Search: “model pruning”

Smart search ranks by meaning as well as keywords (one row per story, last 45 days).

Understanding the Impact of Model Pruning on Long-Tail Forgetting and Explanation Reliability in Medical Imaging

Systematic study finds model pruning causes frequency-dependent long-tail forgetting in medical imaging and that gradient-informed methods best preserve explanations.

Across two long-tailed medical imaging datasets, two CNN architectures, four pruning methods, and sparsity up to 95%, the study measures predictive performance, explanation stability, and faithfulness. Rare classes degrade earlier and more severely than frequent ones, while explanation reliability depends mainly on the pruning strategy, with gradient-informed methods degrading least. Mechanistic analysis ties explanation collapse to loss of class-discriminative gradients rather than vanishing feature activations, recommending class- and explanation-aware evaluation of compression.

arXiv cs.AI / cs.LG / cs.CL·Sep 7, 17:44 UTC · 9d agoAI research

27

What Breaks Under Pruning in Smart Homes, and When? Evaluating LLM Degradation Across Architectures and Task Complexity

Pruning study across four LLM architectures finds dense models degrade sharply on smart-home tool calling while MoE models tolerate far more.

Researchers systematically study pruning-induced degradation in smart-home tool calling across four LLMs spanning dense Transformer, dense hybrid, and mixture-of-experts architectures, combining depth, width, hybrid, and expert pruning methods, and evaluate over 19,500 instances from three datasets after post-pruning supervised fine-tuning. Dense models show narrow safe pruning regions followed by sharp degradation, while MoE models tolerate substantially more pruning. Pruning degrades grounded specificity (operation, device, argument, value) before schema-level intent, and aggressive dense pruning can induce systematic over-refusal.

arXiv cs.AI / cs.LG / cs.CL·Sep 15, 17:51 UTC · 1d agoAI research

20

Lightweight Vision Transformer Compression for On-Device Plant Disease Detection in Resource-Constrained Agricultural Field Conditions

A unified ViT compression pipeline (H-BAC pruning, quantization, distillation) cuts plant-disease models 54.5x to 6.01 MB while keeping 95.13% accuracy.

Researchers combined Hessian-Balanced Adaptive Block Pruning (H-BAC), guided by second-order sensitivity estimation, with quantization and attention-based knowledge distillation to compress Vision Transformers for on-device chilli plant disease detection in India. On a 3-class cross-village, cross-device out-of-distribution dataset, the integrated pipeline reduced model size from 327.42 MB to 6.01 MB (54.5x) at 95.13 +/- 2.32% accuracy, matching the 95.13% FP32 baseline. Ablations also show a directly trained 6.01 MB INT8 student reaches 94.87% accuracy, indicating where pruning and distillation add limited value.

arXiv cs.AI / cs.LG / cs.CL·Sep 4, 16:40 UTC · 12d agoAI research

22

Don't Send What You Don't Need: Question-Guided Token Pruning as a Privacy Defense for Vision-Language Models

QPriv-VL prunes privacy-sensitive visual tokens in federated/split VQA, cutting membership-inference success on VQA-RAD from 0.99 to 0.76-0.79 using ~40% of tokens.

The paper proposes QPriv-VL, a question-guided token-pruning framework for federated, split, and U-shaped split learning that suppresses privacy-sensitive visual patches before transmission. Its Dynamic Threshold Predictor combines cross-modal question relevance with frozen DINOv2-derived sensitivity to compute a per-sample pruning ratio and retention mask in one forward pass, without sensitivity labels. Evaluated on GQA, OK-VQA, VQAv2, SLAKE, VQA-RAD, and PathVQA against FSHA, FORA, iDLG, and attribute-inference membership inference attacks, it matches or beats fixed-ratio pruning. On VQA-RAD it reduces membership-inference success from 0.99 to 0.76-0.79 while preserving competitive accuracy with about 40% of the original token budget.

arXiv cs.CR·Sep 14, 14:48 UTC · 2d agoResearch

35

X-AuT: Progressive Audio-Encoder Compression for Speech LLMs with Cross-Scale Distillation

XPeng AI's X-AuT prunes speech LLM audio encoders, cutting Qwen3-ASR-0.6B error from 5.61% to 5.27% with fewer parameters.

