LLM Inference Efficiency

Diffusion Large Language Models (dLLMs) have recently emerged as a promising alternative to autoregressive LLMs by enabling non-autoregressive text generation. However, their practical deployment remains limited by inefficient inference, largely due to the absence of effective Key-Value (KV) caching and scalable parallel decoding mechanisms. Existing acceleration methods typically study KV caching and parallel decoding in isolation, overlooking the I/O bottlenecks that arise when cache reuse and parallel token verification are jointly applied. In this work, we introduce $\textbf{Flash-dLLM}$, a training-free inference acceleration framework for fast and memory-efficient dLLMs. Flash-dLLM first identifies GPU memory I/O as a dominant bottleneck in KV-cache-enabled dLLM inference and addresses it with an I/O-aware fused KV-cache kernel that reduces redundant memory movement. Building on this optimized cache mechanism, Flash-dLLM further proposes an efficient KV-cache-driven draft-and-verify decoding strategy, where the dLLM itself serves as both drafter and verifier without requiring an auxiliary model. This unified design enables faster decoding while preserving generation quality and improving scalability to longer sequences and larger batch size. Extensive experiments on mathematical reasoning and code-generation benchmarks demonstrate that Flash-dLLM consistently outperforms existing state-of-the-art dLLM acceleration methods in both inference speed and memory efficiency. In particular, it achieves $5.1\times$ and $11.0\times$ speedups over prior strongest baseline Elastic-Cache on GSM8K and HumanEval, respectively.

Greedy decoding from large language models is commonly treated as deterministic. We show it is not precision-invariant: the same model, prompt, and decoding algorithm produce different outputs in BF16 versus FP16 on identical hardware. Across our evaluations of six models (1.1B-7B parameters, four families; divergence additionally characterised at 12B) and three benchmarks, 49-100% of prompts diverge; a single token flip often cascades into trajectory-level divergence. We develop an empirical error-propagation analysis and find that 22 layers of accumulated body error do not distinguish flipping from non-flipping steps; the outcome depends primarily on the top-two logit margin at the LM head relative to the directional perturbation between the top-two candidates. The analysis makes five testable predictions about intervention outcomes, including that applying more FP32 compute (broader scope) makes agreement worse. The experiments match all five predictions. The best-performing low-overhead intervention we evaluate, selective FP32 LM head recomputation, triggered only when the margin falls below a threshold, delivers +22-36 pp exact agreement on A10G (+12-21 pp on L4 and A100) at less than 4% latency overhead in low-batch (batch size <=4) single-stream inference. We map the applicability boundary across six models and four batch sizes, and hypothesise that training-time precision stability is a determining factor. The method is a partial mitigation rather than a universal determinism guarantee: its benefit vanishes when body-originated error dominates, including at batch size >=8 and under end-to-end FP8 in our tests.

Prefill and decode reward different approaches to quantization: low-precision arithmetic accelerates prompt processing, while compact weights reduce memory traffic during generation. We propose "disaggregated quantization" (DQ), which specializes computation formats, weights and storage placement to both of these phases. On Qwen 3 and Gemma 3, removing activation quantization specifically on decode improves accuracy on decode-heavy tasks without increasing inference cost. Training separate compute-native prefill weights accelerates prompt processing relative to weight-only inference while matching or exceeding its accuracy at 2-3-bit decode on both decode-heavy and prefill-heavy tasks. With released Qwen3.8-27B GGUF decoders, training an NVFP4 prefiller improves 1-bit accuracy by 32.5 points on MMLU-Pro and 35.3 on MMMU-Pro without modifying the decode checkpoint. To accommodate the additional checkpoint on a single device, offloaded disaggregated prefill (ODP) streams its weights from SSD, amortizing loading over prompt length. On the same 27B model, ODP delivers a 1.78x time-to-first-token speedup over the weight-only baseline at 8K prompt length in llama.cpp. We evaluate accuracy under disaggregated serving in vLLM and further validate shared-weight format disaggregation through post-training quantization on models up to 2.8T parameters.

Despite their strong performance, large language models (LLMs) are bottlenecked by KV cache memory traffic during long-context inference. Sparse attention is widely used to accelerate LLM inference by computing exact attention over a selected subset of tokens. To recover the contribution of tokens excluded from exact attention, recent methods apply coarse-grained compensation to the omitted attention tail. However, existing methods typically select tokens based on attention mass and only then compensate for the unselected tokens. This decoupled design overlooks their interaction: selection should prioritize tokens that would leave the largest compensation error if omitted. To address this limitation, we introduce CompKV, the first compensation-aware sparse attention framework that divides tokens into blocks and explicitly optimizes selection for the downstream compensation mechanism. Our theoretical analysis shows that the residual left by block-level mean compensation is governed by both block attention mass and within-block logit variation. We approximate this residual using compact block-level statistics, yielding a deployable selection criterion. We further develop an efficient asynchronous implementation. Experiments on RULER and LongBench-Pro show that CompKV performs best among the evaluated sparse baselines while delivering up to a $6.85\times$ self-attention speedup over full attention.

Large language model (LLM) outputs are expected to be reproducible under greedy decoding, yet in practice the same model, prompt, and software stack produce different outputs on different GPUs. The root cause is floating-point non-associativity combined with hardware-dependent kernel selection. Inference frameworks select different matrix-multiplication kernels on each architecture, with different parallel reduction orders and unspecified tensor-core arithmetic, and the resulting rounding differences can flip output tokens. Existing solutions have imperfect cross-architecture reproducibility and incur a significant performance penalty. We present a solution employing a set of fixed-configuration fused-upcast GEMM kernels that load 16-bit weights from memory, upcast them to FP32 in registers, and accumulate with IEEE-754 arithmetic in a reduction order that is a pure function of the problem shape and is therefore independent of the device, its SM count, or kernel scheduling. By fixing the floating-point reduction order as a function of problem shape alone, every GPU runs the same operation sequence, so cross-architecture reproducibility of the linear layers reduces to correct IEEE-754 arithmetic rather than to rounding differences staying below a tie-flip threshold. We confirm our solution's linear-layer outputs are bitwise identical across NVIDIA Ampere, Ada, and Hopper GPUs, while running $1.17$ to $3.1\times$ faster end-to-end than the state-of-the-art solution and cutting weight-memory traffic in half.

Large language model (LLM) inference is constrained by the quadratic scaling of self-attention and the linear scaling of the KV cache, increasing latency, energy consumption, and GPU memory demand as context length scales. Existing soft-compression methods either lack query-guided memory selection at inference time, train without answer-targeted supervision, or couple compression tightly to a specific decoder architecture. We propose a Context-to-Answer-Aligned Memory Compression (CMC) framework, which compresses long input contexts into compact Context Memory Embeddings (CMEs) aligned to any frozen decoder's embedding space, reducing inference costs without modifying decoder weights. CMC introduces a two-tier KV cache that combines question-guided CME selection with a local context window, and trains the compressor with answer-targeted distillation from a frozen LLM. Experiments across nine encoder-decoder combinations and four QA benchmarks show that CMC consistently outperforms the baseline, achieving up to 7.3 EM and 4.0 F1 point gains on SQuAD, while reducing inference time and energy consumption by up to 20% and peak reserved GPU memory by up to 50% at 3,000 generation tokens. Ablation studies confirm that each architectural component and training objective contributes to the performance.

When a language model reads an operation such as "Swap the contents of Box F and Box B", its forward pass writes keys and values for those tokens into the KV cache. Prior work on entity tracking establishes what models use: bindings are resolved at query time rather than stored as explicit latent state. We ask what they write at the operation span and how it is accessed. We split a forward pass into a frozen writer and a reader: the writer's cache is recomputed without gradients, while the reader sees only the instruction and operation tokens, with all state descriptions hidden, and is trained in isolation. Anything the reader recovers was therefore already present in the unmodified cache. On a synthetic boxes task, a base reader recovers $\leq 0.06$ of queried bindings against $0.75$--$1.00$ after training, and recoverability tracks the operation's read/write footprint. We find two modes of access. Across Llama-3.1-8B and Mistral-7B, operation-span transplants causally redirect which visible state is read even when the two worlds hold identical values, revealing a routing record. Isolation training preserves routing and adds direct access to the payload, the value the operation read, from the single operand-name token in a narrow mid-depth band (layers 12--15 of 32 in Llama-3.1-8B, 14--17 in Mistral-7B) --- the same site that holds the routing record. The same recipe extends to further operations, ToMi and GSM8K, but is bounded by training coverage and costs open-book accuracy. Operation tokens thus leave localized, causally recoverable records that support both routing and direct payload access, though the model that writes them reads mainly the address they carry and not the value.

