overfeed.news
Anuncio

SRExpert

Deja de saltar entre 6 herramientas de Kubernetes.

Monitoreo, alertas, seguridad, despliegues Helm y operaciones con IA en una sola plataforma.

Probar 30 días gratis

Empresa

Alibaba

182documentos

7últimos 7 días

arXiv AI Papers

Who Asks, Who Answers: Linear Attention Builds the Query in Hybrid Language Models

Hybrid language models interleave linear-attention layers with a small number of softmax-attention layers. Prior work largely attributes in-context retrieval to softmax attention, treating linear attention primarily as a mechanism for compressing past context. We revisit this division of labor and show that linear attention plays a direct role in retrieval by constructing the query used by softmax attention. Studying Qwen3.5 (Gated DeltaNet) and Granite4.0 (Mamba-2), we identify an Encode--Address circuit. An early Encode block preserves token identity, while an Address block constructs the final-position representation from which a softmax reader forms its attention query. Transplanting this representation between prompts with different queried keys redirects retrieval in 0.97--0.99 of cases. The Address block's recurrent state carries the retrieval target, and its computation transfers across architectures: replacing Granite's Address block with Qwen's restores retrieval accuracy to 0.92--0.94 without updating either model's weights. Across six Qwen3.5 models, Address computation becomes distributed at larger scales over a chain of blocks, each guiding a successive reader, as the query no longer passes from one reader to the next through attention. These findings reveal a division of labor in hybrid retrieval: linear attention constructs the query, while softmax attention performs the lookup. Code is available at https://github.com/yuang-chen/LinearQuery

en
arXiv AI Papers

Quantizing Recurrent Language Models Is Easier Than You Think

Linear recurrences are becoming an increasingly common building block in large language models, replacing the growing KV cache of softmax attention with a fixed-size recurrent state that enables efficient decoding. However, the quantization stack has largely been developed around softmax Transformers, whose numerical properties and inference state differ from those of recurrent models. Our goal is to understand these differences and study the design space of quantization methods for recurrent models. In our analysis, we differentiate between two targets. First, extra-state quantization concerns weights and activations outside the recurrent state, whose distributions are shaped indirectly by the sequence mixer during training. In controlled 1.3B parameter models, we find that decayed recurrences develop weaker residual outlier channels than softmax attention and are therefore easier to quantize without incoherence processing. Second, intra-state quantization concerns the recurrent state itself during decode, where repeated updates create a different numerical problem from KV cache quantization. We instantiate our findings as LoReS, a state quantization pipeline combining value-axis scale grouping to match output channel magnitude, decay-based mixed-precision to protect slowly decaying heads or channels, and stochastic rounding to limit error accumulation under repeated requantization. Across Qwen-3.5 9B, Nemotron-H 8B, and Kimi-Linear 48B-A3B, LoReS supports robust 8-bit state quantization and lower effective precision on shorter decode horizons. We implement LoReS in vLLM and demonstrate end-to-end decode speedups.

en
arXiv AI Papers

Geometry of Confabulation and Refusal in Small Language Models

Faced with an unanswerable question, small language models can either refuse or confabulate: confidently produce a wrong answer. Can the hidden state before the first generated token provide a clue? Can we flip confabulations into refusals? I introduce ConfabQA, a 784-item benchmark whose obscure-pre-cutoff category breaks the typical correctness/knowledge-cutoff confound and map the geometry of the hidden state with linear probes, a sparse autoencoder (SAE), and direct optimization of intervention directions. On Qwen3-1.7B and Gemma 2 2B the probe separates upcoming refusals from confabulations well beyond prompt features (+7.4/+16.3 pp); on Llama 3.2 3B the ConfabQA margin is within fold-to-fold spread, but a direct refusal probe on PopQA and TriviaQA recovers +18-25 pp. On Qwen I closely inspect three candidate directions for the refusal vector: probe-recovered, Qwen-Scope SAE feature, and directly optimized. Although the three directions are pairwise 50-81 degrees apart, a push at or below the state's own norm along any of them flips most confabulations into refusal openings dominated by the "As..." family. Forcing the opener without touching the hidden state yields the same one-in-three judged-refusal rate as any push. The opposite push collapses refusals mostly into confabulations, yet leaves originally correct answers intact, so the direction gates the opening speech act rather than the knowledge behind it. Confabulations do not resolve into a concentrated feature in a dictionary or a token signature on Qwen3-1.7B. Scoring every probe on the margin over the strongest of four prompt-feature baselines extracts the hidden state's own contribution. Correctness signal is model-dependent and portable only for Llama 3.2 3B: 81% of the probe's margin over test majority survives transfer between ConfabQA, PopQA, and TriviaQA. Benchmark, judge, and code are released.