X-AuT is a progressive compression framework for speech LLM audio encoders that selects layer combinations via short behavioral probes and restores pruned models using cross-scale distillation and LoRA finetuning while keeping the language-model backbone frozen. Compressing Qwen3-ASR-0.6B from 18 to 16 audio-encoder layers lowered macro-average error from 5.61% to 5.27% on ten Chinese-English benchmarks. A 14-layer model reached 5.75% error with 20.7% fewer audio-tower parameters, and progressive pruning outperformed direct pruning (5.75% vs 6.73%).

Hugging Face daily papers·Sep 9, 20:00 UTC · 7d agoAI research

26

CoVeR: Coverage-Based Token Pruning for Multi-View 3D Reasoning in VLMs

CoVeR, a training-free coverage-based token pruner, preserves 93.5% of VLM 3D-reasoning performance using only about 8% of visual tokens.

Researchers introduce CoVeR, a deterministic, training-free selector that chooses visual tokens to cover every region of a multi-view 3D scene using only token coordinates. Unlike learned-importance and voxelization pruners, it enforces an exact per-scene token budget, avoids saturation plateaus, and prevents near-duplicate selections. Experiments across four vision-language models show it surpasses prior state of the art by 3.9 percentage points on average across three 3D reasoning benchmarks.

Hugging Face daily papers·Sep 7, 20:00 UTC · 9d agoAI research

45

WarmBloodAban/Minimax-h3_Singularity — new model trending #22 on Hugging Face

Community fine-tune Minimax-h3_Singularity enhances MiniMax-H3 video generation with HDR quality, distant face restoration, and improved motion, trending #22 on Hugging Face.

Minimax-h3_Singularity is a community fusion fine-tune of the MiniMax-H3 multimodal video generation model, built from multiple checkpoints and refined with pruning and weight optimization. It supports Text-to-Video, Image-to-Video, Reference-to-Video, and Video-to-Video workflows in ComfyUI, and claims improvements in HDR clarity, distant face restoration, motion fluidity, and fantasy VFX. The authors recommend pairing it with the minimax_h3_ref2v_turbo_4step_v0.1 LoRA for four-step accelerated inference, and an online demo is available via RunningHub.

Hugging Face trending models·Sep 5, 18:51 UTC · 11d agoModel release7· 1 read

28

RelateAnything: Real-Time Open-Vocabulary Relation Prediction From Any Inputs

RelateAnything is a 53M-parameter open-vocabulary relation prediction model running at 20 ms/frame, with 2.3-3.5x higher mean recall than comparable open-vocabulary methods.

RelateAnything predicts scored relations between image regions using any predicate vocabulary supplied at inference as text embeddings, with object labels never required as input, so region sources can change without retraining. Training covers 19,103 predicates using positive-unlabeled supervision; the authors release RA-4M (474k images, 4.3M geometrically verified relations over 10,102 free-text predicates) and the OV-SGG-Bench evaluation suite. The 53M-parameter model runs at 20 ms/frame and achieves 2.3-3.5x the mean recall of the strongest comparable open-vocabulary method across cross-dataset and zero-shot benchmarks. Model, corpus, and benchmark are public.

Hugging Face daily papers·Sep 10, 20:00 UTC · 6d agoAI research1

40

Learning 3D Editing without Paired Supervision via Generative Prior Distillation

New framework distills 2D editing and VLM priors into a feed-forward 3D editing model without paired 3D training data.

The method, PriorEdit3D, learns feed-forward instruction-guided 3D editing by distilling knowledge from foundation models instead of using ground-truth 3D pairs. Through a differentiable rendering pipeline it supervises a 2D visual prior from an image editing model at the main view and a Vision-Language Model semantic prior at novel views for instruction fidelity and identity preservation. A 3D-aware Distribution Matching regularization constrains outputs to the manifold of realistic 3D assets defined by a pretrained image-to-3D teacher. Experiments report superior instruction fidelity and cross-view consistency over state-of-the-art baselines, with code released on GitHub.

Hugging Face daily papers·Sep 3, 20:00 UTC · 13d agoAI research

30

SQS: Bayesian DNN Compression through Sparse Quantized Sub-distributions

SQS unifies weight pruning and low-bit quantization via Bayesian variational learning, compressing Llama3.2 and Qwen2.5 at higher rates with comparable accuracy.