Despite advances in long-context inference, large language models (LLMs) remain fundamentally limited by the key-value (KV) caching mechanisms that are necessary for stable computation. Techniques such as selective token eviction and pruning have vastly mitigated these issues, but often discard core information to manage the growing cache. In this paper, we propose Attention with Routed Memory (ARM) a novel KV caching structure that introduces a fully differentiable, fixed-size memory system organized as a hierarchical router. Via a Gumbel-Softmax, ARM learns to select memory slots and perform sigmoid-gated updates that softly combine new and stored information, avoiding hard eviction and reducing information loss. By further training a policy to dynamically select varying amounts of memory at inference, ARM adapts its accesses for both simple contexts and inputs that require deeper reasoning, enabling more scalable and effective retrieval on both short- and long-contexts. Experimental results on standard commonsense and long-context reasoning benchmarks demonstrate that ARM achieves superior performance and efficiency compared to fixed KV-caching approaches, while remaining efficient and scalable in terms of both memory and generation latency.

Speculative decoding losslessly accelerates large language model inference by having a lightweight draft model predict future tokens for verification by the target model. Recent block diffusion drafters further reduce drafting latency by predicting multiple tokens in parallel. However, existing block drafters project target hidden states at every input position into a separate drafter-side KV cache, incurring per-request memory and KV-write overhead that grow with concurrency; directly reusing target KVs in place removes this cache but fails to sustain draft quality throughout the block. We propose a hybrid target-context injection method that complements direct target KV reuse with target hidden states only at the last input position, requiring no separate drafter-side KV cache. Building on this design, we propose H-Spec, a hybrid Mamba-attention parallel drafter that consumes the two target-context sources through complementary modules. Mamba modules are initialized with projected last-token target hidden states, while attention modules reuse target KVs in place. Despite its recurrent formulation, Mamba's parallel scan allows H-Spec to preserve block-parallel drafting. Across three target models and diverse tasks, H-Spec improves over the best baseline by 5.0--13.3% in mean accepted length and 5.3--12.6% in batch-size-1 inter-token latency speedup. Under concurrent serving, H-Spec consistently achieves higher throughput while maintaining lower KV cache utilization than baselines across evaluated concurrency levels.

Reasoning and agentic workloads increasingly demand efficient long-context inference. Yet full-attention decoding reads the growing history at every step, regardless of its benefit to the next prediction. We show that a pretrained model's decoding states already contain information predictive of this benefit, before the global read. Building on this finding, we introduce On-Demand Attention (ODA), a local-first decoding method that uses a lightweight recall head to selectively invoke global attention as its predicted benefit changes during generation. ODA trains only the recall head, leaving pretrained weights unchanged and the complete historical KV cache available for future recall. We further implement GPU-side conditional execution in vLLM, translating reduced global reads into practical decoding speedups over full attention at long context lengths. Experiments across Qwen and Gemma models, including hybrid-attention backbones, show that selective recall recovers most of the performance lost under local attention while substantially reducing global reads. These findings support long-context inference in which pretrained models guide their own access to the information they retain.

Speculative Decoding (SD) has significantly accelerated Large Language Model (LLM) inference, yet existing approaches face a fundamental tradeoff between two drafting strategies: neural drafting and context-based copying. Neural drafts (e.g., EAGLE3) provide robust performance across diverse text settings, while copy-based methods achieve higher speedups in copy-intensive regimes by generating candidates faster and exploiting long repetition spans for near-perfect speculation. We analyze existing copy-based methods and find that they are prone to accidental repetitions where surface-level n-gram overlap does not reflect a structural intent to copy, leading to false-positive triggers that ultimately degrade throughput. We introduce SwitchSD, an adaptive framework that treats copying as a latent control signal of the LLM. By training lightweight probes on the target model's internal representations, SwitchSD identifies genuine copy-intent with high precision (AUC > 0.99). This allows the system to dynamically switch between neural drafting (e.g., EAGLE) and context-based copying. Our results across Llama and Qwen families demonstrate throughput gains of up to 15% over state-of-the-art baselines like EAGLE3, effectively turning copying from a noisy heuristic into a principled, model-aware decoding regime.

Per-user LLM inference on transaction histories binds the inference budget linearly to user count, which becomes prohibitive at applied scale. We re-cast attribute inference from per-user to per-transaction-pattern. The pipeline runs in three phases: Resolve abstracts item names with optional web grounding, Profile infers attributes for each frequent pattern, and Tag clusters free-text attributes into a queryable database. In Profile, a single LLM call per pattern emits predefined categorical labels, free-text attributes, and per-attribute prevalence estimates. Because inference runs over patterns rather than users, the budget grows with the pattern count rather than the user count. On the public Open e-commerce corpus, the database is statistically indistinguishable from an LLM that reads each user's raw history directly in AUC across the evaluated attributes, and the prevalence estimates carry discriminative signal between positive and negative users. The pipeline is deployed at a major Japanese bank profiling on the order of tens of millions of users, with close to a three-order-of-magnitude reduction in LLM inference targets versus a per-user pipeline. The code is publicly available on https://github.com/CyberAgentAILab/profiling-agent-open-ecommerce.

Travel-planning agents generate itineraries that may become infeasible after acceptance because of flight cancellations, hotel unavailability, or attraction closures. Revising these itineraries involves full replanning, classical plan repair, and LLM-based travel-agent revision, whose differing task formulations and evaluation protocols hinder comparison. We conduct a systematic empirical study using two TREK-derived benchmark sets: 500 single-disruption cases, including feasible and infeasible instances, and 200 feasible simultaneous compound-disruption cases. We compare LLM-Z3 full replanning, IPyHOPPER hierarchical repair, and an iTIMO local-revision adapter across effectiveness, plan stability, and computational cost. LLM-Z3 with Gemini achieved the highest observed compound-disruption success. IPyHOPPER nearly matched that configuration's single-disruption overall success, while preserving substantially more of the accepted itinerary on successful repairs. Successful hierarchical and local repairs made fewer edits and retained more accepted commitments than full replanning. Computational profiles differed: IPyHOPPER used no LLM inference, the evaluated LLM-Z3 adapter used compact one-call inference, and the iTIMO adapter consumed substantially more tokens. The study provides practical guidelines for balancing feasibility recovery, commitment preservation, and computational cost within evaluated settings.

Modern applications generate massive volumes of raw telemetry data, but translating those noisy, heterogeneous event streams into actionable business insights remains a fundamental challenge. Data engineers and analysts expend substantial effort reconciling semantic discrepancies, hand-crafting parsing logics, and maintaining fragile mappings between raw data and business KPIs. In this paper, we present an end-to-end framework that fully automates the construction of a business semantic layer from application raw logs. Our approach introduces a two-stage semantic abstraction: first, high-level business features are identified via LLM inference augmented with domain-specific industry knowledge; second, fine-grained business nodes are derived through a structured pipeline comprising data refinement, hybrid retrieval, multi-stage filtering, semantic clustering, and canonical naming. Evaluation on production-scale telemetry demonstrates that our system improves human-assessed semantic quality from 50 to 80+ on a 100-point scale, reduces maintenance effort by 80%, filters out 74% of noise, and achieves 0.87 Cohen's kappa via an integrated LLM-as-Judge evaluation, enabling continuous, scalable quality assurance. Overall, our work distinguishes itself from prior work by addressing the novel problem of business semantic layer induction from raw telemetry, operating without labeled training data or manual rule engineering.

We introduce Mahalanobis-Ensemble Decoding (ME-Decoding), a novel Large Language Model (LLM) decoding framework that frames candidate token selection as ensemble pruning. Existing selection strategies rely predominantly on scalar probabilities, ignoring geometric semantic relationships and causing candidate redundancy. Meanwhile, current geometry-aware methods often require complex optimization or directly reweighting the original token probabilities, leading to significant computational overhead or inference instability. To address this, we formulate decoding as a subset optimization problem using a Mahalanobis distance-driven objective to enhance semantic diversity while preserving high probabilities. Specifically, we dynamically discount redundant generation paths using a token similarity matrix, constructed via an adaptive-bandwidth kernel over token embeddings. We further devise an efficient greedy selection algorithm with near-linear complexity in the candidate size under early stopping, while establishing its theoretical approximation guarantees. This renders ME-Decoding a robust, plug-and-play module with negligible inference overhead. Extensive experiments across diverse reasoning and generation tasks demonstrate that our method consistently achieves strong performance.

