Search: “parallel-decoding”

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

Disentangling Representation Evolution in Transformers through Directional Decomposition

Decomposes transformer representation updates into parallel and perpendicular components, linking geometry to editing robustness and better pretraining.

The paper decomposes learned transformer updates into parallel and perpendicular components relative to the hidden state, finding substantial parallel components beyond the residual identity path across pretrained models. Targeted edits reveal exclude-self value-space parallel manipulation is more robust than residual-space or perpendicular alternatives, and perpendicular error separates compression methods more clearly. Applying full-aggregate parallel suppression during from-scratch pretraining lowers validation loss and improves downstream averages, with the value-space variant strongest. Code is released on GitHub.

arXiv cs.AI / cs.LG / cs.CL·Sep 14, 17:56 UTC · 1d agoAI research 2 sources1

35

Online Draft Co-Training for Speculative Decoding in Large-Scale, Long-Context RL Post-Training

NVIDIA researchers detail an end-to-end system for online draft co-training that speeds speculative decoding in large-scale long-context RL post-training.

The paper tackles scaling online draft co-training for speculative decoding in RL post-training, where rollout generation dominates cost. It extends packed, load-balanced zigzag ring attention to merge rank-local branch attention with causal main-sequence attention for context parallelism, and introduces TapChannel to transport target features across pipeline-parallel stages without changing the schedule. Experiments show co-trained drafts tracking the policy baseline with substantial rollout and end-to-end speedups up to 122B parameters and strong scaling at 256K tokens.

Hugging Face daily papers·Sep 6, 20:00 UTC · 9d agoAI research

26

How Lossless Is Lossless Speculative Decoding? The Role of Numerical Precision in Orthrus

Reproduction study finds Orthrus speculative-decoding trajectories match the reference model in only ~45% of cases under BF16, but 100% under FP32.

Researchers independently reproduced Orthrus, a hybrid autoregressive-diffusion architecture claiming lossless speculative decoding via intra-model consensus, testing exact trajectory matching on 1,190 prompts across 12 domains. Under BF16, exact matching occurred in only 45% of cases for the authors' checkpoint and 43% for an independently trained model, with matching probability strongly tied to reference-model response-conditional perplexity. Despite trajectory divergence, downstream lm-eval-harness benchmarks showed no systematic degradation, while FP32 evaluation yielded exact matching on all prompts.

Hugging Face daily papers·Sep 13, 20:00 UTC · 2d agoAI research1

25

A Princeton Researcher Proposes Recurrent Looped Transformer (RLT) that Carries Decoder State across Every Token, Fixing 96 Blocks per Token with Unbounded Temporal Depth

Princeton researcher Yifan Zhang proposes Recurrent Looped Transformer, carrying full decoder state across every token for unbounded temporal depth.

Yifan Zhang's technical report defines the Recurrent Looped Transformer (RLT), pairing a causal encoder with a recurrent decoder whose final output and layerwise sliding-window attention cache carry into every subsequent token with no prompt-response boundary reset. The reference configuration ties 48 encoder and 48 decoder layers, executing 96 logical blocks per token while the state path grows to 48t blocks after t tokens at fixed per-token compute. The report details RL replay contracts that rebuild all states under current parameters and exact prefix snapshots for multi-turn serving, but explicitly reports no measured efficiency, reasoning quality, or scaling results.

MarkTechPost·Sep 13, 17:02 UTC · 2d agoAI research1

40

PARSER: Read in Parallel, Reason in Depth for Long-Context LLM Agents

PARSER uses parallel reader subagents and an RL-trained lead agent for long-context QA, beating baselines and cutting latency up to 11x.

The PARSER paper decouples reading from reasoning: frozen subagents each read one document chunk in parallel while an RL-optimized lead agent iteratively broadcasts queries and aggregates evidence in scatter-gather rounds. On multi-hop QA with 7K to 896K token contexts, a 4B-backbone PARSER beats the strongest sequential memory baseline by 5.7 points on average and 12.0 points at 896K tokens, and a 9B version surpasses DeepSeek-V4-Pro by 6.3 points. Controlled experiments show robustness to evidence position, order, and distance perturbations, with inference latency reduced by up to 11x.

Hugging Face daily papers·Sep 5, 20:00 UTC · 10d agoAI research1

52

The Semantic Bottleneck: Leveraging Semantic Representations for Non-Invasive Speech Decoding

Researchers introduced Brain2Semantics2Text, decoding sentence meaning from non-invasive MEG brain recordings via a semantic bottleneck, improving on prior Brain2Text methods.