en
arXiv AI Papers

LoRAForge: Co-Designing Post-Hoc Compression and Execution for Efficient Multi-LoRA Serving

Multi-LoRA serving allows many task-specific adapters to share a common base model, but scaling to large adapter fleets introduces substantial storage, loading, and execution overheads. Post-hoc compression can reduce the memory footprint of trained LoRA adapters, yet smaller ranks do not automatically translate into faster serving. For instance, padding, factor access, and generic kernels can obscure the retained low-rank structure, while independently recompressing thousands of modules also incurs significant preparation overhead. We present LoRAForge, a system connecting batched factor construction to rank-specialized execution through an explicit retained-rank representation. It supports two allocation policies: uniform retention ratios (LoRAForge-U) and spectral-energy budget allocation (LoRAForge-S). Both share FlashTSQR factor construction and the same serving backend. FlashTSQR exploits parallelism across modules and row tiles, accelerating factor construction by up to 2.28over matched library QR. In a fixed-work Qwen control, the shared runtime adds 1.17--1.22throughput across both compressed policies, with negligible gains on original adapters. Across 85 trained open-weighted Qwen and Llama adapters, the spectral policy's floored quality configuration retains about 7.7% of adapter parameters with fleet-average score drops of at most 1.2 points on a 0--100 scale. The same rank-floored 1/12 spectral exports achieve substantial token-throughput gains on both AMD and NVIDIA GPUs.

en
arXiv AI Papers

Latent Sparse Attention: Towards Subquadratic Ultra-Long Context Attention

Agentic workflows are increasingly shifting towards long-horizon tasks, such as open-ended scientific work, where models benefit from long context windows. However, at ultra-long context (above 1M tokens), full attention is prohibitively expensive. A promising approach is sparse attention, in which a dedicated indexer selects a subset of keys for the attention. Yet current methods of indexing either still have quadratic cost in context size or lose accuracy at ultra-long context scale. We address this with Latent Sparse Attention (LSA), an efficient attention method that maintains accuracy at ultra-long context. In LSA, keys are hierarchically clustered in a learned latent space, and queries score the hierarchy of clusters instead of scoring every key, making prefill, decode and training faster at long context. We propose two variants, Distilled-LSA (D-LSA), which can be applied to pre-trained models via distillation, and Native-LSA (N-LSA), which is trained natively on the next-token loss alone. On GLM-5.3-FLASH (QWEN-3.8-27B) at 10M tokens, D-LSA speeds up one attention layer by 22.7x (72.4x) in prefill and 4.6x (42.5x) in decode over its original sparse attention (full attention), while keeping similar accuracy across a variety of tasks. N-LSA matches the accuracy of full attention with a training speedup of 1.6x per layer at 128k tokens. LSA offers practical and efficient attention for models to benefit from ultra-long context.

en
arXiv AI Papers

From Evaluation to Evolution: Benchmarking Image Editing and Evolving Image Models