Long-context LLM inference is bottlenecked by attention, whose repeated KV-cache reads make decoding memory-bound. Self-speculative decoding alleviates this by drafting tokens with sparse attention and verifying them with full attention, but existing batched methods remain synchronized: all requests in a batch share a single draft-verify schedule, even though the optimal draft length varies widely across requests and changes dynamically within each request. We propose ASPIRE, a non-synchronized batched self-speculative decoding framework built on three components. First, a unified mixed forward allows drafting and verifying requests to coexist in the same batched forward pass, removing the need for global draft-verify phases. Second, a lightweight online speculation scheduler uses per-request acceptance-rate estimates and a batch-aware cost model to let each request independently choose when to verify. Third, an intra-draft refresh layer performs full attention at a single designated layer during drafting, updating the sparse context at every draft step to reduce staleness during drafting. Across three models and five reasoning and long-context benchmarks, ASPIRE achieves $1.70$-$4.58\times$ speedup in decoding throughput over autoregressive baselines and improves average speedup by approximately $27\%$ over the strongest prior self-speculative baselines.

We study the problem of demonstration selection, which involves selecting a subset of examples for prepending to a query to a language model. This problem is closely related to in-context learning and language model inference. Since the inference cost of a transformer model scales quadratically with sequence length, the selection problem becomes especially challenging in a long-context scenario. In this paper, we tackle this problem by building on state space models (SSMs), which require only linear inference time given the input. Our approach involves two algorithms. The first learns a small set of SSMs through distillation of a (trained) transformer model. We partition all the layers into consecutive groups. Then for each group, we estimate a separate state space model to replicate the input-output behavior within the adjacent layers. Second, we map the distilled model outputs to a small set of tokens, and apply these embeddings for demonstration selection in downstream applications. We perform extensive experiments in both synthetic and real-world datasets to validate our approach. We demonstrate that the distilled SSMs only incur an approximation error of less than $0.7\%$ relative to the true output. In downstream evaluation, we show that on several text classification and reasoning tasks, our approach reduces FLOPs by $14.2\times$ and improves accuracy by $6.48\%$ relative to baseline demonstration selection methods.

LLM inference optimizations report speedups on different models, GPUs, prompts, and quality metrics, making them hard to compare or combine. We build a cost, quality, and latency Pareto atlas to identify the best configurations for different deployment constraints. Since exhaustive testing is impractical, we measure 54 configurations of Qwen2.5-7B-Instruct running on vLLM 0.12 across L4, A100, and H100 GPUs and use these anchors to calibrate a simulator. It reproduces measurements at anchored batch sizes, with cross campaign drift below 1.5 percent. A separate quality evaluation tests FP16, AWQ 4bit, FP8 weights, and FP8 KV cache on 200 GSM8K questions with five examples per prompt. Sparse attention is evaluated only in simulation. On the calibrated grid, 18 of 36 configurations reach the Pareto frontier. Combined methods reach it more often than individual methods, with 9 of 15 combinations versus 9 of 21 single methods. Quality testing changes the winners. AWQ 4bit reduces per token latency to 0.34 times baseline on L4 but loses 5.9 percent of strict GSM8K accuracy, narrowly missing the 95 percent quality floor within sampling uncertainty. Flexible answer extraction matches FP16 accuracy, suggesting the loss comes from formatting rather than arithmetic. FP8 weights retain 99.4 percent of baseline accuracy at 0.61 to 0.65 times baseline latency across all three GPUs and appear in three of four regime winners. A naive FP8 KV cache maintains normal throughput but answers none of the 200 questions correctly, showing why speed alone is insufficient. Under two prompt designs, n gram speculative decoding measures at 0.90 to 0.98 times baseline and adds no benefit on this stack. The best choice depends on the constraint and GPU: H100 wins for tight latency, while A100 wins for throughput and low cost at 0.106 dollars per million tokens.

While draft-model-free speculative decoding offers a promising path to efficient LLM inference, it is frequently constrained by stale draft candidates and the high computational cost of the verification. To address these challenges, we propose ECHO, a hierarchical dual-loop framework that exploits the functional asymmetry between LLM layers. Leveraging the high discriminative efficiency of early layers and the authoritative distribution of final layers, ECHO bifurcates inference into a high-frequency inner loop and a low-frequency outer loop. Within the inner loop, early-layer bonus logits drive rapid, multi-step draft-tree exploration at a minimal cost. Simultaneously, the outer loop performs authoritative full-model verification through a state-reuse mechanism. Crucially, the outer loop also utilizes final-layer bonus logits to correct existing paths and supplement the tree with high-confidence candidates for subsequent cycles. Experimental results across diverse benchmarks demonstrate that ECHO significantly boosts mean accepted tokens and achieves a 2.4$\times$ to 2.9$\times$ speedup, outperforming existing state-of-the-art baselines with negligible engineering overhead and no extra deployment parameters, albeit with a one-shot fine-tuning dependency for optimal acceleration. The code is available at https://github.com/whucs21Mzy/ECHO.

Diffusion large language models (dLLMs) offer a promising parallel decoding paradigm as an alternative to autoregressive generation through iterative unmasking. However, dLLMs typically require many steps before token confidence reaches the decoding threshold, resulting in inefficient inference even with block-wise KV caching. To accelerate dLLM inference, we for the first time propose an "early-bird (EB)" decoding framework, motivated by the observation that tokens with similarly low entropy tend to cluster and can be jointly decoded earlier, before reaching the confidence threshold. In particular, our EB-Decode framework integrates two key enablers: (1) a learnable network that adaptively groups tokens with similar uncertainty into variable-length blocks, rather than relying on fixed block sizes; (2) a position-aware sampler that learns to unmask tokens in parallel using fewer decoding steps within predicted variable-length blocks. Both components are developed without modifying pretrained dLLM weights and can therefore be directly deployed as plug-ins during serving, with negligible training and inference overhead. Extensive experiments across three models and four benchmarks consistently validate our observation and the effectiveness of EB-Decode, achieving 3.53-18.76$\times$ higher throughput than the vanilla decoding method and up to 1.58$\times$ higher throughput over the strongest baseline, Fast-dLLM, with comparable accuracy.

Large Language Models (LLMs) have shown impressive capabilities across a range of natural language processing tasks, and LLM inference has emerged as a critical workload for enabling downstream applications. The demands of serving LLM inference are becoming increasingly challenging as requests shift toward longer sequences and heavier inference, driven by retrieval-augmented generation, inference-time compute scaling, and long-context applications. Additionally, these challenges are compounded by hardware trends, as memory capacity and communication bandwidth are not scaling as fast as increases in workload complexity. Compute-in-Flash is a promising solution to address memory bandwidth limitations by moving computation close to memory, and to exploit the large capacity of SSD technologies. However, it is challenging to deploy LLMs on these systems as they lack support for high-precision floating point operations and have limited write endurance. In our work, we aim to address these challenges by designing inference algorithms to enable LLM inference on Flash compute-in-memory devices. We present an end-to-end integer-only quantization approach to eliminate expensive floating-point computations. To address the limited write endurance, we design a dictionary-based KV cache compression strategy based on sparse dictionary coding that represents each KV vector as a linear combination of static dictionary vectors. These algorithmic improvements enable us to exploit the benefits of Compute-in-Flash for both model weights and KV cache, and to minimize expensive data transfer operations. Across Llama-3.1-8B and Qwen-2.5-7B, our combined method exhibits limited accuracy degradation while reducing dynamic KV cache traffic by 15$\times$.

Orthrus is a hybrid autoregressive-diffusion architecture that accelerates autoregressive language-model inference by generating multiple tokens in parallel while using a frozen autoregressive backbone. Its central claim is that an intra-model consensus mechanism enables lossless speculative decoding, producing the same output sequence as the autoregressive model. We independently reproduce Orthrus and examine this claim under different numerical precisions. Under BF16 inference, exact trajectory matching occurs in only 45% of cases for the authors' checkpoint and 43% for our independently trained model across 1,190 prompts from 12 domains. The probability of exact matching is also strongly associated with the response-conditional perplexity of the reference model. Despite this trajectory divergence, Orthrus does not show systematic degradation on downstream lm-eval-harness benchmarks. In contrast, repeating the trajectory evaluation with FP32 yields exact trajectory matching on all evaluated prompts. These results show that the practical losslessness of Orthrus depends on numerical precision and that exact trajectory equivalence should be evaluated separately from downstream task performance.

