Quantization, distillation, pruning and serving work aimed at the same accuracy for less memory, latency and money.
Using a Large Language Model (LLM) as the clusterer at production scale is hard: prompts cannot hold the entire label space, and per-document serial processing does not deliver the throughput real wor...
Mechanistic interpretability of vision transformers seeks to decompose model computation into human-readable units, but learned representations entangle many concepts in each neuron. Feature superposi...
An Intraoperative Hypotension (IOH) event is a frequent complication during administration of general anaesthesia with serious downstream consequences, yet clinical management remains reactive and not...
We show that a quantized model that keeps its classification accuracy still changes $14$ to $46\%$ of its top-1 retrieval results, and that aggregate ranking metrics reveal only part of this damage. W...
The Platonic Representation Hypothesis (PRH) claims that independently trained models converge on a shared statistical model of reality, yet recent work finds only weak pointwise similarity between mo...
The paper introduces Jarvis, an offline, edge‑deployable voice assistant designed for autonomous racecars. It combines speech recognition, synthesis, and a lightweight text‑to‑command classifier fine‑tuned from the Mistral 7B model to provide high‑level behavioral commands. Experiments show 97.63 % intent recognition accuracy with an average latency of 1.39 s, outperforming larger online‑hosted models and enabling quick response times for time‑critical driving tasks.
The paper introduces GUARD, a method for natural forgetting in large reasoning models that transforms unsafe disclosures into safe-exit trajectories using guided answer‑reasoning distillation. It aligns a frozen model with guidance tokens and distills this behavior into the parameters, aiming for a coherent, non‑disclosing chain of thought followed by a refusal‑style answer. The authors also propose the Natural Forgetting Reasoning Score (NFRS) to evaluate structural stability, fluency, and unsupported substitutes, and demonstrate GUARD’s effectiveness on R‑TOFU and a STAR‑1‑derived harmful‑intent setting.
The paper introduces a layerwise, decoupled approach to structurally sparsify fully connected layers in pretrained neural networks. By extracting shallow two‑layer subnetworks, normalizing inner weights, and applying a structured group penalty to each block’s outer weight matrix, the method prunes neurons sequentially and reduces layer widths. The authors prove equivalence to a joint penalty for positively homogeneous activations, and demonstrate that this decoupled formulation is more robust, offering a broader regularization range and lower catastrophic over‑pruning while preserving accuracy in classification, sparse‑recovery, PINN, and OPT‑1.3B experiments.
EFQ-Softmax is a low‑bit probability‑generation technique that replaces the traditional exp‑then‑quantize path in Transformer attention. It maps shifted attention scores directly to block‑scaled E2M1 operands using an exponent‑only scale and a single affine rule, allowing the same low‑bit representation to be used for both numerator and denominator updates. Experiments on Qwen3‑8B, Qwen3‑VL‑8B‑Instruct, and WAN2.2‑TI2V‑5B show that EFQ‑Softmax maintains or improves model quality while reducing vector‑stage latency by about 40% on the A5 vector unit.
Ripple-Pivot Search (RPS) is a training‑free decoding method for Diffusion Large Language Models that identifies mid‑entropy pivot positions to reduce uncertainty across remaining masked tokens. By proactively committing these pivots and evaluating token assignments via lookahead, RPS enables more tokens to be unmasked in parallel, speeding up decoding. Experiments on three dLLMs and four reasoning/code‑generation benchmarks show 4–10× wall‑clock speedup over standard decoding, up to 18× with KV caching, while maintaining or improving generation quality.
The paper introduces QuAKE, a Quantization-Aware Kalman Estimator designed to correct errors in diffusion model sampling when using quantized denoisers. By treating sampling as an online estimation problem, QuAKE leverages the history of quantized outputs to recover full-precision estimates, updating a posterior in closed form at each step. The method is lightweight, plug‑and‑play, and works with any high‑order multistep ODE sampler, outperforming existing correction techniques on W4A4‑quantized text‑to‑image diffusion models.