Real-world image editing requires coordinating multiple visual transformations with content preservation and reference constraints, yet aggregate scores can readily obscure unmet local requirements. Effective evaluation must therefore identify local failures and translate them into priorities for subsequent model improvement. We introduce WeGenEditBench, comprising 2,740 evaluation instances and 148 task types across single-image, text-intensive, and multi-reference editing. Pre-generated checklists shared across models decompose each request into atomic requirements; item-level deductions and visual evidence are established first and then summarized into scores for edit execution, preservation, visual quality, and task-specific correctness. To support the experience-acquisition component of recursive self-improvement (RSI), we develop WeGenEvoStudio, a unified multi-agent harness for RSI-oriented learning that integrates model-conditioned failure diagnosis, retrieval-grounded data construction, supervised adaptation, and re-evaluation. Among 15 evaluated configurations, GPT-Image-2 leads single-image editing with an Overall score of 4.076, while Seedream-5.0-Pro leads text-intensive editing at 4.291. Evolution experiments across four generation and editing benchmarks demonstrate substantial targeted gains: on their report-selected Top-5 single-image targets, FLUX.2 [dev] improves Overall from 1.083 to 3.517, while Qwen-Image-Edit-2511 gains 13.2%. Together, our benchmark and harness provide a foundation for developing more reliable image models through interpretable evaluation and feedback-driven learning.

en
arXiv AI Papers

sDMD: Sinkhorn Distribution Matching Distillation

Distribution Matching Distillation (DMD) distills multi-step diffusion models into few-step generators by minimizing reverse KL divergence. Despite its strong performance, DMD guides samples by local density alone, which introduces two limitations: teacher scores can be unreliable in regions of low teacher density, and the mode-seeking behavior of reverse KL can collapse the student onto a subset of the teacher modes. We first connect DMD to Wasserstein gradient flow, showing that its gradient follows from a fixed-point formulation with integral KL (IKL) as the energy functional. This perspective motivates augmenting IKL with an optimal transport term that addresses these limitations by guiding outlier samples toward the target and promoting mode coverage. We realize this term using Sinkhorn divergence estimated from finite batches. In practice, however, errors in the estimated Sinkhorn velocity can perturb in-domain samples for which IKL guidance is already accurate. To this end, we introduce sDMD, which adaptively activates the Sinkhorn correction based on discriminator outputs, focusing on outlier samples where IKL may be unreliable. With 2--4 sampling steps on FLUX.1-dev, Qwen-Image-20B, and Wan2.1-14B, sDMD outperforms distillation baselines on GenEval, DPG, WISE, and VBench while improving sample diversity over DMD2.

en
arXiv AI Papers

SatOps-Decision: Low-Latency Discriminative Decisions for Onboard Satellite Communication Operations

We propose SatOps-Decision, a method for low-latency decisions in satellite communication operations. To the best of our knowledge, it is the first application of Jev-style general-purpose discriminative decision models to satellite operations, and the accompanying benchmark is the first dedicated to this approach. The learned scorer, a locally adapted Laya encoder with its native decision head, selects among supplied operational answers in a single forward pass. The benchmark combines authored operational contracts, controlled state changes and independently checked labels in an auditable construction. Evaluation prioritizes complete-request latency and raw decision accuracy; all main accuracy comparisons use the same 1,024 held-out questions, separate from training and development. SatOps-Decision's local scorer has a median latency of 90.3 ms per decision, versus 493.9 ms for Qwen3.5-0.8B generation and 388.4 ms for OneJev1.6-0.8B: 5.47and 4.30speedups based on median request times in the measured FP32/MPS implementations. Its mean raw accuracy is 85.3%, versus 42.5% for frozen Laya, 40.7% for Qwen generation and 58.6% for OneJev. As cloud accuracy references, original Jev 1.14/1.13 reach 87.4/81.9%.

en
arXiv AI Papers

"Close enough?" An investigation into multiplication errors in LLMs

Large language models (LLMs) frequently fail at multi-digit arithmetic, but the underlying mechanisms governing these failures remain poorly understood. While recent work demonstrates that addition errors follow a "noisy quantisation" model localised around carry boundaries, whether this framework generalises to multiplication remains an open question. In this work, we evaluate multiplication capabilities across several small base LLMs on a dataset of 16,991 problems spanning varied operand lengths. We show that multiplication errors do not conform to the addition-style noisy quantisation model. Instead, we uncover a novel error regime in small Qwen models: errors in lower-order columns predominantly preserve parity. Conversely, higher-order digit errors shift toward off-by-one carry failures. Our findings suggest that arithmetic computation in LLMs relies on distinct, multi-periodic internal representations where parity is tracked more robustly than exact digit identity, providing new mechanistic insights into numerical reasoning in transformers.

