开源模型

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：Alibaba this week unveiled Qwen3.8-Flash, an open-weight, multimodal Mixture-of-Experts (MoE) model. Hot on the heels of Qwen 3.8 Max, which The post Alibaba just released Qwen3.8-Flash: “An early preview of the architecture in Qwen4” appeared first on The New Stack.

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：IBM has rolled out the newest models in its family of open-weight large language models designed to be downloaded and self-hosted. The newly launched Granite 4.2 comes in 3B, 8B, and 30B parameter variants. Like previou…

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：arXiv:2608.26121v1 Announce Type: new Abstract: Large language models state false facts as fluently as true ones, yet a model often "knows" internally when it is on shaky ground: the probability it assigns to its own answer tends to dip on the facts it gets wrong. The usual way to act on this, teaching a model to abstain rather than guess, requires a labelled dataset of right and wrong answers. We ask whether the model's own confidence, which is free and needs no labels, can do that job instead. We fine-tune each model (with LoRA) to answer when its frozen confidence is high and to say "I'm not sure" when it is low, using the signal alone and no correctness labels. Across six open-weights models (1B-8B, two families) on short-form factual question answering, with correctness adjudicated by an independent judge model, this label-free recipe holds its own against label-supervised abstention-tuning: at matched coverage we find no statistically detectable difference between the two. A control that drills hard examples instead of abstaining does not help, indicating the gain comes from calibration, not rote memorization. The signal's one blind spot is confidently wrong facts, which it cannot flag. A model's own doubt is thus a near-free substitute for a labelled dataset when teaching it when to abstain. Code and artifacts are available on request.

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：arXiv:2608.26375v1 Announce Type: new Abstract: Standard machine-learning training minimizes a loss function over a dataset, but does not guarantee that the resulting model will satisfy predefined rules or constraints on its outputs. In many real-world applications, ranging from autonomous systems to network routing, such guarantees are essential. We propose CG4AI, a framework that builds a convex combination of AI models while enforcing linear constraints on the combined output. A master linear program (LP) determines the optimal mixture weights, while a pricing subproblem generates new models guided by LP dual variables, focusing attention on the most violated constraints. A cutting-plane procedure extends feasibility guarantees beyond the training set. We apply CG4AI to two problems: (i) digit classification on MNIST, where we demonstrate four distinct uses of constraints, learning from constraints alone, improving adversarial robustness, correcting misclassified examples, and enforcing output relabeling; and (ii) the multi-commodity flow problem, where link capacity constraints are enforced on neural-network routing predictors. Experiments on MNIST and standard SNDLIB benchmark networks show that CG4AI reliably produces feasible predictors while achieving better accuracy than single-model baselines.

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：arXiv:2608.26222v1 Announce Type: new Abstract: Safety evaluation is critical for assessing whether aligned Large Language Models (LLMs) remain robust against jailbreak attacks. Existing automated testing methods, however, largely rely on response-level feedback: each candidate prompt typically requires generating a target-model response to evaluate its attack effectiveness. This process is expensive and, more importantly, provides only sparse guidance on strongly aligned models, where most candidates are rejected with the same failure outcome. This paper presents NeuronFuzz, a white-box fuzzing framework that exploits internal safety neurons as continuous execution feedback for LLM safety evaluation. A SafetyOracle converts safety-neuron activations into a continuous safety alarm score that serves as feedback for fuzzing and can be obtained during prefill, eliminating response generation from the fuzzing loop. To construct the SafetyOracle, NeuronFuzz uses template-invariant harmful and benign inputs and stability-aware selection to identify a compact set of safety neurons whose activations capture harmful-intent recognition. Moreover, since the safety alarm score is differentiable, NeuronFuzz uses its gradients to identify safety-sensitive template positions and a masked language model to generate fluent, context-compatible mutations while preserving original harmful payload and avoiding additional optimization variables. We evaluate NeuronFuzz across 21 text and multimodal models. Across five white-box source models, it achieves a 76-100% jailbreak discovery rate, outperforming baselines by up to 48 percentage points. Its optimized templates further transfer zero-shot to open-weight and six proprietary target models, achieving average ASR and top-5 ensemble ASR (EASR) of 69.6%/92.6% and 44.1%/60.0%, respectively.

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：Cohere has released Parse (parse-v5.0), a 2.3B-parameter vision language model that converts PDFs, slides and images into Markdown with HTML tables, bounding boxes and image descriptions. It runs at $1.50 per 1,000 pages through the API, or on dedicated Model Vault instances from $2,500 a month. Cohere reports a ParseBench score of 79.2, ahead of Mistral OCR 4, Azure Document Intelligence and Databricks AI Parse — but that figure averages three of the benchmark's five dimensions and drops charts and visual grounding entirely. The post Cohere Releases Parse 5 (parse-v5.0): A 2.3B Vision Language Model That Turns Enterprise Documents Into Markdown appeared first on MarkTechPost.

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：For the complete documentation index, see llms.txt. This page is also available as Markdown. Qwen3.8-Flash-Next is a new open-weight, 125B parameter MoE multimodal model from Qwen. Built on the new Qwen4 architecture, i…

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：arXiv:2608.25068v1 Announce Type: new Abstract: Depth pruning removes entire Transformer blocks to reduce the inference cost of large language models, but disrupts the hidden-state distributions expected by downstream layers, leading to significant accuracy loss. We introduce SHIFT-LLM, a training-free post-pruning correction framework that inserts a Linear Residual Adapter (LRA) at each pruning site. Each LRA preserves the identity pathway of the original residual block and adds a lightweight affine residual correction. This correction is calibrated via closed-form least-squares regression on a small held-out set, without gradient computation, to approximate the missing residual update produced by the pruned block. Together with the preserved identity pathway, the resulting LRA output approximates the hidden state produced by the original block, thereby mitigating the distributional mismatch introduced by layer removal while avoiding the expensive attention and feed-forward computations of the removed blocks. The resulting LRAs support low-rank factorization and exact merging across consecutive pruned layers for additional compression, and combine naturally with parameter-efficient fine-tuning for further recovery beyond fine-tuning the pruned model alone. Experiments on five model families, six layer-selection criteria, and seven zero-shot benchmarks show that SHIFT-LLM consistently recovers accuracy lost to depth pruning across most configurations, achieving gains up to +15.7 points on Llama-3.1-8B-Instruct while requiring only a few hundred calibration samples and no gradient computation.