SQS introduces a unified Bayesian variational framework performing simultaneous pruning and low-bit quantization, using a spike-and-slab prior for sparsity and Gaussian Mixture Models to model quantized weights. The authors derive an efficient approximation for the intractable objective and provide a consistency result for the variational approach. Experiments on ResNet, BERT-base, Llama3.2, and Qwen2.5 show higher compression rates than prior baselines with comparable performance drops.

Hugging Face daily papers·Sep 6, 20:00 UTC · 10d agoAI research

28

Training and Finetuning Multi-Vector Embedding Models with Sentence Transformers

Hugging Face published a tutorial on training and finetuning multi-vector embedding models using the Sentence Transformers library.

Hugging Face's blog walks through training and finetuning multi-vector embedding models with Sentence Transformers. Multi-vector approaches store multiple vectors per document to support late-interaction retrieval. The post is a practical guide for developers building retrieval pipelines with the library.

Hugging Face Blog·Aug 26, 00:00 UTC · 21d agoAI tools & infra1

20

ShallowStream: Index Shallow then Answer Deep for Streaming Video Understanding

ShallowStream builds streaming-video retrieval indexes from shallow MLLM layers, cutting per-frame prefill latency by up to 52.1x.

ShallowStream is a framework for streaming video understanding with multimodal LLMs that uses the model's shallow layers to simultaneously encode frames and maintain an always-on lightweight retrieval index via shallow-layer KV caches, avoiding full-depth prefill for every incoming frame. At query time, shallow-layer attention scores plus a diversity-aware selection strategy retrieve relevant context frames. It reports performance on par with the strongest existing streaming methods while reducing per-frame prefill latency by up to 52.1x and 10-second end-to-end latency by up to 11.9x, with code released on GitHub.

Hugging Face daily papers·Sep 1, 20:00 UTC · 15d agoAI research

24

Multi-Task Learning for Sparsely-Labeled Time Series: A Case Study on Cold-Hardiness Modeling

Multi-task RNN architectures pooling sparse cultivar data improve grape cold-hardiness and budbreak prediction over single-task and scientific baselines.

Researchers apply recurrent neural networks to daily grape cold-hardiness prediction from weather time series, where per-cultivar labels are temporally sparse and limited. They design multiple multi-task learning architectures that treat cultivars as tasks and evaluate them in both MTL and transfer learning settings. Certain architectures consistently outperform single-task learning and state-of-the-art scientific models, and a single MTL model jointly learning cold hardiness and budbreak improves accuracy on both tasks.

arXiv cs.AI / cs.LG / cs.CL·Sep 8, 17:17 UTC · 8d agoAI research

12

TRACE: Trajectory-robust Admission with Evidence Ordering for Efficient GUI Agents

TRACE, a training-free visual token pruning framework, cuts GUI agent inference latency and memory while keeping trajectory-wide visual evidence reusable.

TRACE is a training-free framework for trajectory-robust admission and coverage-aware evidence ordering that prunes high-resolution screenshot tokens accumulated in GUI agent trajectories. It ranks visual evidence using a query-independent layout-derived interaction prior combined with instruction relevance and feature novelty, and reserves part of the budget for native tokens distributed across the screen to repair spatial coverage. A monotone KV contraction incrementally compresses retired frames into compact session state, avoiding repeated visual encoding or pruning. Experiments across six GUI benchmarks and diverse models verify effectiveness under tight budgets, with source code to be released.

Hugging Face daily papers·Sep 8, 20:00 UTC · 8d agoAI research

28

Kaininja: Extending Native 3D Generators to the Part Level

KaiNinja extends TRELLIS.2 native 3D generation to part-level assets via a dual-volume O-Voxel representation, cutting whole-object Chamfer distance by 40%.

KaiNinja extends the TRELLIS.2 native 3D generator to produce part-level assets instead of one fused mesh, enabling downstream editing, rigging, and simulation. A dual-volume form of the O-Voxel representation solves the problem that a single volume cannot represent interfaces where two parts touch. The model needs no segmentation network, is partly trained on LLM-agent-authored part data, lowers whole-object Chamfer distance by 40%, and raises strict part F-score by 16% versus other part-generation pipelines.

