Search AI news

BlockLock: Reproducible LLM Inference with Fixed Batch Geometry

arXiv:2611.06977v1 Announce Type: new Abstract: Reproducible LLM inference requires that a request produces the same output regardless of the other requests in the batch. Existing solutions pay for this in different ways: batch-invariant kernels constrain every operator, and isolating prefill with fixed-shape decode ({our CoRun-style baseline}) gives up shared prefill compute and pads decode. We present BlockLock, which keeps shared batched prefill and pins only the numerical path of the protected request: it fixes the geometry that the request sees, with {two modes: BlockLock-Rows uses fixed-width rows and retains batched decode of all requests; BlockLock-Pack uses terminal-anchored packing with an isolated dummy segment, then extracts the protected request's KV state for target-only decode. Under a stated per-deployment contract, both modes preserve the target's output bitwise when its companions change.} Across five {dense} models (Qwen3-8B/14B, Llama-3.1-8B-Instruct, and Gemma-3-4B/12B-IT) and three single-GPU deployments (NVIDIA A6000, A100, and H100), we evaluate our approach on both tinyBenchmarks and BFCL{; we also include initial tensor-parallel experiments on two NVIDIA A6000 GPUs. Both modes preserve target sequences and Top20-logit bits in the reported tests.} At equal delivered work, {BlockLock-Rows} reduces latency by 53.5--55.0\% and improves throughput by 115.2--122.2\% relative to vLLM Batch-Invariant; compared with batch-invariant kernels, it needs no custom kernel implementations and constrains only the protected requests.

FlatCake: A Flat KVCache Storage Architecture for LLM Serving

arXiv:2610.38355v1 Announce Type: new Abstract: KV cache reuse accelerates LLM serving, but existing systems manage host DRAM and NVMe SSD as a tiered hierarchy where DRAM acts as an intermediate cache. Production LLM workloads break both premises of this hierarchy. First, bimodal reuse defeats DRAM temporal filtering, since workloads are dominated either by one-shot blocks with long reuse distances or by multi-gigabyte agentic sessions that exceed DRAM capacity, putting SSD on the critical read path. Second, DRAM staging and eviction-ordered layout bottleneck SSD restoration. Holding host staging buffers for in-flight prefixes causes contention under load that aborts prefetches, while incremental eviction scatters a session's blocks across the SSD and prevents large sequential reads. We p

Merge-Based Vocabulary Extension Stalls at the Pre-Tokenizer: An English-Preserving Repair for Nepali, Tested in Continued Pretraining

arXiv:2610.33016v1 Announce Type: new Abstract: Llama-3, Qwen3 and many other open LLMs pre-tokenize text with a regular expression whose word class admits only Unicode letters. Devanagari vowel signs are combining marks, so this regex splits Nepali words at every vowel sign, and models that use it spend 2.6 to 5.8 times as many tokens on Nepali as on the same English text. Merge-based vocabulary extension cannot join the split pieces: for Llama-3 and Qwen3 no number of new merges brings this premium below about 2.2, and ten released extensions already end within 4% of this pre-token floor. We give a small edit to the deployed regex that provably keeps every token id of text without combining marks or Devanagari characters. With 32,000 new merges it brings Nepali to 1.01 times English, abo

Prefix Keys and Values Matter Late: RIPE Caching for Self-Speculative Decoding in Looped Language Models

arXiv:2610.34436v1 Announce Type: new Abstract: A looped language model runs one block of layers $R$ times per token, so its key-value (KV) cache holds $R$ planes per recurrent layer, and self-speculative decoding (draft with a few loops, verify with all $R$) reads them all. We find that prefix keys and values matter late: at equal perturbation size, loops 25-32 are $e^{15.0}$ (Huginn-0125) and $e^{12.8}$ (Raven-Llama) times more sensitive than loops 2-9, and the planes approach the last one geometrically, at 0.88 and 0.81 per loop. Whether the deepest plane alone suffices then depends on how ripe the late planes are: on Raven, storing only it is indistinguishable from a bf16-to-fp32 change; on Huginn, 1% of positions carry 74% of its error. We turn this into RIPE, which stores the deepest