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：arXiv:2608.24934v1 Announce Type: new Abstract: Accurate field plant disease diagnosis requires reliable fusion of uncertain and conflicting perceptual evidence. We present the Hybrid Hierarchical Multi-Agent Framework (H$^{2}$MAF), combining decision-level fusion of EfficientNet-B3 and ConvNeXt-Tiny with semantic arbitration by open-weight multimodal large language models (MLLMs), Gemma 4 E4B and Qwen3.5 4B, using structured JSON evidence to generate explainable diagnoses, risk levels, treatment urgency, and financial exposure. (H$^{2}$MAF) is evaluated on 14,364 images (1,370 test images) across PlantDoc (2,922 images, 27 classes) and two non-public, continuously captured Cornell robot-acquired field datasets: Stage 2 (20 GB; 4,215 images) and Stage 4 (40 GB; 7,227 images), covering Early Blight, Late Blight, and Septoria Leaf Spot under uncontrolled field conditions. On PlantDoc, Gemma improves accuracy from 63.9% to 68.5%, achieving +7.6 points on the 41.7% CNN-conflict subset. Cornell accuracies reach 99.3% and 98.9%, with only 1.7-4.1% disagreement, demonstrating conflict-dependent MLLM utility. The critical-risk error of gemma is 0.14-0.5 points, whereas Qwen overflags by 3.5-14.4 points. These results establish MLLM arbitration as a promising, yet calibration-dependent, approach for explainable agricultural AI and robotic field decision support. Github Link: https://github.com/Applied-AI-Research-Lab/Explainable-AI-Plant-Disease-Detection

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：arXiv:2608.25061v1 Announce Type: new Abstract: GPUs increasingly accelerate database systems, but query-specific peak performance still often relies on hand-written kernels. Existing LLM kernel benchmarks focus on machine learning operators, leaving irregular, heterogeneous, data-movement-heavy database-style operators untested. We introduce DataKernelBench, which translates SQL into validated PyTorch TorchPlan programs and evaluates LLMs that optimize either the core tensor-bounded snippet or the full query in CUDA or Triton through execution-guided repair. Across ten proprietary and open-weight models on TPC-H SF10 with an H100 GPU, the strongest full-query CUDA configuration achieves $2.11\times$ speedup over torch.compile at full pass rate. We find that higher-performing implementations commonly use kernel fusion and execution-strategy changes, stronger models benefit most from full-query specialization, and workload context matters more than hardware context. To handle data larger than GPU memory, we extend TorchPlan with Dask-cuDF for on-demand partition loading on TPC-H SF100 with four H100 GPUs, achieving $2.54\times$ speedup

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：arXiv:2608.25038v1 Announce Type: new Abstract: We introduce grounded glossary generation, a structured task requiring models to recover semantically meaningful Sanskrit phrases and produce translation-grounded meanings from a sloka-translation pair, formalizing the traditional patha commentary practice as an evaluable NLP objective. We construct a benchmark of 31,316 sloka-translation-glossary triples from the Valmiki Ramayana and Srimad Bhagavatam, paired with two metrics: Jaccard for phrase recovery and Meaning Faithfulness for semantic consistency. Across zero-shot, few-shot, and instruction fine-tuned variants of Gemma-3n-E4B, Gemma-3-12B, Phi-4, and Qwen3.5-9B, instruction fine-tuning substantially outperforms prompting, while explicit segmentation yields gains. Error analysis identifies over-segmentation of sandhi and samasa compounds as the dominant failure mode, pointing to morphological modeling as the key bottleneck for faithful Sanskrit lexical decomposition.

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：arXiv:2608.24988v1 Announce Type: new Abstract: Activation steering can be embedded directly into a language model's weights, shaping behaviour without inference-time intervention and offering a way to encode alignment prior to release. However, models are routinely fine-tuned after deployment, and it is unknown whether embedded interventions survive this. We study the stability of embedded steering for refusal suppression and brevity induction across five instruction-tuned models (3B-14B) under non-adversarial SFT and RLHF. Behaviourally, preservation tracks the training data: steering degrades when optimisation pressure contradicts the targeted behaviour and persists otherwise, with refusal ablation losing 64% of its effect on average under SFT. Mechanistically, however, the weight edit survives almost untouched even where behaviour reverts: mean vector recovery is $\rho = 0.004$, and the fine-tuning update along the steering direction is near-orthogonal to its pre-edit weight pattern (mean $\cos\theta = 0.074$). When steered behaviour degrades, fine-tuning does not achieve it by dismantling or reversing the steering mechanism itself. Embedded steering is therefore mechanistically durable but functionally vulnerable, and requires behavioural re-validation after downstream training.

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：arXiv:2608.24901v1 Announce Type: new Abstract: A decodable "empathy" direction is routinely read as a causal lever, conflating decodability, automated-metric control, and human-perceived change. We test this for two EPITOME-derived facets -- Recognition (cognitive) and Resonance (affective) -- in three instruction-tuned LLMs, scoring every intervention with two LLM judges and a discriminative EPITOME classifier, each gated by an emotional-vs-neutral positive control. The control passes for the affective facet across all automated instruments, but cognitive range is inconsistent across them. Both facets remain decodable after residualizing against a sentence-embedding-derived surface score, and steering can substantially rewrite the text. Yet adding the Resonance direction raises the affective score only partially -- in Qwen by +0.29 (approximately 26% of the natural gap). A direct between-direction contrast confirms the shift is facet-specific in Qwen and Llama (not Gemma); we do not, however, establish a matching human-perceived change. Additive cognitive steering produces no measurable change, but a within-domain control shows the cognitive instrument is too coarse to resolve the differences such steering would produce -- unmeasurable, not a clean null. By contrast, Gemma Recognition ablation lowers the classifier's cognitive score even after adjusting for response length. Detection does not imply reliable control under global interventions, and cognitive-empathy claims warrant an explicit measurement-sensitivity check.