Hugging Face daily papers·Sep 13, 20:00 UTC · 3d agoAI research

18

A*-Thought-V2: Efficient Latent Reasoning via Geometric Dynamics of LLM

A*-Thought-V2 compresses chain-of-thought into latent tokens using geometric hidden-state dynamics, cutting computation while improving accuracy on Qwen models.

A*-Thought-V2 models chain-of-thought as a hidden-state trajectory and interleaves explicit text with continuous latent tokens, compressing steps whose transitions deviate from the question-to-solution direction. Trained via stepwise embedding forcing and label forcing with soft multi-modal supervision, it was evaluated on Qwen3.5-9B and Qwen3.6-27B across six benchmarks. Reported results include up to 2.6% average accuracy gain, up to 50% shorter responses, 2.29x higher Accuracy per Computation Unit, 94.6% faster preprocessing, and up to 80.3% faster training.

Hugging Face daily papers·Sep 7, 13:56 UTC · 9d agoAI research

48

Likelihood-free inference with nuisance parameters through normalizing flows

Researchers decompose normalizing flows to derive near-pivotal statistics for likelihood-free inference with nuisance parameters, recovering the t-test and beating Welch limits.

A new paper decomposes neural-network normalizing flows to uncover pivotal statistics in the presence of nuisance parameters using only a sample generator from the distribution of interest. The statistic is near-pivotal in the sense of minimum average KL-divergence of its p-values and can incorporate prior knowledge of group invariances such as translation and scale. Experiments show it recovers the one-sample t-test almost exactly, outperforms the Welch test on worst-case size over a constrained variance-ratio range, and delivers higher power and much faster runtime than profile likelihood-ratio techniques on small-to-moderate samples.

arXiv cs.AI / cs.LG / cs.CL·Sep 9, 17:58 UTC · 7d agoAI research

16

A*-Thought-V2: Efficient Latent Reasoning via Geometric Dynamics of LLM

A*-Thought-V2 compresses redundant chain-of-thought steps into latent tokens guided by hidden-state geometry, improving accuracy up to 2.6% while halving response length.

A*-Thought-V2 models chain-of-thought as a hidden-state trajectory projected into a 3D PCA space and compresses steps whose transitions deviate from the question-to-solution direction into continuous latent tokens, keeping aligned steps explicit. Training uses stepwise embedding forcing and label forcing with soft multi-modal vocabulary supervision. On Qwen3.5-9B and Qwen3.6-27B across six in-domain and out-of-domain benchmarks it improves average accuracy by up to 2.6%, cuts response length by up to half, and raises Accuracy per Computation Unit 2.29x while reducing preprocessing and training time by 94.6% and up to 80.3%.

arXiv cs.AI / cs.LG / cs.CL·Sep 7, 17:56 UTC · 9d agoAI research

36

Data Scarcity and Model Sparsity: Mixtures-of-Experts Overfit More to Repeated Data

Study finds Mixture-of-Experts models overfit faster than dense Transformers under repeated training data, with degradation tied to total parameter sparsity.

Across models from 80M to 1B active parameters (8.5B total), MoE architectures degrade more rapidly than dense models when training data is repeated, with the effect increasing with sparsity as dictated by total parameters. Dense 80M models tolerate 8x repetition with minimal loss while MoEs suffer at 4x and underperform dense models beyond 32x. Masking-based regularization such as dropout mitigates overfitting, letting MoEs beat dense models even at over 64x repetition, though no method matches all-unique training data. Routing stabilizes early and expert specialization correlates with overfitting to repeated data.

arXiv cs.AI / cs.LG / cs.CL·Sep 10, 17:57 UTC · 6d agoAI research

36

ZGCM-1: A Fully Open and Extremely Efficient Foundation Model for Math and Agentic Search

ZGCM-1 is a fully open 7B foundation model with 256K context that stays competitive with frontier models on math reasoning and agentic search.

ZGCM-1 is a fully open 7B dense foundation model trained from scratch using an efficiency-focused recipe: interleaved gated sliding-window and full attention, a stable FP8 Muon optimizer, and MDP-based mid-training with context scaling across 16K, 64K, and 256K. On mathematical reasoning and agentic search suites it remains competitive with much larger frontier models such as Qwen3-235B-A22B and GLM-5.1. The recipe yields a ~4.2x improvement in 16K pre-training time-to-loss, and all weights, checkpoints, training code, data recipes, and W&B logs are open-sourced.