Large Language Models (LLMs) primarily perform inference at the token level, resulting in substantial memory overhead and compromised computational efficiency. In this paper, we propose a Dynamic Semantic Extraction and Inference (DSEI) framework, which achieves segment-level inference within the latent space through a two-stage training strategy. First, we construct a Dynamic Semantic Autoencoder (DSAE) via self-supervised learning. DSAE dynamically extracts segment-level semantics and compresses them into compact latent representations via adaptive semantic weighting and gated fusion. Subsequently, we integrate the DSAE into the LLM architecture and train the model to infer over dense latent space. DSEI substantially reduces both input and generation sequences and significantly enhances inference efficiency. Extensive experiments conducted on the Wanjuan dataset demonstrate that DSEI reduces perplexity by 48% compared to static sentence-level latent inference baseline. Furthermore, compared to standard LLMs using token-level inference, DSEI accelerates inference speed by 2.5$\times$ and reduces memory overhead by 90%.

Multimodal large language models (MLLMs) require substantial computation to process numerous visual tokens across all transformer layers. Most methods for efficient MLLM inference exploit horizontal redundancy by compressing visual tokens. Beyond token reduction, recent studies exploit vertical redundancy through early exit or fixed-layer skipping. However, we find that the extent and distribution of this redundancy vary across inputs and differ between self-attention and MLP modules. Motivated by these observations, we propose AdaVSkip, which equips each layer with two lightweight routers that independently determine whether visual tokens pass through by or skip the self-attention and MLP modules. These decisions collectively define an input-specific visual-computation path, but their discrete and non-differentiable nature makes learning effective paths challenging. To address this challenge, we develop a progressive two-stage training framework that updates only the routers while keeping the backbone frozen. Stage I establishes an initial routing policy through supervised training with input-specific targets derived from module-wise necessity scores. To further align the routing policy with task performance, Stage II uses reinforcement learning to optimize routing decisions with direct feedback from generated answers. It combines an answer correctness reward with a skip-consistency reward that discourages excessive retention of visual-token computation. Across three MLLM backbones, AdaVSkip maintains strong task performance with substantially less computation. On LLaVA-NeXT-7B, AdaVSkip reduces FLOPs by 53.2% while preserving the original model's average performance. Combining it with visual token compression increases this reduction to 91.2%, while retaining 97.2% of the original performance on average.

User prompts provided to large language models (LLMs) may contain sensitive or private information that can be misused by remotely deployed models, such as through inadvertent memorization during retraining. One way to protect user prompts is to execute the LLM inside a trusted execution environment (TEE), with the guarantee that the service provider has no access to computations performed within or information exchanged with the TEE. However, current TEEs are primarily CPU-based and significantly slower than GPUs optimized for LLM inference. To circumvent this, Tramer and Boneh (2019) proposed Slalom, which splits neural network inference between a TEE and an untrusted GPU and encrypts intermediate inputs sent to the GPU. We extend this split-inference architecture to LLM inference and instead protect intermediate inputs using differential privacy. We show that masking intermediate representations is necessary by showing that a prompt-reconstruction attack can recover prompts from these representations with nearly 80% accuracy. Our main contribution is a global sensitivity analysis of key LLM functions, which bounds the required scale of differentially private noise. Unlike encryption, differential privacy avoids quantization, allowing the LLM to remain in the floating-point domain. We also derive an upper bound on floating-point error from masking and noise cancellation in the TEE as a function of the privacy parameter epsilon. We implement our architecture using Intel TDX and evaluate it with two LLMs: Llama-3.2-3B and Qwen3-4B. Our split execution is nearly twice as fast as fully CPU-based inference inside TDX and 5-15 seconds faster than encryption-based Slalom while achieving higher accuracy. Finally, we demonstrate that prompt reconstruction, even with knowledge of the differential privacy mechanism, cannot recover more information than is contained in an unrelated prompt.

Shared host-memory caching can avoid repeated prefill when a request moves between inference replicas. Its usefulness depends on both correct state transfer and lost prefix locality. We study two single-GPU 27B vLLM replicas sharing a 256 GiB LMCache pool. After adopting an existing packed-page patch, we isolate a raw-pointer fallback that omits the dependency on the current CUDA stream. Controlled byte tests fail under an imposed delay and pass when the dependency is restored; the existing mixed allocator provides a working deployment path. Full-pool allocation checks and service regression complete the validation. A four-block OFF-ON-ON-OFF comparison contains 768 measured requests within two block pairs. Median cross-replica time to first content token falls from 31.715 to 0.605 seconds at 128k input and from 92.047 to 0.790 seconds at 256k. Six-turn synthetic sessions alternating replicas improve by approximately 35% and 45% at initial contexts of 32k and 128k, while fixed placement shows little benefit. This engineering case study identifies practical validation steps and the locality conditions in which shared caching pays off.

The rapid diffusion of generative artificial intelligence raises privacy, latency, and performance concerns that motivate a shift toward "local-first" AI, where inferences are performed on the user's device instead of on remote cloud servers. This paradigm also places a significant computational load on battery-powered smartphones, potentially shortening battery life and increasing the overall replacement rate of mobile devices. This paper presents a systematic study of the energy consumption, performance, and accuracy of on-device large language model (LLM) inference. We evaluate 18 models from different model families, sizes, and quantization levels, on two modern smartphones and on a server, using the respective state-of-the-art for such deployments. We measure the energy per generated token, inter-token latency, model accuracy, and battery-cycle consumption. Our results show that (i) on-device inference is on average 3 times less energy-efficient than batched server inference; (ii) the relationship between quantization bit-width and energy per token is non-monotonic, with energy sweet spots on both tested smartphones; (iii) eight out of 18 model configurations lie on the Pareto front of accuracy and energy-efficiency, allowing practitioners to build battery-aware model routers; and (iv) realistic modeling assumptions do not allow local inference to be less environmentally impacting per token than batched server inference, with 88--90% of that impact attributable to device embodied carbon rather than electricity consumption. These findings challenge the premise that local AI is more sustainable than cloud inference, and motivate the need for context-aware and life-cycle-aware model selection when deploying edge AI on battery-powered mobile platforms.

A single SM utilization percentage can make an LLM inference workload look compute-saturated while hiding how much useful work is being done. The problem is not that the counter is wrong, but that it collapses several different mechanisms into one number. This is most severe during decode, where each request contributes only one new token and dense projection GEMMs become small-row matrix multiplications. On Hopper, the bfloat16 GMMA path executes these operations in fixed 64-row matrix fragments, so small-batch decode can fill only a small fraction of each fragment with real token rows. In this paper, we profile vLLM with FlashAttention-3 and cuBLASLt on an H100 NVL across cold prefill, warm prefill, and decode, sweeping sequence length and batch size. We replace the usual single utilization number with eight counter-validated views derived from raw Nsight Compute reports, each pinned to an NCU counter or explicit formula. Together, these views map utilization gaps to concrete mechanisms - fragment fill, occupancy limits, stall signatures, wave quantization, and kernel selection - across four production models and six per-layer kernel roles.

Diffusion language models decode tokens in parallel, but their bidirectional denoiser rules out the naive key--value (KV) cache behind fast autoregressive inference. Block diffusion restores caching by decoding block-by-block, and the block caches deployed on it so far are tied to attention: O(L)in memory and, if used as training-free retrofits, only an approximation of the model's computation. Both constraints can be overcome: sequence mixers that summarize finalized blocks into a reusable state support block caching, and the corresponding block-causal training objective makes the cache exact. We study this recipe at scale, pretraining three 3B block-diffusion denoisers (attention, mamba, and hybrid) on 300B tokens under one single-frontier objective and decoding all three through a single cached interface. Only the state-space cache is O(1) in sequence length: its memory and per-step latency stay constant at any context length, while an attention cache remains O(L). At 256k tokens (where attention has grown to 82GB and 29 ms/step), the Mamba cache delivers 4.3x lower latency, 11x less memory, and 2.6x higher single-stream throughput; and because that footprint is constant it scales with batch as well, reaching 14x the aggregate throughput, where attention cannot run beyond a single stream. The same linear-state bias lets the Mamba and hybrid backbones keep retrieving out to 8-16x their training length, whereas attention's retrieval collapses at 2x, at no measured quality cost.

AI is beginning to make substantive contributions to LLM inference optimization. Existing AI optimizations are predominantly profiling-based. Profiling-bound feedback confines the search to the capabilities and performance of an existing software stack, preventing a fundamentally better architecture of LLM inference systems from being identified. To enable the AI-driven LLM inference system architecting loop, we argue that a general workload representation, a verifiable mutation space, and an implementation-independent evaluator are required. We present the RoofLang domain-specific language (DSL) that provides these features. In our evaluation, RoofLang reveals that DeepSeek V4-series models could achieve 3.5-39.5$\times$ higher peak decode throughput than other representative models. This gap is disproportionate to their total parameter counts and arises largely from compact KV-cache designs that support larger batches and reduce memory traffic. A persistent optimizer agent further discovered several new architectures that improved both throughput and interactivity of DeepSeek V4 Pro on NVIDIA B300 by 6.23-50.1%.

