distillation Efficient Language Model Inference LLM Alignment

Mar 13, 2026

TERMINATOR: Learning Optimal Exit Points for Early Stopping in Chain-of-Thought Reasoning

Large Reasoning Models (LRMs) achieve impressive performance on complex reasoning tasks via Chain-of-Thought (CoT) reasoning, which enables them to generate intermediate thinking tokens before arriving at the final answer. However, LRMs often suffer from significant overthinking, spending excessive compute time even after the answer is generated early on. Prior work has identified the existence of an optimal reasoning length such that truncating reasoning at this point significantly shortens CoT outputs with virtually no change in performance. However, determining optimal CoT lengths for practical datasets is highly non-trivial as they are fully task and model-dependent. In this paper, we precisely address this and design Terminator, an early-exit strategy for LRMs at inference to mitigate overthinking. The central idea underpinning Terminator is that the first arrival of an LRM's final answer is often predictable, and we leverage these first answer positions to create a novel dataset of optimal reasoning lengths to train Terminator. Powered by this approach, Terminator achieves significant reductions in CoT lengths of 14%-55% on average across four challenging practical datasets: MATH-500, AIME 2025, HumanEval, and GPQA, while outperforming current state-of-the-art methods and reducing inference latency by more than 2x compared to the original LRM.

Alliot Nagle, Jakhongir Saydaliev, Dhia Garbaya +3

UT Austin · EPFL · ENS Paris-Saclay +1

Early Stopping for Language Model Reasoning Efficient Language Model Inference LLM Inference Acceleration Overthinking in Language Models CoT Reasoning Large Reasoning Models

Mar 10, 2026

EnComp: Lightweight Encoder-Only Context Compression for Retrieval-Augmented Question Answering

Efficient context compression is critical for retrieval-augmented question answering in resource-constrained settings, where long retrieved contexts increase latency, memory use, and LLM reader cost. We propose a lightweight encoder-only framework for query-driven sentence pruning that preserves answer-critical evidence while aggressively reducing irrelevant context. Our method learns marginal contribution scores for sentences using counterfactual training signals and optimizes a contrastive ranking objective that separates critical evidence from noncritical context. Our approach scores all sentences from a single full-context encoding, enabling fast inference with low computational overhead. Experiments show that it maintains accuracy comparable to the strongest baseline while using 3.7

\times

less peak memory and achieving nearly 3

\times

lower compression latency, demonstrating an effective quality--efficiency trade-off for practical resource-constrained deployment.

Thao Do, Dinh Phu Tran, An Vo +2

Korea Advanced Institute of Science and Technology (KAIST), Daejeon 34141, South Korea

Efficient Language Model Inference Knowledge-Intensive QA Context Compression RAG Context Compression LLM Context Compression

Mar 6, 2026

The Compression Paradox in LLM Inference: Provider-Dependent Energy Effects of Prompt Compression

The rapid proliferation of Large Language Models has created an environmental paradox: the very technology that could help solve climate challenges is itself becoming a significant contributor to global carbon emissions. We test whether prompt compression improves inference energy efficiency in 28,421 successful API trials (28,428 planned) across three providers (OpenAI GPT-4o-mini, Anthropic Claude-3.5-Sonnet, and DeepSeek-Chat), five benchmarks (HumanEval, MBPP, GSM8K, MATH, MMLU), and four compression ratios (r in {1.0, 0.7, 0.5, 0.3}). Energy is estimated with a token-based proxy calibrated against local direct measurements, and quality is tracked with benchmark pass rates. Compression produced substantial quality loss (overall pass rate 26.0% at baseline vs. 1.5% at r=0.7) and strongly provider-dependent energy behavior. DeepSeek exhibited output expansion under compression (21 to 798 tokens at r=0.3), corresponding to energy increases up to +2,140%, while GPT-4o-mini showed mixed effects including a reduction at r=0.5. These results indicate that input-token reduction alone is not a reliable energy optimization strategy in production inference. For the evaluated settings, model selection and output-length control provided more consistent energy-quality tradeoffs than prompt compression.