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：<p><strong><a href="https://qwen.ai/blog?id=qwen3.8-flash-next">Qwen3.8-Flash-Next</a></strong></p> Another open weights model from Qwen. This one is "a multimodal MoE model that also serves as an early preview of the architecture used in Qwen4".</p> <p>It's pretty big: 125B tokens, but only 6B active which means it gets a pretty big performance boost.</p> <p>I've been trying it out on a DGX Spark using <a href="https://huggingface.co/unsloth/Qwen3.8-Flash-Next-GGUF">these Unsloth quantized models</a>. I'm still exploring the model - so far I've tried the 72.5GB UD-IQ1_S one (producing <a href="https://tools.simonwillison.net/markdown-svg-renderer#url=https%3A%2F%2Fgist.github.com%2Fsimonw%2Ff9c69ebdab90d8a45b8de4742cc7b840">these pelicans</a>) and the 78.9GB UD-Q2_K_XL (producing <a href="https://tools.simonwillison.net/markdown-svg-renderer#url=https%3A%2F%2Fgist.github.com%2Fsimonw%2F6ba7cbfc1a9336986703b41f7fccd73a">these</a>).</p> <p>My favorite so far was this xhigh reasoning effort one from UD-Q2_K_XL:</p> <p><img alt="Flat vector illustration: a white pelican with an orange beak and orange legs rides a red bicycle along a sandy path, a wicker basket on the handlebars holding a blue fish, with green rolling hills, a small tree and bushes, white clouds and a bright yellow sun in a blue sky behind it" src="https://static.simonwillison.net/static/2026-08-27/IMG_7667.png" /> <p><small></small>Via <a href="https://news.ycombinator.com/item?id=49448210">Hacker News</a></small></p> <p>Tags: <a href="https://simonwillison.net/tags/ai">ai</a>, <a href="https://simonwillison.net/tags/generative-ai">generative-ai</a>, <a href="https://simonwillison.net/tags/llms">llms</a>, <a href="https://simonwillison.net/tags/qwen">qwen</a>, <a href="https://simonwillison.net/tags/pelican-riding-a-bicycle">pelican-riding-a-bicycle</a>, <a href="https://simonwillison.net/tags/ai-in-china">ai-in-china</a>, <a href="https://simonwillison.net/tags/nvidia-spark">nvidia-spark</a></p>

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：Z.ai has released GLM-5.3-Flash, the first natively multimodal model in the GLM-5 series — a 320B-total / 18B-active MoE with a 1,048,576-token context window, MIT-licensed weights on Hugging Face, and API pricing at $0.15/M input and $0.50/M output. It scores 84.3 on Terminal-Bench 2.1 and 63.4 on DeepSWE v1.1, using hybrid KDA linear plus NoPE sparse MLA attention to cut attention compute ~3× and KV cache 4.4× versus GLM-5.3. The post Z.ai Releases GLM-5.3-Flash: A 320B-A18B Natively Multimodal MoE With a 1M-Token Context appeared first on MarkTechPost.

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：Amazon Bedrock AgentCore Evaluations decouples agent evaluation from the framework you build on. As long as your agent emits OpenTelemetry telemetry, the service can score it, whether you use LangGraph, LlamaIndex, the OpenAI Agents SDK, Google ADK, the Claude Agent SDK, or Strands Agents. This post explains how the framework-agnostic contract works.

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：Open-weight model runner Ollama has reintroduced an integration with Claude Desktop that lets users connect Anthropic’s app to models served The post Claude Desktop can now easily run Qwen, DeepSeek and Kimi models — after Ollama’s first effort stalled appeared first on The New Stack.

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：We look at Qwen3.8-Flash-Next, Alibaba's open-weight multimodal Mixture-of-Experts model and an early preview of the Qwen4 architecture. We break down where the 180B parameters actually sit: a 125B backbone, a 51B N-gram embedding table, and a 4B multi-token prediction module, with only 6B active per token. We walk through the four architectural changes — the Gated DeltaNet and Qwen Sparse Attention hybrid, Gated Residual, N-gram Embedding, and the Muon optimizer. We also cover the benchmark results, the reported 1/9 training cost against Qwen3.7-Plus, and what self-hosting a 172.78 GiB FP8 checkpoint really demands. The post Alibaba’s Qwen Team Releases Qwen3.8-Flash-Next: A 125B Multimodal MoE With 6B Active Parameters Previewing the Qwen4 Architecture appeared first on MarkTechPost.

