HeadWiseKV is a training‑free framework that compresses the residual global key–value caches of hybrid long‑context language models by assigning each physical KV head a static, multilevel history window. It formulates cache allocation as a restricted operational rate–distortion problem and uses the SeqCalib algorithm to generate per‑head residency policies that account for interactions across layers. In evaluations on four hybrid models, HeadWiseKV preserves near‑full‑KV quality while reducing peak device memory usage by 8.59% at a 112K context length and extending the largest verified context from 114K to 161K.
By Renjie Xie, Juncheng Yang, Aoting Hu, Mingxi Zhang, Liyao Wu, Zheheng Hong, Wei Xu
ShallowStream is a framework for streaming video understanding that uses the shallow layers of a multimodal large language model (MLLM) to encode frames and build a lightweight index. During streaming, it maintains an always‑on index via the KV cache of shallow layers, and at query time it scores context frames using shallow‑layer attention and selects diverse evidence for answering. The approach matches the performance of leading streaming methods while cutting per‑frame prefill latency and 10‑second end‑to‑end latency by up to 52.1× and 11.9×, respectively.
By Jitai Hao, Ke Yang, Qiang Huang, Jun Yu
Debias‑SparseGPT is a post‑training pruning technique that adds a representational debiasing term based on demographically contrasting inputs to mitigate bias amplification caused by weight sparsification. The method is validated across various generative LLMs and sparsity levels (25%, 50%, and structured 2:4), consistently reducing pruning‑induced bias while maintaining perplexity and zero‑shot accuracy. In the most aggressive 2:4 sparsity regime, enriching the calibration set with long‑context, content‑rich examples further improves both downstream performance and fairness.
By Irina Proskurina, Guillaume Metzler, Antoine Gourru, Julien Velcin
The paper introduces a new 4‑bit floating‑point (FP4) pretraining approach that pairs E2M1 payloads with unsigned E5M3 block scales, enabling periodic tensor scaling and selective stochastic rounding while eliminating the randomized Hadamard transform. Using this method, the authors pretrained a Nemotron‑H 8B model on nearly 190 billion tokens, achieving lower training and validation losses compared to NVIDIA’s Transformer Engine. The approach also improves inference performance and demonstrates a 21.2 % increase in token throughput when certain optimizations are removed.
By Robert Hu, Carlo Luschi, Paul Balanca
Cliff is a reward‑shaping method for reinforcement learning with verifiable rewards that identifies the first mistake in a language model’s reasoning process using an off‑the‑shelf teacher. By splitting each rollout into a correct prefix and an incorrect suffix, Cliff assigns token‑level advantages—positive for correct tokens and negative for the rest—providing fine‑grained supervision. Across 12 scenarios, Cliff outperforms on‑policy distillation by 15% and standard GRPO by 7%, even when the teacher is only modestly capable.
By Peixuan Han, Runhui Wang, Ketan Ramaneti, Jie Hao, Gerald Friedland, Chris Kong
The paper introduces Repo-To-Skill, a method for converting GitHub repositories into reusable AI skills. By distilling operational knowledge from over 1,000 machine‑learning repositories, the authors build the AREX‑Skill Library with more than 5,000 verified skills across 20 areas. Integrating these skills into a research agent—DisCo—yields significant performance boosts on multiple benchmarks, demonstrating the value of reusable, task‑agnostic knowledge.
By Jianlyu Chen, Yuyang Hu, Hongjin Qian, Jiawei Liu, Wenqing Wei, Xiaolong Chen, Defu Lian, Zhicheng Dou, Chaozhuo Li, Qiwei Ye, Zheng Liu
The paper introduces a multi‑shell decoder for Leech‑lattice vector quantization, achieving the best reported 2‑bit quality under its evaluation protocol. It presents a GPU‑friendly layout that fuses dequantization with matrix‑vector multiplication, demonstrating significant speed and memory advantages over traditional one‑hot masks and other 4‑bit methods. Experiments show the new kernel outperforms baseline approaches across multiple model sizes, with measurable gains in throughput and reduced byte traffic.
By Pier-Jean Malandrino (Scub)
The paper presents a systematic framework for large language model (LLM) watermarking as a provenance tool in big data ecosystems. It categorizes existing watermarking methods along four deployment dimensions—insertion point, verification authority, operational state, and transformation threat model—and aligns them with the big data principles of Volume, Velocity, Variety, Veracity, and Value. The authors introduce a readiness framework that maps four key workloads—online generation, streaming detection, transformation pipelines, and ecosystem governance—to system-level requirements such as throughput, false-positive control, robustness, cross-domain reliability, governance, and downstream utility, while highlighting gaps between benchmark performance and real-world deployment readiness.
By Huy Phan, Kieu Dang, Ojaswi Dulal, Aiham AL Shukairi, Abby Shine, Chase Garner, Phung Lai
FORGE is a forward‑only test‑time adaptation technique designed for integer‑only vision models running on microcontrollers. It restores batch‑normalization statistics after BN folding by re‑normalizing each convolution’s per‑channel output using only forward‑pass estimates, enabling adaptation on deployed, folded integer models. The method achieves accuracy gains comparable to gradient‑based TENT, requires adapting only a few layers, works with single‑sample streaming, and has been validated on an ESP32‑S3 with minimal energy and latency overhead.