The paper reports on programming AMD XDNA NPUs for the FlashAttention workload using open‑source IRON and MLIR‑AIR compiler tools. It compares four reference designs on XDNA 1 and XDNA 2, showing that a fused kernel that keeps QKᵀ scores in local memory achieves 3.62 TFLOP/s on XDNA 2, doubling throughput and greatly improving energy efficiency over the IRON design and the integrated GPU. Roofline analysis guides when to fuse or stream operators based on each device’s ridge points, and the authors release the reference designs as open source.
arXiv:2609. 20883v1 Announce Type: new Abstract: Despite the widespread use and success of generative AI techniques today, theoretical guarantees on learning a distribution supported in $d$ dimensions from $n$ samples degrade as $O(n^{-1/\Theta(d)})$, though shown to be minimax optimal.
The study evaluates automatic tooth segmentation on panoramic radiographs using a large annotated corpus of 1,422 images and 42,142 tooth polygons. It finds that increasing input resolution improves boundary precision (mask mAP50‑95 rises from 0.656 to 0.717) while detection performance remains unchanged, and that architectural changes have minimal impact on in‑domain accuracy. Targeted interventions such as LoRA adaptation, promptable foundation models, and anatomical label assignment provide negligible gains, indicating that resolution and acquisition diversity should be prioritized over model novelty.
SpecQuant is a training‑free framework that merges speculative decoding with multi‑parent quantization to enable adaptive, efficient inference of large language models. It generates several quantized variants (INT4, FP8, FP16) from a single base model and routes queries to the appropriate variant based on predicted complexity, using lightweight models for simple tasks and full‑precision models for complex reasoning. Evaluations on Qwen2.5 models across MMLU, AlpacaEval, and GSM8K show 35‑43% speedups with less than 2% accuracy loss, facilitating practical on‑device LLM deployment without specialized infrastructure.
ArenaFlow is a hierarchical credit propagation framework designed to improve reinforcement learning for open-ended agent tasks. It uses tournament-based relative ranking to generate trajectory-level rewards and structured reflective evaluation to identify pivotal success steps, reusable strategy skills, and skill usage attribution. The framework propagates advantages to high-confidence steps and maintains a global skill memory, enabling more targeted optimization and reusable skill priors for future exploration.
The paper presents Sometin Beta Pass Notin (SBPN), a multilingual ASR framework for Nigerian languages that uses a two‑stage knowledge‑distillation approach. First, student‑teacher distillation from monolingual models is conditioned on language‑specific N‑gram language models; second, iterative self‑improvement with pseudo‑labelled data further refines accuracy. The method reduces relative WER by 29% over monolingual baselines and outperforms state‑of‑the‑art multilingual models on Common Voice and FLEURS benchmarks for Yoruba, Hausa, Igbo, Nigerian Pidgin, and Nigerian English.
GestureFAR is a flow‑autoregressive framework that generates natural co‑speech gestures from streaming speech while preserving causality and continuous motion expressiveness. It autoregresses over continuous motion latents using a transformer for audio‑motion context and a flow‑matching head to sample the next latent. A head‑only flow distillation strategy further reduces latency by collapsing multi‑step flow sampling into a single network evaluation, enabling real‑time token‑causal generation with improved quality‑latency trade‑off on the BEAT2 benchmark.
Edit‑VAR is a training‑free, inversion‑free framework that uses a pretrained visual autoregressive video model for text‑guided video editing. It encodes the source video into multi‑scale discrete tokens and applies probability‑guided conditional token replacement, attention‑guided token‑wise and scale‑aware modulation, and scale‑decoupled generation to preserve source appearance while enabling precise edits. The method also includes residual‑guided token pruning to reduce inference cost, and experimental results show it outperforms existing training‑free video editing methods in fidelity, source preservation, temporal coherence, and efficiency.
GEM-MPC is a reinforcement learning method that blends MPPI planning with policy learning to balance exploration and exploitation in high-dimensional continuous control tasks. It trains a policy to clone the planner while also maintaining a KL-regularized policy that explores around the planner’s suggestions, thereby improving the synergy between planning and learning. The approach introduces Gated Prior Distillation, which selectively updates policies from stored planning distributions only when they offer better targets, reducing the influence of stale data without costly reanalysis. Across continuous-control benchmarks, GEM-MPC outperforms existing planning-based baselines while using lower computational budgets.