Hugging Face daily papers·Sep 10, 20:00 UTC · 6d agoModel release

50

Graph Machine: Towards Better Pretraining via Edges

Researchers propose Graph Machine, an O(n)-state sparse architecture that replaces 75% of Qwen3-0.6B dense layers with only slight loss change.

The paper introduces the Graph Machine (GM), an architecture that maintains an O(n)-sized state accessed through sparse, dynamic routing via pointer-like edges updated differentiably by a referral mechanism resembling pointer chasing. The authors replaced 75% of dense Transformer layers in Qwen3-0.6B with GM sparse layers and pretrained from scratch on 15.7B tokens. Retrieving 2 of 4,096 tokens per KV head in each sparse layer degrades loss only slightly, while retrieving 4 marginally improves loss over the dense baseline.

Hugging Face daily papers·Sep 1, 20:00 UTC · 15d agoAI research

30

Ambient @ EgoLongQA 2026: Distilling Long-Video perception into a Sub-2B Model

Ambient team wins EgoLongQA 2026 sub-2B division by distilling an agentic long-video perception pipeline into a 2B vision-language model.

Ambient's entry to the EgoLongQA track of the Wearable-AI Challenge at ECCV 2026 placed first in the <=2B parameter division with 0.8279 on the held-out test set. The system distills the junior perception module of a tool-using agentic pipeline into a 2B student, reaching 89% of the pipeline's accuracy with 1.1% of its parameters and lifting a 27.1% base model to 81.4%. To meet the division limit, the multilingual embedding table is pruned from 248,320 to 143,469 rows, reaching 1.9985B parameters with provably identical logits on retained rows.

Hugging Face daily papers·Sep 9, 20:00 UTC · 7d agoAI research

35

MAxBench: A Multinomial Concept Recovery Benchmark

MAxBench evaluates multinomial concept recovery methods, finding affine subspaces steer most reliably but none consistently beats prompting.

MAxBench is a geometry-agnostic evaluation framework for multinomial concept representations in language models, based on sampling from recovered concept representations. It compares 10 localization methods covering 5 geometry types across 6 concepts and 4 models. Findings show affine subspaces steer more reliably than rank-one or linear subspaces due to better non-zero offsets, manifold steering is competitive where applicable, and no method consistently outperforms prompting.

arXiv cs.AI / cs.LG / cs.CL·Sep 11, 17:08 UTC · 5d agoAI research

18

[AINews] 10% worse, 100x cheaper, 10000x faster: Why Simulation is taking over

Latent Space argues AI training pipeline stages—rewards, data, teachers, curricula, environments—are flipping from human-made to model-made simulation.

Latent Space's AINews essay traces how each component of AI training has turned synthetic since 2022: reward models (InstructGPT, RLAIF), synthetic pretraining data (Microsoft Phi, NVIDIA Nemotron-4 340B), model teachers (Alpaca, DeepSeek-R1 distillation), and self-generated curricula (Self-Rewarding Language Models, SPIN). In 2026 it highlights Karpathy's autoresearch loop—700 experiments yielding 20 kept improvements, cutting GPT-2 training time from 2.02 to 1.80 hours—and Z.ai's GLM-5.3 fully synthetic RL environment, judging, and verification stack. It frames these shifts as 'simulation': 10% worse but 100x cheaper and 10,000x faster than human equivalents.

Latent Space·Aug 22, 07:36 UTC · 25d agoAI industry

35

OpenWAM: An Open, Modular Exploration Towards Systematic World-Action Model Pretraining

OpenWAM releases an open modular stack for world-action model pretraining, plus OpenWAM-alpha trained on about 6,400 hours of egocentric and robot data.

OpenWAM is an open research stack that factorizes World-Action Model pretraining into composable infrastructure, study, and model components with unified training, inference, and evaluation. Controlled experiments distill three principles on knowledge inheritance, world-action synergy, and out-of-domain generalization gains from embodied co-training. The resulting OpenWAM-alpha, pretrained on roughly 6,400 hours of egocentric human and robot data, achieves top-tier results across eight simulation benchmarks and real-robot tests spanning single-arm, bimanual, and dexterous embodiments. The full stack, including infrastructure, evaluation protocols, pretrained models, and data recipes, is released openly.