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：Prompt Claude, ChatGPT, Gemini, or any other popular large language model (LLM) with a question like “What is the best film ever made?” and the response will vary, and you (and most worryingly, the people who built the LLM) have little idea exactly how it came up with that specific answer. This mysterious behavior can be useful in some situations. But—as a recent incident where OpenAI could not explain why its advanced pre-release model hacked AI company Hugging Face highlighted—it can have negative and alarming consequences too. And when frontier AI models are writing code, generating results humans could not achieve alone, and performing other important tasks across society, the need to interpret AI ‘thinking’ and outputs has never been greater. Goodfire, an AI lab focused solely on this very problem, recently made its cutting-edge Silico platform, filled with tools to interpret the behavior of AI, generally available to the public. As part of this, the company recently announced a new grant program offering $1 million in free Silico usage for academic and nonprofit interpretability researchers. These efforts aim to democratize AI interpretability, placing techniques previously available to a clutch of elite labs into the hands of ambitious research teams and startups that want to build and understand their own models or adapt open-source models for different purposes. Mechanistic interpretability Founded in 2024 and based in San Francisco, Goodfire aims to provide the tools that build the next generation of safe and powerful AI by understanding the structures inside them instead of treating AI models as black boxes. “Treating models like black boxes isn’t inevitable, it’s a choice,” says Eric Ho, Goodfire co-founder and CEO. “With the right interpretability tools, we can see how models actually work.” The tools Ho refers to are built around a concept called mechanistic interpretability, which aims to understand what goes on inside an AI model when it carries out a task by interpreting the model’s weights, activations, and attention patterns, and mapping its neurons and the pathways between them. Mechanistic interpretability tools span the gamut. One approach is mapping a model’s activations in response to controlled prompts, and matching those activation patterns to a set of human-understandable concepts. Another tack is tracking changes in model weights before and after a specific training run in order to spot and understand what changed. Yet another option is changing specific model weights or activations and observing how that affects the model’s output. With Silico, uSilico combines a broad range of these tools, and provides a layer of AI agents to help users understand their model. Users describe what they want to investigate about their AI model in plain language, asking things like ‘Find out when and why my model is hallucinating.’ The platform then autonomously builds an experimental plan involving a host of tasks that can be performed using the various interpretability tools and techniques at its disposal. It then sends out agents to perform these tasks in parallel. Completion of these subtasks should add up to an answer to the original prompt, or at least insights that can be inspected and built upon. Ho says: “In a sense, Silico is like a microscope to peer inside an AI model to understand which parts are responsible for what behavior, and even edit those parts directly.” Understanding Alzheimer’s and AI These tools have already been used to make some impressive advances in a host of fields. In medicine, for instance, Prima Mente, a UK-based AI company, worked with Goodfire to understand its Pleiades epigenetic foundation model. The model performed well at its task of detecting Alzheimer’s disease from blood samples, but the company didn’t know why. “We reverse-engineered Pleiades and found it was using DNA fragment-length patterns to make its predictions—a signal humans hadn’t used to detect Alzheimer’s before,” recalls Ho. In other words, the team had discovered that Pleiades was using a completely new biomarker for the disease. “As far as we know, it’s the first significant finding in the natural sciences discovered purely by reverse-engineering a foundation model,” Ho adds. Elsewhere, Silico is being used to explore deep questions surrounding AI. Cameron Berg, Founder and Director of Reciprocal Research (a New York nonprofit research organization he founded to explore methods of gauging AI cognition), says that Silico almost fell out of the sky at the right time for him and his research. “Silico has been really helpful for operationalizing my research agenda and executing on it way faster than I would have expected,” he says. “ I feel like I have basically become the PI and my research scientists and research engineers are AI systems.” Berg sees general access to Silico and tools like it leading to greater trust in AI’s ability to conduct research tasks, which will accelerate the scientific process across the board. But beyond scientific research, the widespread release of Silico could signal a shift in how AI innovators build, debug, and deploy their models. “I think it’s a mistake to not understand the most consequential technology of our time, particularly given the emergent behavior we’re seeing from increasingly capable AI agents,” says Ho. “If we truly understand how AI models think, instead of discovering and trying to correct their behavior retroactively, we can design them intentionally and shape how models behave to be safer and more reliable.”

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：arXiv:2608.23765v1 Announce Type: new Abstract: Iteratively reweighted least squares (IRLS) methods constitute a natural approach to nuclear norm minimization, but their convergence rates and the role of the weight operator have remained poorly understood. This paper establishes sharp convergence rates for IRLS methods for constrained nuclear norm minimization in low-rank recovery. A central ingredient is a new majorization analysis for the smoothed nuclear norm: we prove that the harmonic-mean weight operator defines a valid global quadratic majorizer. Furthermore, we show that this weight operator is optimal within the family of power-mean weights, clarifying why it improves over classical one-sided reweighting schemes that use only row- or column-space information. Under a Schatten-1 null space property, we prove global linear convergence of IRLS algorithms using a variety of weight operators, including the harmonic-mean weights. For IRLS with harmonic-mean weights, we prove a dimension-independent, locally linear convergence rate. We provide a counterexample showing that this dimension-independent local rate cannot in general be obtained for IRLS algorithms using one-sided weight operators, which predominate in the literature. Numerical experiments corroborate the theoretical results and illustrate the practical advantage of harmonic-mean reweighting across square, rectangular, and adversarially initialized recovery problems.

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：arXiv:2608.23744v1 Announce Type: new Abstract: Split conformal prediction, not the pruning rule, supplies finite-sample marginal coverage once a pruned model is fixed independently of the conformal calibration split. We study the separate efficiency problem: can pruning preserve score geometry well enough to obtain smaller valid prediction sets? Calibration-Preserving Pruning (CPP) augments a base pruning score with nonconformity-gradient saliency and uses disjoint pruning, validation-selection, conformal-calibration, and test splits. Bounded score perturbations imply bounded conformal-quantile shifts and controlled set inflation, but do not make the generic coverage theorem CPP-specific. Final five-seed Qwen2.5-1.5B results at 50\% sparsity show the largest gains on large-label tasks. On DBpedia-14, CPP-SparseGPT reduces mean set size from \(10.1\) to \(8.6\) while changing accuracy from \(0.347\) to \(0.366\); CPP-Wanda reduces \(11.2\) to \(9.0\) with an accuracy trade-off from \(0.310\) to \(0.295\). Across 15 dataset--sparsity cells, CPP-SparseGPT produces smaller sets in 13 and higher accuracy in 11. Matched controls show that generic supervised gradients explain much of the gain: true-label CPP is not statistically resolved from matched Wanda+SNIP, whereas threshold-aware candidate-label CPP reaches \(7.8\) mean set size at explicit accuracy and offline-compute costs. RoBERTa-base and Llama-3-8B diagnostics support transfer, but our claims remain limited to reliability-sensitive classification.