Large language models (LLMs) demonstrate impressive capabilities, but their deployment presents significant efficiency challenges. Autoregressive decoding imposes substantial inference latency and under-utilizes hardware accelerators in low batch size regimes. Discrete diffusion models can generate in parallel but struggle to match autoregressive quality without many diffusion denoising steps. Long-context reasoning creates memory bottlenecks that strain even state-of-the-art accelerators. My thesis is that language models can direct their own inference execution strategy by annotating semantic dependence -- which tokens depend on which others -- in their generation. I call such models self-orchestrating language models. For each system, I design a runtime that acts on these annotations to parallelize autoregressive decoding, evict intermediate context, or derive denoising orders, achieving Pareto-optimal quality-efficiency trade-offs. I demonstrate this approach through three self-orchestrating systems. First, PASTA uses semantic dependence to parallelize autoregressive decoding, training the model to annotate which output chunks can generate independently. Second, TIP uses semantic dependence to evict intermediate reasoning steps from the KV cache, reducing memory consumption while preserving accuracy. Third, Planned Diffusion uses semantic dependence to derive a denoising order for discrete diffusion, autoregressively generating a plan that specifies which chunks to denoise in parallel.

Soft context compression condenses a context into a few memory tokens that a frozen LLM consumes in place of the raw text, but existing compressors fix the compression ratio at training and inference: each deployed ratio requires a separately trained model, and the chosen ratio is applied uniformly to all inputs, whose actual needs vary drastically. We propose FlexComp, a method-agnostic framework that decouples the ratio from both training and deployment: Matryoshka-style training samples the memory budget $K$ per instance, turning one model into an any-ratio compressor, and the budget is then chosen per input by: (1) confidence-based cascade routing or (2) a lightweight learned $K$ predictor. Across ICAE, 500xCompressor, and SAC on MRQA, a single FlexComp model matches separately trained fixed-ratio specialists with minimal degradation. Cascade routing preserves over 98% of the mildest ratio's accuracy at up to 266x average compression; the $K$ predictor, in a single compression-decoding pass, reaches 158-236x within 0.7 F1 of the mildest ratio. At serving-scale batch sizes, the $K$ predictor cuts context KV cache by 50% and improves decoding throughput by 47%.

Recent advances in multimodal large language mod- els (MLLMs) have opened new opportunities for edge intelligence by enabling reasoning across heterogeneous sensor modalities, such as vision, text, and telemetry data. However, deploying these capabilities on resource-constrained edge platforms remains challenging due to the substantial computational, memory, and communication demands of modern MLLMs. Rather than transmitting raw sensor observations or partitioning neural networks at intermediate layers, Edge Multi-Modal Intelligence (EMMI) communicates a compact representation between edge devices and server resources, enabling communication-efficient edge MLLM inference. To achieve this, EMMI performs modality-specific encoding, cross-modal representation fusion, and learned compression at the edge, transmitting only a compact latent representation to server-side resources for high-capacity MLLM reasoning. This representation-centric design reduces communication overhead, preserves local data privacy, and provides a fixed-size interface between heterogeneous edge devices and server-side MLLMs. Evaluation on a representative multimodal benchmark demonstrates that EMMI can reduce the communication payload by 32x while maintaining comparable downstream accuracy, resulting in up to a 3.4x reduction in estimated end-to-end inference latency under bandwidth-constrained edge conditions.

Large language models are often fine-tuned, shared, or downloaded from third parties, so a deployed model may carry a hidden backdoor that behaves normally on benign inputs but switches to attacker-controlled behavior when a secret trigger appears. While backdoors can be audited before deployment, runtime monitoring remains important for models that are frequently updated. The challenge is that LLM serving is latency-sensitive: existing inference-time detectors either rely on assumptions about the trigger form, which can fail on stealthy attacks, or require extra model computation, such as input perturbations or an additional generation pass. We introduce SpecGuard, an inference-time backdoor detector that repurposes speculative decoding at zero added model-computation cost. Speculative decoding speeds up inference by using a small draft model to propose tokens and a target model to verify them. We observe that this verification process already exposes a useful signal: when a backdoor is triggered, the target model shifts toward the attacker's behavior, while a clean draft model does not predict this shift, causing the draft-token acceptance rate to change. We formalize when this signal appears and show that an attacker who suppresses it must also weaken the backdoor. Across diverse backdoor types and model families, SpecGuard reliably detects triggered behavior, including stealthy cases where input-level filters are blind, while avoiding the extra generation cost of existing runtime detectors. Speculative decoding therefore doubles as a free, always-on signal for detecting backdoored LLM behavior.

As Omni-modal large language models (Omni-LLMs) take in audio, video and text together, their KV cache memory cost grows. KV cache quantization is the de facto approach in text-only LLMs, but its application to Omni-LLMs remains unexplored. In this paper, we analyze how TurboQuant, a representative rotation-based KV cache quantization method, behaves on multimodal caches and identify two critical issues: temporal key drift and heterogeneous value geometry. To address these, we propose OmniKVQuant, a training-free framework that (i) sets the key quantization range over each short window of the input stream; and (ii) rotates values separately per modality. On Qwen2.5-Omni and Qwen3-Omni, OmniKVQuant enables 2-bit KV caches while substantially preserving performance across seven audio-visual benchmarks. We further provide a fused Triton decode kernel that unpacks the 2-bit cache during attention, so no dense FP16 cache is ever built. Code: https://github.com/kaistmm/OmniKVQuant

Open-source large language models (LLMs) are increasingly competitive with closed-source models while offering transparency and the ability to run inference without exposing user inputs to a service provider. However, running large-scale models locally requires substantial computational resources. In practice, users may still resort to a third-party provider, giving rise to privacy and correctness concerns. Existing solutions that address these problems often impose substantial server overhead or introduce additional trust assumptions. In this paper, we present Maverick, a novel approach to private and verifiable LLM inference based on a protocol for delegating matrix-vector multiplication, a dominant operation in LLMs. At its core, Maverick provides, to our knowledge, the first information-theoretically sound verification protocol for matrix-vector multiplication delegation with transparent preprocessing, efficient (batch) verification, and virtually no server overhead. We combine this verification primitive with LPN-based pseudorandom masking to provide input privacy. We implement our matrix-vector delegation primitive and use it to build an end-to-end prototype of Maverick, which we evaluate on Qwen3-4B by measuring throughput in tokens per second. We evaluate client configurations with 1-8 threads. With one client thread and a CPU server using up to 128 threads, Maverick achieves throughput gains over local inference of up to 17x when privacy masks are generated online, 45x when they are precomputed, and 44x when only verification is required. With four client threads, the corresponding gains are 13x, 18x, and 17x. When server computation is no longer the bottleneck, client-side microbenchmarks with simulated network delay show speedups of 12x-20x, 34x-135x, and 38x-157x.

Depth pruning reduces large language model (LLM) inference cost by removing complete Transformer blocks. Activation-based methods collect hidden states through forward passes on calibration data, while existing forward-free methods score each Transformer block separately without measuring similarity between blocks. We propose Weight-Redundancy Pruning (WRP), a forward-free depth-pruning method that estimates inter-layer redundancy from checkpoint weights to select blocks without calibration data or model forward passes. WRP compares attention output and MLP down-projection weights across layers and combines their pairwise similarities with relative projection-scale information. The resulting all-pairs similarity matrix guides layer grouping and block selection. Across multiple pruning settings, model families, and downstream tasks, WRP consistently outperforms existing forward-free magnitude pruning and approaches the performance of activation-based methods.