Hugging Face daily papers·Sep 6, 20:00 UTC · 10d agoAI research

44

Man Charged With 3 Felonies For Breaking 3D

Oviedo, Florida police charged a man with three felonies for cutting down an officer's 3D-printed decoy Flock surveillance camera.

After several real Flock Safety cameras were stolen in Oviedo between July 23 and August 3, 2026, police replaced them with 3D-printed decoys built by an officer at home and monitored the fakes. Evan Meyer was arrested after midnight and charged with attempted grand theft, criminal mischief over $1,000, and property crimes against computer equipment, despite the decoy costing only a few dollars of filament. Mayor Megan Sladek said she had no idea the sting was underway, and the department claims no records of the decoy's creation exist, citing an ongoing investigation.

404 Media·Aug 25, 15:44 UTC · 22d agoPolicy & legal

22

Studying Image Tokenizers as Visual Languages in Unified Multimodal Models

A controlled autoregressive testbed shows validation losses must be analyzed per task, and image tokenizer choice affects joint multimodal text modeling.

Researchers built a pure-autoregressive testbed to study image tokenizers as the 'visual language' of unified multimodal models, tracking task-specific validation losses during multimodal continual pretraining across text, image, text-to-image (T2I), and image-to-text (I2T) prediction. They found that losses exhibit distinct scaling behavior per task and rank tokenizers differently, and that I2T loss over a shared text vocabulary gives a more consistent loss–performance signal than T2I loss. Better reconstruction does not necessarily yield lower task-specific losses or stronger downstream performance, and tokenizer choice can affect text modeling under joint optimization. Case studies examine the discriminator, semantic supervision, and vocabulary size design axes.

Hugging Face daily papers·Sep 7, 20:00 UTC · 9d agoAI research1

40

GE-Act 2.0: Pretraining and Scaling a World-Action Model for Robotic Manipulation

GE-Act 2.0 is a from-scratch pretrained world-action model for robotic manipulation, with success rising from 17.1% to 44.1% as co-training data scales to 30,000 hours.

Genie Envisioner Act 2.0 (GE-Act 2.0) is a world-action model whose generative and action components are all initialized from scratch on manipulation data, combining a control-oriented autoencoder (CoAE), single-step visual planner (SVP), and inverse dynamics model (IDM) trained jointly via knowledge-aligned selective optimization (KASO). Scaling co-training data from 300 to 30,000 hours raises zero-shot success from 17.1% to 44.1% on G1-OP and 13.4% to 31.1% on G2-90D, despite the latter comprising under 2% of data, suggesting cross-embodiment transfer. Gains span 19/20 and 18/20 skill groups, and skill-specific coverage correlates with zero-shot OOD success (Pearson r=0.80).

Hugging Face daily papers·Sep 3, 20:00 UTC · 13d agoAI research

30

Forgetting Only What Matters: Layer-Selective Unlearning toward Robust LLMs

Researchers introduce FOM-UL, a layer-selective machine unlearning framework that improves forgetting-utility trade-offs and resists knowledge recovery after quantization.

FOM-UL selects transformer layers for unlearning using a forget-to-retain significance score, concentrating parameter updates on layers highly influential for the forget set while leaving most of the model unchanged. It reduces residual memorization versus GA, NPO, KLD, SURE, ReLearn, and LUNAR-based baselines on TOFU, KnowUnDo, and MUSE-style evaluations while preserving retain-set utility. Under 8-bit and 4-bit post-training quantization and adversarial prompts, it maintains stronger suppression of forgotten content, addressing brittleness of diffuse unlearning updates.

arXiv cs.AI / cs.LG / cs.CL·Sep 9, 16:50 UTC · 7d agoAI research1

36

Augustinian BabyLM: What Ostensive Definition Can and Cannot Teach a Small Language Model

Study shows visually grounded token embeddings in a small masked LM persist through training and improve object-property knowledge, but escape standard BabyLM benchmarks.