Toward Open ISA Supercomputing: A Comparative Study of RISC-V, Arm, and x86 Architectures on HPC Workloads

arXiv:2610.32632v1 Announce Type: new Abstract: RISC-V is emerging as a candidate architecture for high-performance computing (HPC), but its competitiveness against established x86 and Arm processors remains difficult to assess because existing studies use different platforms, workloads, and methodologies, and rarely consider energy efficiency. Our work provides an evaluation of contemporary HPC RISC-V, Arm, and x86 processors using a common software stack and methodology. We characterize the Sophgo SG2044 and compare it with the NVIDIA Grace CPU Superchip and Intel Xeon Platinum 8480+ across four HPC benchmarks (STREAM, HPL, HPL-MxP, and HPCG) and four representative applications (iPIC3D, PLUTO, OpenFOAM, and llama.cpp). The version of HPL-MxP linked to OpenBLAS built and used for this wo

A new arXiv paper proposes an intelligent agent framework for traceable multimodal emotion reasoning, supporting evidence verification and prompt repair without weight updates.

On October 26, 2026, arXiv published a preprint paper (ID 2610.31685v1) that proposes a traceable multimodal emotion‑reasoning intelligent agent architecture based on frozen large models. By leveraging modality‑specific perception agents, an integration agent, an MLLM critic, and an evidence‑verification stage, the framework achieves error attribution and adaptive optimization of the prompt space, improving the F‑score for emotion recognition across several open‑source multimodal large models (with 2B–38B parameters) by 0.6–2.2 points on the MER2025 dataset.

Why it matters：This study addresses the trade-off between high instruction fine-tuning costs and the non‑debuggable pipeline of intelligent agents in current multimodal emotion‑reasoning systems by proposing the first end-to-end interpretable framework that integrates evidence citation, certainty verification, failure attribution, and targeted prompt repair. All improvements are implemented without modifying model weights and have been empirically validated across four open-source MLLMs of varying scales, demonstrating both methodological innovation and practical engineering utility.

Democratizing LLM Training with the Edge: Challenges, Systems, and the Road Ahead

arXiv:2610.32215v1 Announce Type: new Abstract: The growing computational power of modern GPU data centers has let a few industrial giants train and deploy large language models (LLMs) on massive amounts of data. Yet the sheer scale of these models and their training data makes centralized development increasingly hard to sustain, and at the same time shuts out smaller institutions that hold relevant private data but lack comparable infrastructure. In this paper, we organize our positions on distributed LLM training around the pretraining and fine-tuning paradigms, covering federated learning, split learning, and decentralized approaches, and we assess their readiness for geo-distributed settings with heterogeneous hardware and unreliable connectivity. We then present a {\em unified} syste

ULME Framework Achieves True Unification of Text Embedding and Language Modeling: A New Paper on arXiv Proposes a Predictive Coupling Mechanism

On October 24, 2026, arXiv published the paper arXiv:2610.30705v1, introducing ULME—a post-training framework for causal-decoder–based large language models that, through a predictive coupling mechanism, achieves, for the first time, a mutually reinforcing unification of text embedding and language-generation capabilities. This approach was validated on the Qwen2.5/Qwen3 series (with 0.5B–8B parameters), yielding MTEB improvements of 3.3–4.1 points while maintaining performance on generation tasks such as MMLU.

Why it matters：This paper directly addresses the core issue of fragmented multimodal LLM capabilities: embeddings and generation have long been handled either by separate models or by “pseudo-unified” models that rely solely on parameter sharing. ULME achieves a substantial architectural breakthrough by employing learnable slots and joint training with three objectives—including a stop-gradient prediction loss—enabling the generation process to actively shape the embedding representations. The study features broad experimental coverage, strong baselines, comprehensive metrics, and was published on a reputable preprint platform, balancing timeliness with technical rigor.

Research reveals the local geometric structure of semantic meaning in large language models

A paper published on October 24, 2026, on arXiv (ID 2610.30715v1) characterizes the local geometric properties of semantic meaning in embedding space across four decoder-only large language models—Gemma 4, Llama 3.2, Qwen 2.5, and Pythia—using cross-lingual probes and fine-grained probes.