SymbolicLight V2 combines sparse event computation with continuous-state processing in a hybrid neuromorphic language architecture. Extending V1's spike-gated dual paths, it adds graded signed events at further projections and softmax-free local attention. We implement the 194M-parameter model on an Alveo U50C FPGA using digital fixed-point arithmetic and on an ARM CPU using sparse integer execution. Across three same-checkpoint FPGA implementations at 175 MHz, active-row weight gathering and valid-state KV loading raise decode throughput from 474.6 to 643.2 tokens/s for a 32-token prefix and 128 outputs. Estimated gross card energy falls from 0.06087 to 0.04407 J per generated token, a 27.6% reduction. Complete-request energy, including prefill, falls by 24.4-27.7% across three prefix lengths. An independent idle split attributes 82.8% of gross card energy to loaded idle, explaining the benefit of shorter token latency. Against the recorded RTX 5090 compiled-FP32 baseline, integer FPGA execution uses 89.1% less estimated card energy during short-context decode; arithmetic precisions differ, and the GPU baseline is not the lowest-energy tested configuration. On four Cortex-A76 cores of a ROCK 5T, complete requests reach 65.4 tokens/s at 9.80 W and 0.151 J per generated token at the adapter's AC input. These results connect event sparsity to omitted computation and data movement. The mechanisms also support other dedicated V2 implementations: increasing throughput by a greater factor than active power lowers energy per generated token. Evaluation holds the deployed checkpoint fixed; its quality trails a same-budget dense control, so the results do not establish equal-quality efficiency.

Deploying Large Language Models (LLMs) directly on mobile platforms at the edge is gaining traction due to a myriad of benefits, such as increased privacy, personalization, and reduced latency. However, LLMs have heavy computational requirements, which are difficult for resource-constrained mobile and edge platforms to fulfill. In addition to limited compute resources, mobile and edge systems often have a compact form factor and lack physical mechanisms to dissipate heat generated from high processor usage rates (e.g., fans) to prevent throttling and reduced processing power, which LLMs can easily cause. To mitigate these effects, prior works have proposed various power governing strategies, such as dynamic voltage and frequency scaling (DVFS), for reducing power and heat generation for heavy computational tasks on mobile platforms. Recently, DVFS methods tailored for mobile LLMs have also been proposed. However, these methods mostly focus on optimizing hardware parameters and processor frequencies, and they fall short under some thermally constrained scenarios. Drawing from recent advances in machine learning, we identify and take advantage of the key insight that not all tokens require full-depth inference to maintain high-quality generation. Motivated by this, we present PELM, a solution that augments traditional DVFS processor frequency tuning with two additional workload-specific knobs: 1) speculative decoding and 2) variable verification depth to expand the optimization space to multiple dimensions for more power efficient on-device LLM inference. In extensive evaluations across hardware platforms and datasets, PELM demonstrates superior performance compared to state-of-the-art power governing methods, with up to 23.1% speedup and 52.4% reduction in energy consumption, while maintaining comparable task performance. The source code is available at https://github.com/imec-nu/PELM.

Inferring speaker relationships from spoken conversations is an important step towards socially aware speech understanding. However, this task remains underexplored, and supervised modeling is costly to train and scale. At the same time, existing inference-time LLM approaches provide limited structure for handling subtle, distributed, and multimodal relational cues that may support multiple plausible interpretations. To address these limitations, we introduce a training-free multi-agent reasoning framework that organizes inference through structured interaction among LLM agents, allowing relationship judgments to be proposed, challenged, and adjudicated without task-specific training. We instantiate this framework with two complementary designs. We propose Multi-Role Multi-Agent Debate as a task-specific adaptation of standard multi-agent debate for speaker relationship inference, assigning agents complementary roles or social-theory-grounded perspectives rather than a single undifferentiated viewpoint. In contrast, we introduce Multi-Agent Compete, a competition-based protocol that compares agent judgments through pairwise adjudication, eliminates weaker candidates, and retains the most defensible one. We evaluate these methods on the Seamless Interaction dataset across different modality settings, covering both binary classification and fine-grained relationship-detail prediction. Results suggest that they improve over zero-shot and existing multi-agent baselines in most cases. Human evaluation further suggests that this task is challenging even for people. LLM methods can sometimes outperform human annotators in text-included settings but are less competitive in the audio setting. Together, these findings suggest that relationship inference benefits from structured inference-time interaction among agents, while acoustic cues are not yet fully captured by current models.

Speculative decoding is critical for accelerating LLM inference. However, the speedup is fragile: drafters are typically trained against a narrow distribution for a single target model, and their acceptance rate collapses under workload shifts. This is a striking inversion of modern LLM development, where target models are valued precisely for the broad generalization they acquire through large-scale pretraining. We argue that the natural remedy, pretraining, has been hard to apply to drafters because existing recipes are target-specific: the drafter consumes the target's hidden states and is distilled on the target's logits, so pretraining must be repeated for each target. We introduce Osprey, which instead bootstraps drafters from off-the-shelf pretrained small language models, treating broad pretraining as a reusable, target-agnostic asset and reducing per-target work to a lightweight adaptation step. Realizing this requires overcoming two challenges: small LMs are far deeper than a latency-bound drafter can afford, and their pretrained computation must remain intact while the drafter learns to ingest target hidden states and emit tokens in the target's vocabulary. Osprey addresses both by pruning to a shallow backbone, restoring its language-modeling capability with target-agnostic next-token pretraining, and adapting it to each target through vocabulary alignment, zero-initialized QKV expansion, and distillation from the target model's output distribution. Empirically, a single pretrained Osprey backbone transfers across targets and improves mean acceptance length by 16.1% for Qwen3-8B, 21.2% for Llama-3.3-70B-Instruct, and 22.7% for the 229B MiniMax-M2.5 (with 17.5% higher tokens per second), with the largest gains on out-of-domain and multilingual data. Our code is available at https://github.com/LeanModels/Osprey.

Large language models (LLMs) are increasingly used as backends for intelligent web services, but serving them across the edge continuum requires balancing quality, latency, model footprint, and energy. This paper presents a controlled measurement study of self-hosted LLM inference across edge and near-edge deployment nodes: an NVIDIA Jetson AGX Orin and a near-edge server with CPU-only and GPU-enabled inference modes. We evaluate multiple open-weight LLMs and quantization variants using a fixed question-answering workload, and compare them against GPT-4o as a cloud-hosted accuracy and latency reference. Our benchmarking pipeline reports accuracy, model footprint, per-token decoding latency, prefill latency, and overall execution energy. The results show that GPU-enabled server execution provides the lowest compute-side latency, while Jetson Orin shows lower measured energy, consistent with its lower platform power under our setup. CPU-only execution is consistently dominated in latency for our workload and shows higher measured energy. We also show that parameter count and downloaded weight-file size alone do not reliably predict observed accuracy or latency. Finally, using Pareto-frontier analysis, we study how deployment decisions may change under possible streamed-token delivery overheads, highlighting that compute-side inference metrics alone can lead to suboptimal placement for latency-sensitive interactive web services.

Dynamic layer routing reduces the inference cost of Large Language Models (LLMs) by learning to skip layers for individual tokens. Existing methods, however, treat each routing decision as a local operation conditioned solely on the current hidden state which is a formulation that overlooks the sequential, path-dependent nature of routing across depth: earlier decisions shape the representations seen by downstream routers, and the layer-usage objective couples all decisions jointly. We propose History-Aware Routing (HeRo), a dynamic routing framework that resolves this mismatch by introducing a router memory mechanism to maintain an explicit routing state across model depth. The memory is constructed via linear attention, incrementally aggregating preceding routing scores and their induced residual updates into a compact history representation. At each routed layer, the router conditions jointly on this accumulated state and the current hidden representation to select the executed branch. Instantiated for token-wise FFN routing, HeRo trains only lightweight routers and adapters on a frozen backbone, requiring no modification to pretrained parameters. Across Llama 3.1-8B, Llama 2-7B, and Llama 2-13B, HeRo consistently achieves the highest aggregate performance retention among ten baselines. On Llama 3.1-8B, it bypasses 26.87% of model parameters while achieving 100.24% of dense model performance across seven benchmarks, and retains 97.01% while bypassing 38.82% of model parameters under a tighter computation budget. Ablation studies confirm that removing routing history consistently degrades performance, most notably on multistep reasoning and code generation, validating that explicit routing memory enables more accurate and adaptive dynamic routing than solely conditioning on hidden state.