By Muhammad Rehan, Haider Ali, Muhammad Ali Munir, Moaz Amjad
The paper introduces Epipolar Distillation (EpiDistill), a method that transfers scale‑aware geometric priors from multi‑view models to monocular depth foundation models using Rectified Stereo Tokens. By preserving epipolar attention patterns, the single‑view model maintains geometric consistency without needing multi‑view inputs during inference. Experiments show significant improvements in zero‑shot metric depth estimation on challenging datasets such as ETH3D and DIODE, and the approach consistently boosts performance of state‑of‑the‑art ViT‑based models like UniDepthV2 and DepthPro.
By Jung-Hee Kim, Xiaoming Liu
SGD-KV is a head‑aware framework that compresses key‑value caches in large language models by using a chunk‑summarization diagnostic task to identify attention heads that specialize in hierarchical information aggregation. It prioritizes these heads during compression, achieving state‑of‑the‑art performance on long‑context benchmarks with up to 1M tokens while cutting KV cache memory usage by as much as 75%. Experiments on Qwen2.5‑7B‑1M and Qwen3‑32B confirm that allocating cache budget based on summarization scores yields a superior efficiency‑accuracy trade‑off for long‑context inference.
The paper introduces MIRC, an overfitted image codec that quantizes all components—including latents, synthesis network, and entropy models—within a single rate‑distortion objective, following the neural video representation codec NVRC. It adds a multi‑scale representation with cross‑stage parameter sharing to capture cross‑scale redundancy, achieving a 10.5% BD‑rate saving over VVC on the CLIC2020 professional validation set. MIRC also offers configurable decoding complexity ranging from 1.2 to 2.9 kMAC per pixel, allowing deployment to match specific resource budgets.
Stable 4-bit floating-point (FP4) pretraining is difficult because the E2M1 payload represents only a narrow range of magnitudes. NVIDIA's Transformer Engine \nv{} recipe addresses this with current-tensor scaling, a randomized Hadamard transform (RHT), and bfloat16 (BF16) final layers, adding work outside the FP4 matrix multiplications.
Cliff is a reward‑shaping method for reinforcement learning with verifiable rewards that identifies the first mistake in a language model’s reasoning process using an off‑the‑shelf teacher. It splits each rollout into a correct prefix and an incorrect suffix, assigning positive token‑level advantages to the prefix and negative feedback to the suffix. Across 12 scenarios, Cliff improves reasoning performance, outperforming on‑policy distillation by 15% and standard GRPO by 7%, even with modest teachers.
Repo-To-Skill: Distilling GitHub Repositories Into AI4AI Skills introduces DisCo, a research agent that extracts and verifies operational knowledge from GitHub repositories to create reusable AI skills. The agent produces both task‑agnostic skills—compiled into the AREX‑Skill Library of over 5,000 verified skills from 1,000 repositories—and task‑oriented skills tailored to specific research tasks. When equipped with these skills, the agent achieves significant performance gains across multiple benchmarks, outperforming a skill‑free version by 134.3% on MLE‑bench, 34.4% on PaperBench, 9.2% on FrontierCS, and 14.0% on PassNet.
H3DNAS is a hardware‑aware compression framework that operates directly on ONNX computational graphs, eliminating the need for original source code or gradient access. It introduces a Channel Dependency Graph to classify operators and compute a provable compression ceiling, and a two‑stage hierarchical search that prunes architectures via L1‑importance channel selection and applies GhostConv mutations. On the ModelNet40 dataset, H3DNAS achieves significant parameter reductions and inference speedups for PointNet, PointNet++, and PointMLP with minimal accuracy loss.
Loom is a generative consensus framework designed for real‑world root‑cause analysis (RCA) that combines open‑form hypotheses from modular heuristics with a lightweight large language model (LLM). It projects hypotheses into a continuous embedding space and uses an iterative centroid‑based reweighting algorithm to resolve conflicts, producing a single consensus that is then synthesized by one LLM call. On the OpenRCA benchmark, Loom achieves state‑of‑the‑art accuracy on Bank and Market‑2 while being significantly faster and more efficient than existing autonomous agents.
The paper introduces Multi-Teacher Self-Distillation Policy Optimization (MT‑SDPO), an on‑policy distillation method that combines multiple frozen teachers into a single student model. MT‑SDPO uses self‑anchors, answer‑verified eligibility, and privileged distillation to identify the most reliable teacher for each sample rather than relying on domain labels. Experiments on five students from three model families show that MT‑SDPO improves the weakest domain of Qwen3‑8B by 14.79 points and reduces its domain gap by 74.7%, achieving a better balance than matching a single teacher per domain.
The paper investigates where post‑training quantization (PTQ) harms large language models (LLMs) and how to best allocate a limited precision budget. By causally raising each layer to 8‑bit precision across nine open‑weight models, the authors find that quantization damage is diffuse rather than concentrated in specific task circuits or weight statistics, and that globally refining quantization granularity outperforms selectively protecting the most recoverable layers. They also observe that the residual accuracy loss is budget‑limited and that peak recovery locations correlate with architecture within families but not across families.
By Jundong Hu, Shekar Ramachandran
arXiv:2603.05887v2 Announce Type: replace-cross
Abstract: Neural audio codecs optimized for mel-spectrogram reconstruction often fail to preserve intelligibility. While semantic encoder distillation...
By Junhyeok Lee, Xiluo He, Jihwan Lee, Helin Wang, Shrikanth Narayanan, Thomas Thebaud, Laureano Moro-Velazquez, Jes\'us Villalba, Najim Dehak