Warren Johnson

Plexor Labs Principal Researcher

LLM Inference Efficiency Energy-Aware LLM Inference Efficient Language Model Inference Prompt Compression

Mar 6, 2026

Compression Method Matters: Benchmark-Dependent Output Dynamics in LLM Prompt Compression

Prompt compression is often evaluated by input-token reduction, but its real deployment impact depends on how compression changes output length and total inference cost. We present a controlled replication and extension study of benchmark-dependent output dynamics under aggressive compression, covering 5,400 API calls across three benchmarks and multiple providers. To explain conflicting prior observations, we formalize instruction survival probability (Psi), a structural metric that captures whether task-critical prompt segments remain after truncation. Results show a strong benchmark effect: under r=0.3, DeepSeek exhibits severe output expansion on MBPP (56x, Psi approx 0.15) but substantially lower expansion on HumanEval (5x, Psi approx 0.72), while GPT-4o-mini is comparatively stable across benchmarks. This reconciles the apparent discrepancy between previously reported extreme explosion and lower replication effects by identifying prompt structure, not provider identity alone, as the primary moderator. We introduce the Compression Robustness Index (CRI) for cross-benchmark evaluation and show that single-benchmark assessments can produce misleading conclusions about compression safety and efficiency. To contextualize energy claims, we incorporate companion direct NVML measurements from rented RunPod GPUs and show that token savings can overstate joule savings. These findings motivate benchmark-diverse testing and structure-aware compression policies for reliable, energy-conscious LLM deployment.

Warren Johnson

Plexor Labs Principal Researcher

Prompt Compression LLM Inference Efficiency Efficient Language Model Inference LLM Evaluation

Mar 6, 2026

Prompt Compression in Production Task Orchestration: A Pre-Registered Randomized Trial

The economics of prompt compression depend not only on reducing input tokens but on how compression changes output length, which is typically priced several times higher. We evaluate this in a pre-registered six-arm randomized controlled trial of prompt compression on production multi-agent task-orchestration, analyzing 358 successful Claude Sonnet 4.5 runs (59-61 per arm) drawn from a randomized corpus of 1,199 real orchestration instructions. We compare an uncompressed control with three uniform retention rates (r=0.8, 0.5, 0.2) and two structure-aware strategies (entropy-adaptive and recency-weighted), measuring total inference cost (input+output) and embedding-based response similarity. Moderate compression (r=0.5) reduced mean total cost by 27.9%, while aggressive compression (r=0.2) increased mean cost by 1.8% despite substantial input reduction, consistent with small mean output expansion (1.03x vs. control) and heavy-tailed uncertainty. Recency-weighted compression achieved 23.5% savings and, together with moderate compression, occupied the empirical cost-similarity Pareto frontier, whereas aggressive compression was dominated on both cost and similarity. These results show that "compress more" is not a reliable production heuristic and that output tokens must be treated as a first-class outcome when designing compression policies.

Warren Johnson, Charles Lee

Plexor Labs, Issaquah, WA, USA · Project Autobots, Seattle, WA, USA

Prompt Compression Token Compression Efficient Language Model Inference Cost-Aware Inference

Mar 6, 2026

Diffusion Language Models Are Natively Length-Aware

Unlike autoregressive language models, which terminate variable-length generation upon predicting an End-of-Sequence (EoS) token, Diffusion Language Models (DLMs) operate over a fixed maximum-length context window for a predetermined number of denoising steps. However, this process is independent of the required response length, resulting in computational waste for the majority of short responses common in reasoning and chat tasks. To address this problem, we conjecture that the latent prompt representation contains sufficient information to estimate the required output length. We provide empirical evidence for this phenomenon and propose a zero-shot mechanism to dynamically crop the context window before generation begins, leading to fewer diffusion steps and substantial computational savings. We evaluate our approach on four benchmarks with diverse tasks -- GSM8K (reasoning), HumanEval (code generation), IfEval (instruction following), and LongFormQA (question answering) -- revealing massive efficiency gains at minimal performance impact. We report significant reductions in FLOPs across all tasks, with no statistically significant performance degradation, and significant performance improvements in 2 out of 4 tasks.

Vittorio Rossi, Giacomo Cirò, Davide Beltrame +3

Department of Computing Sciences, Bocconi University, Milan, Italy.

Adaptive Computation for Diffusion Models Diffusion Language Models Diffusion Language Model Inference Efficient Language Model Inference Diffusion Language Model Decoding

Mar 5, 2026

Reasoning Theater: Disentangling Model Beliefs from Chain-of-Thought

We provide evidence of performative chain-of-thought (CoT) in reasoning models, where a model becomes strongly confident in its final answer, but continues generating tokens without revealing its internal belief. Our analysis compares activation probing, early forced answering, and a CoT monitor across two large models (DeepSeek-R1 671B & GPT-OSS 120B) and find task difficulty-specific differences: The model's final answer is decodable from activations far earlier in CoT than a monitor is able to say, especially for easy recall-based MMLU questions. We contrast this with genuine reasoning in difficult multihop GPQA-Diamond questions. Despite this, inflection points (e.g., backtracking, 'aha' moments) occur almost exclusively in responses where probes show large belief shifts, suggesting these behaviors track genuine uncertainty rather than learned "reasoning theater." Finally, probe-guided early exit reduces tokens by up to 80% on MMLU and 30% on GPQA-Diamond with similar accuracy, positioning attention probing as an efficient tool for detecting performative reasoning and enabling adaptive computation.