Key--value (KV) cache compression is an effective way to reduce the memory overhead of large language model (LLM) inference, particularly for long-context workloads. However, existing compression methods make different trade-offs among accuracy, inference latency, and peak KV cache memory utilization, making a single fixed configuration unsuitable across different prompts and resource constraints. We introduce MetaKV, an adaptive framework that selects a KV cache compression configuration for each input prompt based on user-specified latency and peak memory budgets. MetaKV uses lightweight prediction models to estimate the end-to-end latency, peak memory, and probability of a correct response for each candidate configuration, and selects the configuration that best satisfies the latency-memory constraints while preserving accuracy. We evaluate MetaKV across ten configurations from three representative KV cache compression methods, KVQuant, H$_2$O, and RocketKV, together with an uncompressed FP16 configuration, on four datasets covering mathematics, science, commonsense reasoning, and reading comprehension. Across a wide range of latency and peak memory constraints, MetaKV consistently outperforms the best static configuration, improving constrained success rate (CSR), the fraction of prompts answered correctly while satisfying both constraints, by approximately 0.07 on average and up to 0.135. These results demonstrate the benefit of adapting KV cache compression to individual prompts and latency-memory constraints. Code is available at https://github.com/MichaelWang0505/MetaKV.git

Large language model (LLM) cascades answer easy requests with a small model and escalate selected requests to a larger model. Most routers prioritize examples on which the small model appears uncertain or likely to be wrong. This proxy ignores a decisive fact: escalation is useful only when the large model corrects the small model, and it is harmful when the large model replaces a correct answer with an incorrect one. We introduce Signed Rescue Routing (SRR), a budgeted routing method that predicts these two events separately and ranks requests by their difference. We show that this signed conditional gain is the Bayes-optimal routing score under a fixed escalation budget. SRR requires only the small model's output statistics at deployment and adds a lightweight two-head router. We evaluate SRR with Qwen3-4B and Qwen3-8B on TBD examples from MMLU, HellaSwag, and ARC-Challenge. Across the accuracy-compute curve, SRR reaches an area of TBD, compared with TBD for a learned small-model error predictor and TBD for entropy routing. These results show that predicting incremental value, rather than model uncertainty, is a simple and effective objective for efficient LLM cascades.

Large language models (LLMs) face severe memory bottlenecks in long-context inference due to the linearly growing size of key-value (KV) caches. Existing KV cache compression techniques typically rely on simple heuristics, overlooking the distinct functional roles of different attention heads. We present SGD-KV (Summarization-Guided KV Cache Compression), a head-aware framework that leverages a novel chunk-summarization diagnostic task to systematically identify and prioritize attention heads specialized in hierarchical information aggregation. Experiments on Qwen2.5-7B-1M and Qwen3-32B across diverse long-context benchmarks demonstrate that SGD-KV achieves state-of-the-art performance with contexts up to 1M tokens, while reducing KV cache memory usage by up to 75%. Our findings show that strategically allocating the KV cache budget based on the summarization score distribution of attention heads yields a superior efficiency-accuracy trade-off for long-context inference.

On-device LLM inference is attractive for privacy and responsiveness, but remains challenging on mobile and embedded devices because model weights far exceed available DRAM. Prior systems exploit activation sparsity and offload weights to SSD or flash storage, but face a fundamental systems trade-off: accurate sparse execution decisions require the latest context, whereas efficient computation-I/O overlap requires early prediction. As a result, existing designs either serialize execution or incur redundant weight fetches, extra computation, and large cache overheads. We present LeanStream, a streaming speculate-and-refine framework for efficient on-device LLM inference. LeanStream progressively refines computation, loading, and cache-retention priorities using partial GPU results, enabling fine-grained overlap between GPU execution and storage I/O. We implement LeanStream on both mobile and embedded platforms. Compared with prior on-device LLM inference systems, LeanStream reduces memory usage by 4.8$\times$ to 7.5$\times$ at the best throughput achieved by prior work, while further improving token generation throughput by 1.6$\times$ to 2.1$\times$.

Long-context inference retains a growing key--value (KV) cache during decoding, which consumes substantial GPU memory and can reduce generation throughput. This bottleneck remains in hybrid language models because their residual global-attention layers can dominate context-dependent cache demand. We study how to allocate this state under an aggregate KV-residency budget. We introduce HeadWiseKV, a training-free framework that compresses the residual global KV caches of hybrid language models while preserving their native local, recurrent, and linear paths. It assigns each physical KV head a static, multilevel history window, making cache demand predictable before serving. We formulate this allocation as a restricted operational rate--distortion problem and propose SeqCalib as the core policy-generation algorithm in HeadWiseKV. SeqCalib processes layers in execution order and conditions each decision on the lower-layer policy used at deployment, thereby accounting for interactions across depth. A grouped-cache runtime materializes the selected policy as actual per-head KV residency rather than a mask over a full cache. We evaluate downstream quality across four hybrid long-context models and study physical residency and serving behavior on Qwen3.6-27B. HeadWiseKV retains near-Full-KV RULER and LoCoMo quality across the evaluated models. In the fixed-model systems study, it reduces sampled peak device memory by 8.59% at a 112K context length and extends the largest verified successful context from 114K to 161K.

The attention prefilling phase of long-context LLM inference scales quadratically, making self-attention a severe computational bottleneck. Traditional sparse attention methods mitigate this through fixed patterns or offline profiling, but lack the flexibility to adapt to input-dependent attention structure. Recent dynamic methods address this by routing heads to sparse patterns in real-time, but rely on indirect routing proxies with overhead and budget allocation mechanisms that overlook the post-softmax mass hierarchy. We present CRISP (Cliff-awaRe Input-adaptive Sparse Prefilling), which identifies and addresses two structural challenges in this dynamic routing paradigm. First, we show that the routing decision can be read directly off the structure of the proxy attention map. We replace the Jensen-Shannon Divergence (JSD) routing with C_struct, a structural proxy that measures mass at Vertical-Slash compatible positions and reproduces JSD's routing decisions while eliminating both the pooled matmul and subsequent KL divergence overhead. Second, we formalize the post-softmax mass cliff and demonstrate theoretically that strictly cumulative coverage thresholds accumulate O(n) background noise at long contexts. CRISP navigates this via a sink-aware threshold grounded in the noise floor. Empirically, across InfiniteBench, RULER and LongBench on two model families, CRISP is the strongest sparse method overall and matches or exceeds exact dense attention on retrieval-heavy benchmarks, recovering up to +28.0 pp on retrieval tasks over baselines and achieving up to a 5.30x attention speedup at 512k tokens, driven primarily by our O(n) noise elimination during selection while preserving structural integrity.

Sentence-level recognition of depression symptoms is challenging because similar expressions can differ in symptom relevance, and language-model inference is insufficiently grounded in diagnostic definitions. This study proposes a two-stage framework separating symptom-candidate generation from definition-grounded verification. A contrastively fine-tuned sentence encoder generates a symptom candidate per sentence, and a fine-tuned language model verifies whether the candidate is present or absent using the sentence, its context, and a candidate-specific diagnostic definition, checking its judgment against that definition before answering. Evaluated against encoder, inference-based, medical, and general LLM baselines and a matched single-stage supervised classifier, the proposed pipeline attains the best accuracy and F1 scores of all methods, with rationales matching expert-authored annotations. A preliminary clinical audit indicates moderate alignment with diagnostic definitions, with explanation quality strongly dependent on prediction correctness. The results support decomposing symptom recognition into candidate generation and definition-grounded verification, though performance remains limited for rare categories.

Large language models (LLMs) are increasingly deployed on mobile devices, making energy efficiency a key deployment constraint, yet the energy impact of prompt design remains underexplored. This paper aims to understand how two prompt properties, cognitive load and phrasing pattern, shape the energy behavior of on-device LLM inference. We conduct a broad empirical study covering prompt properties, datasets, models, and devices, with phase-level profiling that separates prefill and decode energy. We find that cognitive load primarily affects the energy cost per token, while phrasing pattern affects energy largely through token usage. Our energy-quality analysis further shows that prompt design reshapes the attainable frontier differently across models, highlighting the need for model-aware prompt design in energy-efficient on-device LLM inference. Code, datasets, and scripts are available at https://amai-gsu.github.io/PromptProperty/.

Compressed context is usually carried as human-readable text or as rendered images that must be decoded, even when its consumer is a language model. We introduce LatentPress, which writes conversational histories and long documents into a third representation: continuous memory tokens that a frozen decoder reads directly through its input-embedding interface, with no text reconstruction at inference. A small reader-matched writer compresses $4$-$16\times$ while training only an adapter (4.2M-26.2M parameters, $\sim\!0.1\%$ of the decoder). On LongMemEval, LatentPress reaches $0.504$ accuracy at $7.70\times$ compression versus $0.490$ for uncompressed evidence, outperforming text summaries (0.184) and OCR-based compression (0.426 to 0.312). On LongBench-QA, in-domain writers match or exceed raw-context reading at $4$-$8\times$ compression, while $16\times$ trails raw. Writing takes 43ms per conversation, roughly an order of magnitude faster than text summarization or OCR reconstruction, and reading is $5$-$9\times$ faster than raw context or cached OCR. We validate the interface under two transfer settings, zero-shot from UltraChat to LongMemEval memory QA and from LongMemEval-derived QA to unseen LongBench document domains, establishing direct soft tokens as a practical machine-facing context interface beyond text and vision. The implementation of the experiments could be found at: https://github.com/HJSang/LatentPress .