en
arXiv AI Papers

Follow the Gradient: Tracing Safety Degradation Back to Training Data

Fine-tuning Large Language Models (LLMs) on instruction data can quietly erode their safety, even when no training example looks harmful. Because this safety-degrading data looks benign, standard content filters struggle to catch it. We tackle this problem by developing SimpleLoRIF, a scalable gradient embedding method that identifies fine-tuning examples whose gradients align with unsafe behavior. By simplifying a low-rank attribution method and restricting backpropagation to the final four transformer blocks, SimpleLoRIF can embed datasets on the order of a million examples, boosting gradient attribution from a small-scale analysis tool into a practical filter for full post-training datasets. We first validate it on emergent misalignment, where we find that gradients can connect insecure-code training data to the broadly harmful behaviors that appear after fine-tuning, with AUC as high as 0.865 in the gradient space of misaligned models. We then use SimpleLoRIF to build SafeTulu, a version of Tulu 3 with the 20% of examples most gradient-aligned with unsafe responses removed. Training on SafeTulu reduces GCG attack success rate by up to 22% across Qwen, Llama, and Olmo, outperforming both LLM-as-a-judge and prior gradient-based filtering, while largely preserving or even improving capability. Finally, we analyzed the removed data and identify possibly under-studied sources of safety degradation, such as tool-execution failures and narrative arithmetic. We release SafeTulu along with a Hugging Face Space that lets researchers build their own customized versions of SafeTulu to fine-tune on, such as by changing the removal fraction or filtering only certain subsets.

en
arXiv AI Papers

Can JEV Support Clinical Trial Criteria Assessment?

Clinical trial prescreening requires distinguishing unmet conditions from insufficient evidence in patient records. Given patient records, criterion questions, and predefined answer options, we retrospectively compared JEV, a non-generative decision model returning native option probabilities, with schema-constrained Qwen Flash across 2,048 test patient-condition pairs from TrialGPT and question-answer datasets derived from MIMIC-III and MIMIC-IV, all evaluated under strict Zero Data Retention (ZDR) policies to ensure clinical data privacy. Both models assessed conditions individually or in batches sharing the same patient context, with performance measured against reference labels. In batch mode, JEV and Qwen achieved similar macro-F1 scores on TrialGPT (0.652 vs. 0.660) and MIMIC-IV (0.807 vs. 0.814), whereas JEV substantially outperformed Qwen on MIMIC-III (0.768 vs. 0.534; Δ = +0.234, 95% CI: 0.187-0.281). This advantage largely reflected higher recall for conditions labeled as having insufficient information (56.3% vs. 9.9%). Across all three datasets, JEV also had lower observed median job latency (1.18-1.40 vs. 2.80-8.90 s) and 53%-74% lower estimated cost per condition. These findings position JEV as an initial assessment component that provides criterion-level judgments and flags potential information gaps for review. JEV's observed errors in interpreting clinical evidence underscore the need to verify its outputs against patient records before they inform eligibility decisions.

en
arXiv AI Papers

Where Rounding Happens Matters: Rounding-Location Alignment Stabilizes LLM Reinforcement Learning

Reinforcement learning for large language models often uses separate systems for inference and training. Even with identical weights and BF16 tensors, implementation differences can make an intended on-policy update off-policy and destabilize training. Existing remedies align the entire inference and training stacks, change global precision, or modify the learning algorithm. We ask whether aligning selected parts of the execution stack can stabilize BF16 RL while other implementation differences remain. We align training-side rounding locations with inference in three regions---RoPE, SwiGLU, and residual RMSNorm---by retaining selected intermediates in FP32 until their BF16 outputs. Across four Qwen-family models, aligning all three regions prevents the collapse observed under the default BF16 baseline; RoPE-path alignment alone also suffices in the studied runs, despite remaining numerical discrepancies. The interventions keep the inference engine, matrix-multiplication and attention kernels, and learning objective unchanged.