Why it matters：This study is the first to systematically characterize the invariance, variability, and refinement mechanisms within “the same semantic factor” in LLMs, proposing the “local anisotropic neighborhood” model. The approach has been validated through blind human evaluations, demonstrating rigorous methodology, broad model coverage, and cross-architecture consistency in its findings.

Clinicians Lead the Development of Verifiable Safety Guidelines: Establishing a Traceable Layer of Safeguards for Mental Health Chatbots

On October 24, 2026, arXiv published the paper arXiv:2610.30690v1, which proposes a four‑rule safety‑check rulebook defined, finalized, and adjudicated under the guidance of two clinical physicians. After formal verification, this rulebook is deployed as a guardrail layer to govern the outputs of large language models (LLMs) in mental health contexts. All judgments within this guardrail can be traced back to a single naming rule and sixteen factual verifications, aiming to enhance the interpretability and testability of safety assessments.

Why it matters：This study offers a safety‑enhancing approach distinct from mainstream prompt engineering, LLM‑as‑a‑Judge, or Constitutional AI. Its core innovation lies in structuring and formalizing clinical decision‑making while maintaining a complete audit trail throughout the process. All key claims—such as the number of rules, validation methods, factual base, comparative experimental data, and cost metrics—are explicitly stated in the original text, and the publication date (2026‑10‑24) aligns with the current UTC date (2026‑10‑25), ensuring timeliness and a solid foundation of credibility.

New research reveals that open-book assessments primarily measure reading behavior rather than knowledge retention.

An arXiv paper published on October 24, 2026 (2610.31017v1) points out that, in document-level LoRA adapter evaluations, open-book scoring relies heavily on the model’s ability to read on the fly, rather than on whether the adapter has truly encoded knowledge specific to the document; random adapters can reproduce 82% of the performance advantage of expert adapters on open-book tasks, but on cloze tasks involving unstructured text, they only replicate 2%—statistically indistinguishable from zero.

Why it matters：This study, through systematic controlled experiments—comprising 150 Wikipedia articles, the Llama-3.1-8B-Instruct model, and a two-task comparison—reveals the fundamental limitations of the current open-book evaluation paradigm: it fails to effectively distinguish between “reading” and “remembering,” posing methodological challenges to the practical utility of knowledge-editing and retrieval‑oriented adapters. The paper also reports four silent errors that lead to “clean yet incorrect results,” highlighting risks to reproducibility.

FusionEval Released: A New Benchmark for Large Language Model Fusion

On October 24, 2026, researchers published the paper arXiv:2610.30162v1 on arXiv, introducing FusionEval—the first unified benchmark specifically designed to evaluate fusion methods for large language models. It covers vulnerable aspects such as reasoning chain length and tool invocation, and supports comparisons between fine-tuning-based and zero-shot fusion approaches.

Why it matters：This benchmark fills the gap left by existing evaluations, which inadequately cover long-chain reasoning and tool usage; it clearly distinguishes between average performance and worst-case capability retention, provides a reproducible expert pool (based on Qwen3/Llama), multi-dimensional metrics (including GPU memory and runtime), and open-source code, offering strong practical guidance.

Research on Modulating the Capability Distribution of Large Language Models via Singular Value Spectrum Intervention: Empirical Analysis of Llama and Qwen Models

An AI study published on October 24, 2026, on arXiv (ID 2610.30676v1) applied post-hoc singular-value spectrum smoothing and spike perturbations to five open-source Llama and Qwen instruction-tuned models, finding that this lightweight intervention can systematically reshape the trade-offs in model performance across three capability domains: factual knowledge, language understanding, and deductive reasoning.

Why it matters：This paper is the first to empirically demonstrate, within a multi‑model, multi‑task framework, the causal impact of the singular value spectrum of transformer feedforward layers on downstream capabilities, revealing structurally sensitive patterns in capability allocation and offering a new pathway for decoupling model capabilities and enabling controllable editing. All conclusions are based on publicly available models and standard benchmarks, with transparent methods and comprehensive evaluation dimensions.