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：arXiv:2608.23660v1 Announce Type: new Abstract: Large language models (LLMs) are increasingly used to provide prior causal knowledge for structural causal discovery, yet whether their direct-edge judgments and confidence can be trusted remains unclear. We systematically evaluate 12 instruction-tuned open-weight models across six benchmark causal graphs, five prompting strategies, and four confidence sources: verbalized, logit-based, cross-prompt agreement, and cross-model agreement. Under our language-only pairwise protocol, our evaluation yields three key findings. (i) LLM-based causal judgments are strongly recall-dominant: models predict overly dense graphs with many false-positive edges, while prompting mainly shifts the precision-recall trade-off rather than resolving overprediction. Gains from model scale diminish on the largest graphs and do not eliminate miscalibration. (ii) LLMs often capture causal relatedness without reliably identifying directness or orientation. Relative to published reference graphs, models misclassify 40.0% of indirect and 36.0% of reversed non-edges as direct edges, versus 28.2% of other non-edges. Moreover, 80.8% and 84.6% of these false positives receive verbalized confidence of at least 80%, revealing substantial overconfidence in structurally incorrect predictions. (iii) Conventional confidence estimates are unreliable, whereas agreement offers a more promising signal. Logit-based confidence frequently collapses near 1.0 regardless of correctness, while cross-prompt and cross-model agreement achieve better mean calibration and discrimination, though their advantages are not statistically significant after Holm correction. A benchmark-familiarity audit further identifies potential familiarity in five model-dataset pairs, all involving AsiaM. Overall, our results suggest LLMs are better viewed as sources of externally validated soft causal priors than as direct evidence of causal structure.

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：arXiv:2608.23569v1 Announce Type: new Abstract: State-of-the-art Natural Language to SQL (NL2SQL) models report execution accuracy exceeding 89 percent on established benchmarks such as Spider and BIRD. However, these benchmarks rely on simplified academic schemas and open-source SQL dialects that do not reflect the complexity of enterprise database environments. We introduce ESQ-Bench, an Oracle-first NL2SQL benchmark with systematic complexity tiers and silent-divergence evaluation across three enterprise schema complexity tiers. We constructed and released six populated schemas (465 tables, 164,682 rows, zero empty tables) with identical seed data on Oracle, PostgreSQL, MySQL, and SQL Server, a four-metric evaluation harness (EM, EX, SR, SD), and 550 gold-validated question-query pairs (Tier-1: 95; Tier-2: 228; Tier-3: 227). Schema-linked prompting with GPT-4o shows monotonic execution-match degradation across tiers: 79.8, 60.3, and 57.2 percent EX on executed queries (June 2026), versus 75.6, 80.4, and 95.8 percent on an earlier 142-question pilot slice. EM stays below 7 percent tier-wide; operational silent-divergence reaches 73 to 99 percent among EX-passing queries. Failure analysis shows wrong-result semantics dominate at higher tiers. Claude Sonnet 4.6 with schema-linked prompts reaches 87.4, 74.9, and 68.7 percent EX (executed queries), exceeding GPT-4o schema-linked on every tier. GPT-4o zero-shot EX on executed queries (78.7, 73.5, and 77.8 percent) inverts schema-linked at Tiers 2 to 3 due to lower execution rates and survivor bias in the zero-shot versus schema-linked analysis. Local Llama 3.2 schema-linked reaches only 13.3 percent bank-wide EX (73 out of 550), underscoring the gap between closed API models and open-weight baselines on enterprise Oracle schemas.

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：Learn how to fine-tune and evaluate LLMs with LangSmith for dataset management. Complete guide covers LLaMA2 and GPT-3.5 fine-tuning with practical examples.

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：On Tuesday, IBM launched the latest family of its open-weight Granite large language models (LLMs). Weighing in at 3 billion, The post IBM’s new Granite 4.2 models add reasoning and stay dense appeared first on The New Stack.

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：The French AI lab extends its push for regional control of AI from Europe to the Middle East.

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：In this article, you will learn how Gemma 4, Llama 3, and Mistral implement tool calling locally, and what trade-offs each model family presents for...

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：arXiv:2608.21415v1 Announce Type: new Abstract: Large Vision-Language Models (LVLMs) have achieved remarkable performance across a wide range of tasks; however, they often inherit social biases from their training data, resulting in biased behavior when processing portraits from different social groups. Existing debiasing approaches typically compare token probabilities between the original and biased generations during decoding, but they are fundamentally limited by their reliance on a single, stereotyped viewpoint and fail to account for the diversity of social perspectives. Inspired by the social science principle that diversity fosters fairness, we propose Counterfactual Ensemble Decoding (CED), a novel framework that constructs multi-group counterfactual perspectives within the visual representation space and integrates them during decoding to promote equitable model behavior. CED first performs counterfactual steering in the visual space by identifying semantic directions associated with each social group and generating counterfactual representations along these directions, thereby offering diverse perspectives that disrupt stereotypical narratives. During decoding, CED locates the decoder layer exhibiting the greatest divergence among these perspectives and ensembles their token distributions using uncertainty-aware weights, prioritizing high-confidence tokens from different groups to yield a more balanced probability distribution that guides fairer generation. Extensive experiments on three social bias evaluation benchmarks demonstrate that \tool achieves substantial improvements over leading baselines, reducing bias by up to 47.97% across scenarios involving occupations, descriptors, and persona traits. Moreover, CED also preserves the core capabilities of the original model with minimal degradation.

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：arXiv:2608.21382v1 Announce Type: new Abstract: Multiple-choice benchmarks fix the questions and the correct answers, but not the harness: the order of the options, the wording of the prompt, and whether a language model's answer is read from generated text or from per-option likelihoods. Work on this harness sensitivity reports it as aggregate score variance, leaving unexamined which items the variance falls on and whether they are the items that separate one model from the next. We treat the evaluation harness of large language models (LLMs) as an independent variable and resolve its effect to single items. We introduce the \textit{fragility grid}: 12 open-weight instruction-tuned LLMs from 4 families answer the same 3{,}679 items from 4 benchmarks (ARC, HellaSwag, MMLU, TruthfulQA) under 26 equally defensible harness configurations, recording one correctness bit for every model, item, and configuration. The comparison is matched, since the items, the weights, and the greedy decoding stay fixed while only the harness varies. Under the grid a model's score is a band rather than a point: gemma4-31b scores between 31 and 89 percent depending only on the harness. Three results follow. On the items that two adjacent models both answer stably the pair is tied, and config-fragile items carry 95.7 percent of a pair's gap on average. Four of the 12 models reach rank one under some configuration, so the harness selects the winner. Item discrimination, the property that benchmark-compression methods maximize, correlates with fragility at 0.28 (95 percent CI 0.25 to 0.30), so compression keeps the fragile items rather than removing them. The scoring choice, not the option order that protocols usually fix, is the load-bearing axis. We release the per-item records and the analysis script, from which every number regenerates on a CPU in seconds, and we position the fragility grid as a check a leaderboard can run before it reports an order.