en
arXiv AI Papers

Do "True" and "False" Add Up to One? Complement Consistency in LLM-Agent Opinion Simulations

In LLM-agent opinion simulations, each agent reports the probability that a claim is true after seeing its own previous probability and the probabilities reported by its neighbors, and that number becomes its state in the next round. If the number describes the agent's position rather than the sentence it was shown, then rewriting the claim as its complement and replacing every probability p with 1 - p should return the complementary answer, so the two reports should sum to one. We test this with three small open models: Qwen3-4B, Llama-3.2-3B and Gemma-3-4B. On 2,304 agent states taken from a running network simulation, Qwen's paired reports sum to 1.18 on average, and on 48 real fact-checking claims from AVeriTeC they again exceed one. A controlled benchmark of 24 worlds, each with exactly two possible outcomes, shows the same failure in all three models. The discrepancy is two to ten times larger than the change produced by a same-meaning paraphrase, and it persists under repeated sampling and across peer group sizes and report strengths. When the reports are fed back into two populations that start from complementary states, the populations do not remain mirror images: their divergence is positive after five rounds for every model. Averaging each answer with the complement of its paired answer restores consistency by construction and lowers Brier loss averaged over the two wordings, but at the final round it beats the better of the two original wordings for only one of the three models. A schedule-by-targeting experiment on the same model gives a small, stable interaction because such sweeps keep the claim wording fixed. Consistency under rewording therefore needs its own test before simulated probabilities are read as beliefs

en
arXiv AI Papers

PHYRD: A Living Benchmark for Reproducing and Extending Theoretical Physics Research

Large language models and agents have shown significant power in advanced physical research. In this paper, We introduce PHYRD, a living benchmark for theoretical-physics reasoning with 63 Reproduce and 43 Discover problems. Reproduce tests derivation of published results; Discover tests consequences of modified theoretical conditions. FormulaForest constructs the tasks with reference solutions and subquestion-level rubrics. Across five models, whole-problem accuracy ranges from 49.21% to 68.25% on Reproduce and from 18.60% to 74.42% on closed-book Discover. With tools, four models complete more Discover problems, but conceptual errors and missing arguments persist. Besides, our experimental results demonstrate the value of FormulaForest as synthetic data. On FrontierScience, a benchmark developed by OpenAI, it raises the accuracy of a Qwen-35B-A3B model from 61% to 72%.

en
arXiv AI Papers

DuoCook: An Overcooked-Inspired Benchmark for Visual Planning and Coordination

Coordinating two players in a cooking task requires a vision-language agent to track changing scene states, sequence recipe steps, and allocate shared resources under a time budget. We introduce an Overcooked-inspired benchmark for evaluating these capabilities in Unreal Engine environments constructed using coding agents. A single vision-language model controls both players from top-down visual observations, selecting high-level actions while the engine executes navigation. The benchmark contains 20 maps in five categories. These categories support evaluation of centralized coordination under equipment contention, changing terrain, and handoff dependencies. A prop legend and pot ingredient-count indicators provide explicit visual cues, while grid-based destinations support spatial grounding. The evaluation compares four VLMs on all 20 maps, measuring task score, order completion, delivery precision, action acceptance rate, and model API expenditure. A memory ablation on the same 20 maps examines the effect of a shared, self-authored notebook while retaining six-response action history in both conditions. The strongest evaluated model, gpt-6-astra, achieves 26.2\% mean order completion. Notebook memory has mixed score effects across models, with positive mean differences for Claude and Qwen and negative differences for GPT and Gemini.

en
NVIDIA AI Blog

NVIDIA Cosmos Opens New Tools to Train, Evaluate and Deploy World Models

To help developers rapidly build world models, which help physical AI systems understand and simulate the real world, NVIDIA has expanded the Cosmos platform with new training tools, simulation checkpoints and Cosmos Data Factory blueprints. Ranking top five on six different open models leaderboards, the NVIDIA Cosmos platform gives world model builders the frontier open […]

en
Alibaba — overfeed.news