On-device mobile Large Language Model (LLM) inference is gaining significant attention. However, mobile devices operate in highly dynamic multitasking environments where users frequently switch between applications. This creates memory pressure, forcing LLM memory (model weights and KV cache) to be evicted by the operating system. When a new inference request arrives, the inference system must restore the evicted memory through slow storage reads or recompute the entire KV cache, severely degrading responsiveness. To address this, we present mzCache, an on-device LLM inference system with specialized memory management for multitasking environments. Under unpredictable memory pressure, mzCache elastically evicts LLM memory and leverages the unified memory of mobile SoCs to enable zero-wait inference on the GPU with concurrent CPU-side restoration. mzCache realizes this through restoration-oriented memory management: LLM memory is partitioned into fine-grained shared buffers to enable partial eviction and restoration with concurrent cross-processor access, while hybrid swap and backward-out eviction policies ensure low-latency restoration from any eviction state. Implemented on llama.cpp and deployed as an Android application, mzCache achieves 2.1-5.5$\times$ reduction in Time-to-First-Token compared to storage-backed partial offload and demonstrates its effectiveness in real multitasking scenarios.

Mixture-of-Experts (MoE) architectures scale Large Language Model (LLM) capacity efficiently by activating a sparse subset of experts per token. However, modern MoE inference remains heavily constrained by the rigid, whole-expert abstraction. Existing frameworks manage, schedule, or prune experts as atomic execution units, which fixes the optimization boundary too early and leaves fine-grained intra-expert computational redundancy underexplored. In this work, we present PCoMoE, a path-compositional execution framework that shifts MoE inference from coarse-grained expert selection to fine-grained path composition. PCoMoE incorporates a path-level formulation of expert computation, a compatibility-aware layer-wise pruning strategy to suppress low-value path combinations, and a hardware-friendly execution engine to exploit reusable sub-expert structures under strictly bounded overheads. Experimental results demonstrate that PCoMoE achieves up to a 1.31x end-to-end inference speedup while enhancing model accuracy by 10%. The code is available at https://github.com/gzyyy0/PCoMoE

As one of the most critical challenges in large language models, contextual faithfulness directly determines their reliability in knowledge-intensive applications. This task is particularly challenging as it requires balancing factual consistency with generation efficiency. Contrastive decoding methods require dual forward passes (with and without context) to compare model outputs, doubling inference computational overhead, while post-training alignment demands extensive reinforcement learning with substantial computational overhead. To address this challenge, we present SFAD, a speculative decoding framework that enhances contextual faithfulness without inference degradation. We first construct ConFide, a preference dataset with fine-grained atomic perturbations, to train a context-faithful draft model via Direct Preference Optimization. During inference, Epistemic Friction detects potential hallucinations by quantifying distributional tension weighted by specialist certainty. When friction exceeds the threshold, Asymmetric Logit Steering refines the target distribution through residual-based logit injection; otherwise, standard speculation proceeds. Extensive experiments demonstrate that SFAD substantially improves faithfulness while achieving $2.48\times$ speedup, offering a practical solution for efficient LLMs.

Block Quantization (BQ) enables efficient LLM inference by quantizing both weights and activations, but its design space remains underexplored. Through hardware-accuracy design space exploration, we identify block size as a key trade-off: larger blocks improve hardware efficiency by amortizing dequantization and accumulation costs, but degrade accuracy. Motivated by this insight, we propose Hierarchical Block Quantization (HBQ), which combines large blocks with low-overhead significand (SIG) scaling for second-level quantization. SIG scaling effectively compensates for large-block quantization errors while accounting for distinct weight and activation distributions. HBQ-A achieves W4A16-level accuracy with W4A5 and lower area than NVFP4, while HBQ-E further reduces hardware cost by 17% while outperforming existing BQ methods in accuracy. We implement HBQ for weights, activations, and KV cache in a 28nm ASIC accelerator and introduce partial-sum BQ to reduce EMA energy. At comparable accuracy, HBQ achieves 2.3x/4.6x higher area/energy efficiency than state-of-the-art weight-only quantization and 1.6-3.3x lower system energy with 1.5-3x speedup over prior BQ methods. Our implementation is publicly available at: https://github.com/SeoLabCornell/HBQ.git.

The development of long-context Large Language Models (LLMs) is constrained by the memory bandwidth bottleneck and quadratic complexity of the attention mechanism during decoding. To overcome the inherent trade-offs between the memory overhead of metadata-based metrics and the computational inefficiency of adaptive selection strategies, we present Faster Flash Decoding (FFD), a novel hardware-algorithm co-design framework designed to break the memory wall in long-context decoding. FFD integrates the selector and computer into a fully fused kernel, replacing external metadata indices with content-aware scanning via low-bit quantization. Furthermore, we introduce the top-delta strategy, which dynamically filters blocks to achieve distribution-adaptive sparsity without global synchronization. Offering a training-free and plug-and-play solution, FFD also enables the reuse of scanning results for computation, achieving up to 11.6x kernel-level speedup and scaling to 256K context length, with 2.37x end-to-end throughput improvement. Empirical validation on RULER and LongBench confirms that FFD maintains model accuracy while delivering high-ratio sparsity, with code available at https://github.com/qluoluo/faster-flash-decoding

Long-context compression is essential for reducing the cost and latency of large language model inference. However, existing methods can fragment important evidence, require additional training or alignment, and often depend on the target model for effective compression. We introduce TopoCompress, a training-free and model-agnostic framework that compresses long contexts by selecting coherent semantic spans. TopoCompress first scores each span using dense and lexical query relevance together with semantic acceleration. It then constructs a hybrid graph that connects spans based on semantic similarity and sequential adjacency, and propagates the query-guided relevance scores over the graph. Across five long-context tasks-HotpotQA, 2WikiMQA, MuSiQue, Qasper, and MultiFieldQA-en-TopoCompress consistently outperforms strong compression baselines. Notably, TopoCompress achieves performance comparable to the strongest baseline while using a 4x smaller compression budget, and provides a 1.41x smaller compression time over the fastest baseline.

Large Language Models (LLMs) offer new possibilities for scaling qualitative analysis, but existing applications often provide limited methodological transparency regarding how qualitative methods are translated into computational procedures. This paper presents an auditable and privacy-preserving computational operationalization of inductive and latent Thematic Analysis (TA). This paper first derives five design principles from the methodological requirements of TA and the conditions introduced by LLM-based inference: preserving interpretative context, maintaining traceable relationships between empirical material and analytical outputs, representing analytical constructs and reasoning explicitly, constraining LLM inference to interpretative tasks, and enabling privacy-preserving local deployment. Second, it presents a proof-of-concept for a two-phase workflow that operationalizes these principles by combining interpretative LLM inference with deterministic procedural control to generate codes, analytical justifications, themes, and theme descriptions while preserving explicit links to the source material. Third, it proposes an evaluation framework combining structural comparison with human-led TA and independent expert assessment of analytical quality. The evaluation is conducted on semi-structured Danish interview transcripts. and the results shows that the workflow produces code-level outputs with coverage broadly comparable to human annotations and highly rated analytical justifications, while generating a more compressed thematic structure characterized by fewer and broader themes. The findings demonstrate the feasibility of auditable LLM-supported TA through a modular workflow designed to scale to larger datasets, accommodate different LLMs, and support transfer across research domains, with domain adaptation primarily requiring adjustments to the prompting strategy.

Inference with transformer-based large language models (LLMs) is often limited by the memory-bound KV cache and quadratic attention cost. State-space models (SSMs) mitigate this through linear attention and fixed-size recurrent states, but their large dense linear projections remain computationally expensive even after quantization. We introduce a method that induces sparse neural activity in heavily quantized linear-attention models with minimal performance loss. Activations below a per-projection trainable threshold ($\pm Δ$) are nullified while preserving crucial outliers, achieving comparable performance to dense models with up to 4$\times$ fewer effective arithmetic operations. Targeting a multi-core, multi-chip neuromorphic platform, where event-driven execution converts unstructured sparsity into throughput at both the compute and communication levels, a capability GPU architectures fundamentally lack, we project up to 37$\times$ higher throughput and 16$\times$ lower power versus edge GPU inference of a comparable transformer-based model, and up to 5.4$\times$ improvements over the non-sparsified baseline. These results position sparse, quantized linear-attention models as a natural fit for deploying LLMs on event-driven multi-core platforms.