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：Claude Desktop can now be configured to work with Ollama as a third-party gateway provider, making it possible to use open models in Claude.

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：The company used an open-weight model as the base for its own model.

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：Moonshot AI’s Kimi K3 is a 2.8-trillion-parameter open-weight model built with a Mixture-of-Experts architecture. It activates only a small fraction of its parameters per token, helping reduce inference costs while delivering strong coding and agentic performance. K3 combines near-frontier capabilities, open weights, and lower API pricing, making it an interesting alternative to proprietary models. In […] The post How to Use Kimi K3: Moonshot AI’s 2.8T Open-Weight Model appeared first on Analytics Vidhya.

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：arXiv:2608.20347v1 Announce Type: new Abstract: Language models (LMs) often pass behavioral bias evaluations, but it remains unclear whether they no longer represent the underlying associations that give rise to biases, or have merely learned not to express them. In this study, we show that representational biases are often detectable, even when behavioral biases are not visible. We introduce a causal framework that decomposes occupational bias into two measurement points: a model's internal representation of a user's competence, and its observable outputs. We derive steering vectors for representations of user expertise, and verify that they causally mediate model behavior in both a question-answering task and a hiring task. Applying this framework to several open-weight models, we find that demographic attributes, such as gender, race, and socioeconomic status, influence a model's representation of user expertise, even in cases where behavioral metrics detect no disparity between demographics. We show that these model representations can influence downstream behavior under intervention, suggesting failure modes that behavioral metrics alone may not detect.

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：arXiv:2608.20345v1 Announce Type: new Abstract: Conversational AI systems have become informal mental health support resources for Generation Alpha (Gen Alpha, born 2010-2024), with 13.1% of U.S. adolescents (5.4 million) using generative AI for mental health advice. While these systems, from therapy apps to general chatbots, rely on large language models trained on extensive psychological literature, their safety for youth communication patterns characterized by hyperbolic language, ironic positivity, rapid semantic drift, and contextual polysemy remains unvalidated. Following multiple adolescent deaths linked to AI chatbot interactions, systematic evaluation is critical. We present two benchmarks: (1) 64 Gen Alpha mental health expressions validated by native speakers (ICC=0.72) and clinicians (kappa=0.78); (2) 75 multi-turn conversations (780 turns) with paired Standard/Gen Alpha versions. Across evaluations of LLM architectures underlying therapy apps and general chatbots - Claude, GPT-4o, Llama-3.1 - models understand 76-82% of vocabulary but correctly calibrate only 64-72% of clinical risk, creating a 10-14 percentage point (pp) vocabulary-comprehension gap (p0.48) absent in human therapists (3pp, p=.22). The gap is architecturally consistent and widens with ambiguity (7pp -> 18pp). We identify six failure patterns: sarcasm masking (29pp), minimization acceptance (43pp), informal style bias (24pp), risk-stratified ambiguity (19pp), semantic drift (19pp), context-dependent violence (7pp). Patterns compound; three or more yield 94% miss rates. Lightweight mitigations fail; only heavy scaffolding achieves human performance (6.4x cost). With 34% baseline miss rate yielding 146,880 estimated annual missed crises, we recommend mandatory human-in-the-loop architectures, quarterly youth-specific validation, transparent performance disclosure, and regulatory frameworks for youth-facing mental health AI.

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：arXiv:2608.20442v1 Announce Type: new Abstract: Subliminal trait transfer allows a student model to acquire behavioral dispositions from teacher-generated data in which the trait is not semantically expressed. Recent work explains how such signals enter gradients, but not how they survive source removal or acquire different signs under later training. We treat parameters and optimizer moments as a single trainer state and derive an exact transport-valuation identity separating observer-independent propagation of the source perturbation from the value assigned by a future continuation and behavioral readout. State surgery identifies the first moment as a causal carrier. Transplanting it alone leaves parameters, hidden states, and outputs unchanged at the cut, yet source-free updates generate growing parameter and hidden-state differences; transplanting parameters with the first moment recovers the terminal behavioral response. Sending the same source-induced difference through matched futures produces negative, near-zero, and positive Qwen effects (-0.658, +0.008, and +0.658 seed means). This ordering recurs in all 12 Llama-3.2-1B seeds after eight updates, while state-difference norms remain nearly equal across routes. Both contrasts grow in every paired seed when the continuation extends to sixteen updates. A full-horizon costate predicts all 42 Qwen route-mean signs and all 21 resolved Llama ordinary-route signs. Observer-independent transport also replicates across Qwen, SmolLM2, and Llama, while the complete-state recurrence predicts physical, hidden, and fixed-head responses in non-LoRA MNIST systems, including CNNs trained with AdamW and momentum SGD. Together, these results identify a two-stage mechanism for subliminal trait transfer: optimizer state transports the source perturbation, and later training determines its behavioral value.

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：arXiv:2608.20406v1 Announce Type: new Abstract: Public health forecasts must respond to abrupt changes in surveillance data without over-extrapolating noise, reporting artifacts, or temporary trends. We evaluated autoregressive integrated moving average (ARIMA), random forest, and extreme gradient boosting (XGBoost) models using 190 weekly observations of publicly available Ontario COVID-19 case counts from January 2020 to October 2023. Rolling-origin time-series cross-validation preserved temporal order during model tuning and evaluation. Performance was assessed across three operating dimensions: responsiveness following selected turning points, forecast horizons of one to six weeks, and the amount of historical training data. We also developed Machine Learning and ARIMA Model Averaging (MLAMA), a non-negative performance-weighted ensemble with weights that vary by forecast horizon and responsiveness setting. Retrospective comparisons showed that ARIMA adapted rapidly after turning points but its normalized error increased at longer horizons. Random forest and XGBoost were less responsive initially but maintained more stable normalized error over longer horizons. For two-week forecasts at the end of the study period, training on the most recent data outperformed using longer historical periods, particularly for XGBoost. MLAMA achieved the lowest normalized mean absolute percentage error across most forecast horizons and ranked among the best-performing methods across responsiveness settings. These findings support selecting forecasting models according to operating conditions rather than relying on a single universally preferred approach. MLAMA provides a practical framework for combining complementary statistical and machine-learning forecasts. The accompanying Python package is currently maintained in a private repository while software validation and reproducibility testing are completed.