The paper implements ostensive definition for a small DeBERTa masked language model trained on 10M words, seeding visually grounded tokens with embeddings derived from labeled image regions before training. Visual initialization leaves a persistent, seed-replicated advantage on object-property knowledge (COMPS) and a corpus-tailored Visual-Property Swap benchmark covering color, material, size, and shape, but has no effect on most BabyLM grammar benchmarks. Synthetic grounding of previously unseeded words causally transfers the advantage to exactly those words.

arXiv cs.AI / cs.LG / cs.CL·Sep 10, 17:43 UTC · 6d agoAI research

20

Does Your Agent's Memory Survive a Model Upgrade? A Controlled Study of Memory Portability

A controlled study finds agent memory portability varies sharply: fixed-schema knowledge graphs survive model swaps while compressed notes degrade.

The study compares preserving an agent's history as raw long context, RAG chunks, compressed natural-language notes, or fixed-schema knowledge graphs across model upgrades, using 48 synthetic histories and two open-weight sub-10B-parameter models. Fixed-schema KG accuracy changed by only +0.0004 ± 0.0020 after a writer swap, while compressed NOTES shifted asymmetrically by +9.91 or -13.28 percentage points depending on migration direction. Mixed 50/50 embedding migrations captured only 4.96 of an 11.90-point RAG re-embedding gain; 80% of the NOTES deficit came from information lost at construction, and 81% of the RAG deficit from retrieval failures. Store-only repair of NOTES failed to reach 90% recovery in all 48 cases, while retaining raw histories enabled recovery in 34 of 48 for one direction.

arXiv cs.AI / cs.LG / cs.CL·Sep 4, 16:44 UTC · 12d agoAI research1

32

Fine-tuning a 350M Model for Better Structured Outputs in 100 GRPO Steps

Hugging Face guide fine-tunes a 350M-parameter model with 100 GRPO steps to improve structured output reliability.

A Hugging Face blog post demonstrates fine-tuning a 350M-parameter model using GRPO (Group Relative Policy Optimization) with TRL over 100 training steps. The stated goal is more reliable structured outputs from small language models. No article body was available, so details beyond the title are limited.

Hugging Face Blog·Sep 3, 00:00 UTC · 13d agoAI tools & infra

22

Training a coding model to paint watercolours with TRL and OpenEnv

Hugging Face tutorial trains a coding model with TRL and OpenEnv to paint watercolours through generated code.

A Hugging Face blog walkthrough uses the TRL reinforcement learning library and the OpenEnv environment framework to train a coding model. The target task is generating code that produces watercolour-style drawings, serving as a hands-on reinforcement learning training example. No article body was available in the feed, so specifics are limited to the title.

Hugging Face Blog·Sep 3, 00:00 UTC · 13d agoAI tools & infra1

18

Multi-Vector (Late Interaction) Embedding Models with Sentence Transformers

Hugging Face details building and using multi-vector late-interaction embedding models with Sentence Transformers for retrieval workloads.

Hugging Face published a guide on multi-vector, late-interaction embedding models (ColBERT-style) supported through Sentence Transformers. The post covers how practitioners can build and use these models for retrieval and RAG pipelines. It is a developer tooling and technique write-up, not a security advisory.

Hugging Face Blog·Aug 18, 00:00 UTC · 29d agoAI tools & infra1

42

FlashVector: Agent for Hierarchical Model Serving Stack Optimization

FlashVector agent optimizes all layers of Unity's ad-serving stack, delivering up to 2x model-server throughput and 1.98x latency speedup in production.

FlashVector is an agentic system that optimizes performance across GPU kernels, ML framework computation graphs, model servers, and on-demand feature processing. Deployed in Unity's Vector advertising platform, it achieved up to 2x model-server throughput increase, 1.98x latency speedup, and 1.6x feature-store throughput gain. Optimizations spanned NVIDIA Triton's C++ codebase and the Python feature transformation service, demonstrating extensibility beyond single-kernel tuning.

arXiv cs.AI / cs.LG / cs.CL·Sep 15, 16:27 UTC · 1d agoAI research

30

Lightning Weave: Improving the Accuracy-Efficiency Frontier of Reasoning Models through Capability Composition

Lightning Weave composes capabilities from independently post-trained models via on-policy distillation, improving Qwen3.5-4B reasoning accuracy while cutting tokens.