Siddharth Boppana, Annabel Ma, Max Loeffler +5

Goodfire AI · Harvard University, Cambridge, MA

Faithfulness of Language Model Reasoning Language Model Reasoning Evaluation Faithfulness of Language Model Explanations Language Model Probing Efficient Language Model Inference

Mar 5, 2026

Transducing Language Models

Modern language models define distributions over strings, but downstream tasks often require different output formats. For instance, a model that generates byte-pair strings does not directly produce word-level predictions, and a DNA model does not directly produce amino-acid sequences. In such cases, a deterministic string-to-string transformation can convert the model's output to the desired form. This is a familiar pattern in probability theory: applying a function

f

to a random variable

X\sim p

yields a transformed random variable

f(X)

with an induced distribution. While such transformations are occasionally used in language modeling, prior work does not treat them as yielding new, fully functional language models. We formalize this perspective and introduce a general framework for language models derived from deterministic string-to-string transformations. We focus on transformations representable as finite-state transducers -- a commonly used state-machine abstraction for efficient string-to-string mappings. We develop algorithms that compose a language model with an FST to marginalize over source strings mapping to a given target, propagating probabilities through the transducer without altering model parameters and enabling conditioning on transformed outputs. We present an exact algorithm, an efficient approximation, and a theoretical analysis. We conduct experiments in three domains: converting language models from tokens to bytes, from tokens to words, and from DNA to amino acids. These experiments demonstrate inference-time adaptation of pretrained language models to match application-specific output requirements.

Vésteinn Snæbjarnarson, Samuel Kiegeland, Tianyu Liu +3

ETH Zürich · University of Copenhagen · CHI-FRO

Efficient Language Model Inference

Mar 2, 2026

How Small Can 6G Reason? Scaling Tiny-to-Small Language Models for AI-Native Networks

Emerging 6G visions, reflected in ongoing standardization efforts within 3GPP, IETF, ETSI, ITU-T, and the O-RAN Alliance, increasingly characterize networks as AI-native systems in which high-level semantic reasoning layers operate above standardized control and data-plane functions. Although frontier-scale large language models (LLMs) such as Qwen2.5-7B and Olmo-3-7B demonstrate strong reasoning capability, their computational footprint limits deployment in latency-sensitive, edge-native infrastructures. This paper presents a systematic empirical study of the scaling behavior and deployment efficiency of compact language models for network-level semantic reasoning in AI-native 6G systems. Using 6G-Bench, a standardization-aligned benchmark comprising 30 decision-making tasks across five capability domains, we evaluate models ranging from 135M (SmolLM2-135M) to 7B parameters (Qwen2.5-7B), including mid-scale architectures such as Llama-3.2-1B, Granite-1B, and Qwen2.5-3B. Deterministic accuracy (pass@1) increases from 0.224 at 135M to 0.707 at 7B, but scaling gains are highly non-uniform. A pronounced stability transition occurs in the 1 to 1.5B range, where accuracy rises from 0.373 (Llama-3.2-1B) to 0.531 (Qwen2.5-1.5B) and the instability gap Delta_5 contracts from 0.356 to 0.138. Beyond 3B parameters, improvements diminish (+0.064 from 3B to 7B). Through single-query inference profiling and an Edge Score metric that normalizes accuracy by latency and memory footprint, we show that semantic reliability per unit edge resource does not scale monotonically with parameter count. Instead, mid-scale models (approximately 1.5 to 3B) achieve the most favorable balance between deterministic stability and computational efficiency, providing deployment-relevant guidance for AI-native 6G architectures. All scripts and results are publicly available at https://github.com/maferrag/6G-Bench

Mohamed Amine Ferrag, Abderrahmane Lakas, Merouane Debbah

Department of Computer and Network Engineering, United Arab Emirates University, UAE · Research Institute for Digital Future, Khalifa University, UAE

6G Networks Small Language Models Language Model Scaling Wireless Communications Efficient Language Model Inference