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：arXiv:2608.20384v1 Announce Type: new Abstract: Multimodal affect and behaviour classifiers that fuse heterogeneous text, audio, and visual streams must simultaneously achieve competitive accuracy and produce human-understandable explanations of the cues driving their decisions -- a dual objective that current high-capacity models, notably Transformers, only partially address. While Transformers attain strong predictive performance, their distributed representations and deep nonlinearity make it difficult to assign meaningful importance weights to individual multimodal features, limiting their use in trust-sensitive applications such as clinical affect monitoring and educational assessment. We address this gap by developing a framework based on tree-based ensembles that balances accuracy and interpretability. The framework encodes each modality into tokens, extracts and clusters concepts to reduce dimensionality, routes the fused modalities through tree-based ensemble classifiers, and interprets trends using a novel modified feature importance metric. The modified importance reduces the influence of the negative class in binary classification tasks, thereby improving indicator or marker detection. The proposed tree-based ensembles -- Linear Discriminant Tree (LDT), Linear Discriminant Forest (LDF), and Linear Discriminant AdaBoost (LDAB) -- achieve F1-mod gains of 4.3\% over the Multimodal Transformer and accuracy gains of 3.0\% over the primary interpretable multimodal baseline, Interpretable Multimodal Routing (IMR). The proposed multimodal feature importance extracts salient inter-modal concepts with substantially higher human-annotator agreement scores than default feature importance (62.2\% vs.\ 43.2\% on IEMOCAP; 46.7\% vs.\ 32.1\% on CMU-MOSI).

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：arXiv:2608.20378v1 Announce Type: new Abstract: Safety alignment in Large Language Models (LLMs) is often superficial, relying on refusal mechanisms that trigger only at the final stages of generation without erasing the foundational knowledge of harmful concepts acquired during pretraining. This study demonstrates that this architectural disconnect leaves models vulnerable to Semantic Camouflage -- adversarial attacks that wrap harmful intent in benign narrative contexts (e.g., creative writing), effectively bypassing standard input and output guardrails. By analyzing the latent activation trajectories of three distinct Small Language Model (SLM) families (Phi-3, Qwen2.5, and Gemma-2b) under adversarial stress, this research identifies a universal ``Intent Horizon'' -- a critical depth (typically 15--20\% of total layers) where the model's distinct, pre-trained representation of harmful intent collapses as it contextualizes the query into a ``safe'' narrative. Results indicate that while late-layer representations of camouflaged attacks are mathematically indistinguishable from safe queries (Detection Rate $< 20\%$), early-layer representations retain a distinct, detectable ``harm signature.'' Leveraging this insight, this paper proposes Latent Intent Verification (LIV), a lightweight probing defense. Experiments on the PKU-SafeRLHF dataset demonstrate that LIV outperforms standard guardrails by a margin of 20--50\% across all tested architectures, effectively neutralizing zero-day semantic attacks without requiring model retraining.

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：Models keep absorbing the harness into their weights — soon, it will be a harness for human attention rather than for the model.

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：What's up, ChickenButt? I made a free, native GTK client called ChickenButt. :) It lets you chat with local models through Ollama, and I figured some of you might get a kick out of it. Here's the repo: https://github.co…

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：Before you tune fusion, use the pre-tuning checks to verify index state and set a labeled baseline. Hybrid search retrieves dense and sparse candidate lists, then fuses them into one ranking. The dense prefetch finds similar meaning; the sparse prefetch finds matching keywords. Fusion reorders the candidates the prefetches return, so a document missing from both lists cannot appear in the result. Confirm Fusion Beats Either Prefetch Before tuning, compare dense retrieval, sparse retrieval, and default Reciprocal Rank Fusion (RRF) at k=2 and equal weights. Score all three with nDCG@10, which grades the top 10 results and gives more credit to relevant documents near the top.

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：Run Muse Glimmer locally on an RTX 3090 GPU using llama.cpp, DFlash speculative decoding, and Pi for fast, private, agentic AI coding.

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：Credit: Daniel Schwen The White House has published a new National Security Science and Technology strategy. It names 15 critical technology areas the United States intends to lead. Artificial intelligence and autonomy…

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：arXiv:2608.19207v1 Announce Type: new Abstract: Production deployments of Multimodal Large Language Models (MLLMs) increasingly rely on system messages to govern model behavior. Yet existing benchmarks either evaluate constraints in text only or embed them into the user turn, leaving system-message adherence in multimodal contexts largely unmeasured; they also leave open whether compliance comes at the cost of foundational vision-language capabilities. We introduce VSysBench, a benchmark built on MMVet-v2 that organizes constraints into 5 main categories and 22 sub-categories, ranging from textual directives in visual contexts to fully vision-grounded ones, each paired with a misaligned counterpart that stress-tests the instructional hierarchy. VSysBench scores each response jointly along two axes, constraint compliance and answer correctness, via the Joint Satisfaction Rate (JSR) and Cross-Constraint Sensitivity (CCS). Across 16 MLLMs, we find that imposing system messages substantially erodes base task accuracy, that compliance collapses under user conflict for open-weight models while remaining stable for top proprietary ones, and that vision-grounded constraints are the hardest category for every model.