Lightning Weave is a post-training framework that merges accuracy and efficiency capabilities from independently post-trained specialist models into a single student via on-policy distillation. Each capability is represented as a policy shift, combined via aligned log-ratio shifts and Tilted-Target DOPD, enabling training without serving multiple live anchor models concurrently. On Qwen3.5-4B, it raises HMMT 2025 accuracy from 59.2% to 64.0% with 10.7% fewer response tokens, and LiveCodeBench v5 accuracy from 41.7% to 54.2% with 9.6% fewer tokens. The authors report a state-of-the-art accuracy-efficiency Pareto frontier across diverse students and math/code benchmarks, with code planned for release.

Hugging Face daily papers·Sep 12, 20:00 UTC · 4d agoAI research

52

Perplexity Details Its GPU Embedding Stack: How Ivy, Tulip and ROSE Serve pplx-embed

Perplexity details its GPU embedding serving stack (Ivy, Tulip, ROSE), which reuses LLM prefill/decode kernels, CUDA graphs, and LazyTensors to cut launch overhead.

Perplexity engineers published a deep dive on the serving infrastructure behind pplx-embed, used across Perplexity Search and its API platform. The stack comprises Ivy (Rust HTTP gateway), Tulip (gRPC scheduling and batching), and ROSE (Runtime-Optimized Serving Engine), which reuses LLM prefill and decode kernels rather than running a separate embedding engine. Optimizations include whole-model CUDA graphs with lazy capture and a LazyTensor abstraction that overlaps CPU batch preparation with in-flight GPU work. Benchmarks are reported against vLLM v0.22.0 in BF16, with FlashAttention 4 generally fastest but FlashInfer 3 winning on Qwen-based models at very long sequence lengths.

MarkTechPost·Sep 6, 03:20 UTC · 10d agoAI tools & infra1

33

NCP-ArchPreview Technical Report: Moving towards Latent Space Language Models through Next Concept Prediction

An 8.9B-parameter latent-space language model using next-concept prediction matches OLMo-3-7B pretraining loss with only 51.3% of the training tokens.

NCP-ArchPreview augments next-token prediction with Next Concept Prediction over a product-quantized concept vocabulary built from hidden states, trained jointly end-to-end. The 8.9B model was trained on 5.73T tokens from the Dolma-3 dataset, the largest latent-space language model demonstration to date. It consumes 51.3% of the tokens to reach OLMo-3-7B's final pretraining loss and outperforms it by 2.45 points on the downstream macro-average, including a 5.99-point GSM8K gain. The learned latent space also enables lightweight domain adaptation via a 17M-parameter VQ module and improves speculative drafting accepted length by 4.17%.

Hugging Face daily papers·Sep 8, 20:00 UTC · 8d agoAI research1

52

DataFlex-RL: An Evaluation Platform for RLVR Data Policies

DataFlex-RL benchmark of 13 RLVR data policies on Qwen2.5-7B finds none reproducibly beats uniform sampling under matched GRPO training.

DataFlex-RL is an evaluation platform comparing rollout-selection, reweighting, and mixture data policies for RLVR under a common GRPO recipe. Across 13 configurations and 12 matched seeds with Qwen2.5-7B-Base on 12 math, logic, and science benchmarks, uniform GRPO improved domain-balanced accuracy by 7.76 points, but no alternative policy achieved a statistically significant improvement. A corrected 12-seed Llama-3.1-8B-Base extension found no consistent winner, and math-heavy evaluation summaries were negatively correlated (-0.33) with domain-balanced summaries.

Hugging Face daily papers·Sep 4, 20:00 UTC · 12d agoAI research

35

What Else Needs Fixing? Exploring Cost-Effective Test-Time Compute for Revision Propagation in Artifacts Generated Through Conversation

A new benchmark shows LLMs reach 68.3-93% accuracy propagating local revisions across conversationally generated artifacts, with parallel-sample selection most cost-effective.

The paper introduces a benchmark for revision propagation: when users request a local change, LLMs must identify dependencies and update all affected parts of an artifact generated through conversation, where context lives in the chat history. Nine revision methods, including sequential reflection and parallel sampling variants, were evaluated on gpt-oss-20b/120b, gpt-5.4-mini, and qwen3.5-9b/27b/122b. Baselines scored 68.3-93% accuracy, and selecting among three parallel samples via LLM-based or medoid selection improved accuracy by 2.2-9.7% as the most cost-effective test-time compute strategy. Code and dataset are released.

Hugging Face daily papers·Sep 2, 20:00 UTC · 14d agoAI research

30