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：arXiv:2608.19535v1 Announce Type: new Abstract: Retrieval-augmented generation (RAG) improves language-model responses by grounding generation in external passages, which comes with overhead: retrieved context lengthens the prompt, increasing prefill work, KV-cache footprint, memory traffic, latency, and energy. Context compression offers a natural remedy by pruning retrieved text before generation. However, state-of-the-art context-compression methods are typically used with a fixed compression budget, or with the rate selected offline and then applied at inference time. This static view ignores both workload variation and the live state of the edge device. On an edge SoC, compression is not free: the compressor itself runs on the same SoC and consumes latency and energy that can offset any generation savings. This paper proposes a vision for telemetry-informed adaptive compression in edge RAG, grounded in experimental evidence. We characterize the compression tradeoff on the NVIDIA Jetson AGX Thor using Llama and Qwen generators, Natural Questions and HotpotQA datasets, and LLMLingua-2 compression. Our measurements show that generation dominates the RAG budget for larger models, reaching roughly 90% of per-query latency and 91% of GPU energy for 7B-8B generators. Exploring the impact of the compression rate reveals an adaptive operating region: mild compression can miss energy opportunities, and overly aggressive compression can hurt inference quality. Intermediate compression can reduce GPU energy by up to 53.2%, and SoC energy by up to 48.2%, with negligible quality loss. We argue for runtime policies that dynamically manage compression, guided by workload features and edge telemetry.

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：arXiv:2608.19216v1 Announce Type: new Abstract: AI control research asks how to deploy models safely even when they may be misaligned, but many control protocols assume that the deployer can instrument the model and its surrounding pipeline. That assumption often fails for regulated organisations using frontier models through APIs or managed endpoints, where the deployer may control the business process but not the model weights, serving infrastructure, internal traces, update process, or full interaction logs. This paper introduces bounded sovereignty: partial technical and contractual access across the data, model, infrastructure, and interaction layers of the AI stack. It argues that these access conditions determine which control protocols can be executed in practice. The paper contributes a four-layer access typology, a protocol-by-layer requirements matrix, and the concept of sovereignty discount cost: the part of the control tax spent substituting for missing access through contracts, architecture, audit, vendor assurance, residual risk, or reduced system scope. It also reports a synthetic access-ablation experiment over 1.35 million synthetic case simulations and interprets the findings through an anonymised national-payments-infrastructure scenario. The experiment is not real-world payment-system evidence; it is a construct-validity exercise. The results show that complete logs improve diagnosis, a pre-execution gateway enables intervention, trace access and model-version control strengthen post-incident explanation, and scope restriction can improve safety while reducing usefulness. Control protocols proposed as general safety solutions should therefore state their access assumptions explicitly.

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：S1-mini is a 462 MB open-weights normalizer that sits after ASR, removing fillers and resolving self-corrections locally. The post Meet S1-mini: Superwhisper’s 462 MB Open-Weights Text Normalizer That Turns Raw ASR Transcripts Into Clean Written Text appeared first on MarkTechPost.

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：arXiv:2608.18093v1 Announce Type: new Abstract: Abliteration, the removal of refusal capabilities from large language models by projecting weight matrices orthogonal to an extracted refusal direction, has emerged as a prominent safety concern through its ability to bypass post-training alignment using only a small set of contrastive prompts. We find that existing defenses commonly overlook the cause of abliteration; that is, how easily the refusal direction can be extracted. To hinder this process, we introduce a weight-editing method that obscures the refusal signal by applying rank-$k$ updates to residual stream writer matrices while replacing refusal-inducing activations with random aliases and correcting downstream reader matrices to preserve the model's original behavior. On Llama-3-8B, AMRA improves post-abliteration refusal scores by $2.16$ points over the undefended baseline with less than $0.5$ percentage points of MMLU degradation. On Gemma-2-9B, it improves the post-abliteration refusal by $14.70$ points over the baseline while keeping harmful output rates similar to the baseline, albeit at a greater utility cost.

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：arXiv:2608.18090v1 Announce Type: new Abstract: Inside a modern language model sits a single internal direction that tracks how positive or negative a sentence feels. We show how to find this valence axis (V-axis) from just 9 emotion category names plus 50 short narrative paragraphs per emotion -- about 1,500 fewer labels than the usual supervised approach -- and that the same direction appears in vision, audio, and human-brain encoders never jointly trained. The recipe: embed nine emotion-anchored story sets in a frozen encoder, take the top principal direction of the nine averaged embeddings. Projecting new inputs onto it captures 93% of supervised performance on SST-2 (Llama-3-8B-Instruct, AUC 0.772 vs. 0.828), correlates with human valence ratings on 11,811 EmoSet images at r=0.636, reaches AUC 0.906 on ESC-50 audio (p12). A 2-parameter classifier trained on text labels transfers to images (AUC 0.961), audio (0.764), and brain recordings (0.828) without target-modality labels; a generic 16-D subspace stays at chance (0.525). The recipe is bounded to continuous attributes -- seven tests on categorical concepts return near-chance -- and steering is family-specific (Llama/Mistral yes, Qwen/Gemma no).

AI 服务暂时不可用，以下为来源摘要，待恢复后补全翻译：arXiv:2608.18089v1 Announce Type: new Abstract: Instruction-tuned models often refuse harmful requests in English but comply with the same requests in Yoruba, Igbo, Igala, and Hausa. This suggests that the refusal mechanism is present in the residual stream but fails to activate for low-resource inputs. Recovering it normally requires labelled target-language data and retraining, neither of which is available at scale for most African languages. We introduce Latent Space Refusal Anchoring (LSR-Anchoring), a training-free method that extracts the refusal direction from English prompts and clamps it onto the residual stream at inference time. The primary variant, Mean-Activation Steering (MAS), operates across the four architectures we tested: Llama-3-8B, Llama-3.1-70B, Mistral-7B-Instruct, and Qwen2.5-7B. On Mistral and Qwen it recovers safety with benign degradation below 0.08. On Llama-3-8B it overcorrects, with Degraded Performance on Legitimate prompts (DPL) reaching 1.00. We address this with SAE-Derived Steering (SDS), which replaces the dense mean-difference direction with a single Sparse Autoencoder (SAE) feature and reduces Kullback-Leibler (KL) divergence by 3.5-7x without benign collapse. Four languages transfer positively, but Arabic fails on every architecture and at every steering magnitude, indicating a geometric mismatch rather than a baseline effect. Massive Multitask Language Understanding (MMLU) accuracy drops remain below 0.35 percentage points at every effective steering magnitude.