The paper proposes Brain2Semantics2Text, a non-invasive speech decoding method that maps sentence-level magnetoencephalography (MEG) responses into a semantic embedding space and inverts those embeddings into natural language. Motivated by evidence that high-level semantic representations are distributed across cortex and evolve on slower timescales, the approach targets meaning rather than phonemes or words, avoiding the need for word-level alignment. The authors report improved sentence-level results compared to prior non-invasive Brain2Text methods despite the low signal-to-noise ratio of neural recordings.

Hugging Face daily papers·Sep 8, 20:00 UTC · 7d agoAI research2

25

Unlocking Lossless Speedups in LLMs via Discrete Diffusion

Uno pairs autoregressive LLMs with lightweight diffusion weights to draw multiple tokens in parallel, delivering up to 3x lossless speedup without a draft model.

The paper introduces diffusion-augmented LLMs: autoregressive weights trained with the standard next-token objective plus lightweight diffusion weights trained via a Diffusion Distillation phase to emit multiple tokens in parallel. Psi-Spec samplers enable lossless acceleration without the separate draft model required by speculative decoding. The 8B Uno model outperforms the 26B open DiffusionGemma and proprietary Mercury 2 on agentic tool use, coding, and long-context reasoning benchmarks, with up to 3x throughput gains over the base model at all evaluated batch sizes. Code and checkpoints are released publicly.

Hugging Face daily papers·Sep 2, 20:00 UTC · 13d agoAI research

55

Memory-Efficient Designs for Word-Wise Universal Fully Homomorphic Encryption

BXT framework mitigates FHE memory bottlenecks via ciphertext compression, serialization, delayed seeding, and digit pruning, achieving up to 3.8x CNN inference speedup.

A new paper proposes BXT, an optimization framework for word-wise Universal Fully Homomorphic Encryption that targets the memory bottleneck rather than compute. It combines four techniques: ciphertext compression via seed regeneration, bit-packed ciphertext serialization for L2-to-L1 transfers, delayed PRNG-heavy offline seed generation across aggregated operations, and fault-aware ciphertext digit pruning. On CNN inference, the BXT-CSO50 configuration achieves up to 3.8x speedup over a 100x GPU baseline with under 1% accuracy loss at 50% comparison precision.

arXiv cs.CR·Sep 4, 06:03 UTC · 12d agoResearch

24

What Did I Just Say? Self-Listening for Full-Duplex Speech Models

Researchers propose Self-Listening, a full-duplex speech approach feeding realized model speech back as input to improve interruption recovery.

Full-duplex spoken language models can listen and speak simultaneously, but asynchronous text generation, speech synthesis, and playback cause mismatches between what a model believes it said and what the user heard. The paper defines the resulting recovery problem as anchor interruption and proposes Self-Listening, which interleaves user speech, model text, and played speech as input streams. The authors also release AnchorSpeech, a benchmark with homogeneous training and test splits tracking which ordered response items were actually spoken. Experiments show self-listening models achieve better anchoring performance than full-duplex baselines.

Hugging Face daily papers·Sep 3, 20:00 UTC · 12d agoAI research

22

Speculative Decoding in vLLM on AMD GPUs

vLLM benchmarks speculative decoding on AMD Instinct MI300X and MI355X GPUs across five drafting methods including EAGLE-3 and native MTP.

The vLLM project documents draft-and-verify speculative decoding support for AMD GPUs via ROCm, comparing native MTP, Gemma 4 MTP, EAGLE-3, DFlash, and DSpark drafting approaches. Output-token throughput effects varied with drafting method, proposal length, model family, draft checkpoint, workload, and acceptance behavior. The post also covers how to enable each method plus practical tuning and observability considerations.

Hacker News · AI·Sep 7, 09:26 UTC · 9d agoAI tools & infraHN 141↑ · 52 comments1

40

Robust Coverless Linguistic Steganography via Sentence Embedding Space with Global Resynchronization

Researchers propose a coverless steganographic framework encoding messages as hierarchical clustering paths in sentence embedding space with a Global Resynchronization Mechanism for robustness.

An arXiv paper proposes encoding secret messages as hierarchical clustering paths in the sentence embedding space rather than token space, improving decoding stability against word- and sentence-level textual perturbations. A Global Resynchronization Mechanism (GRM) reframes variable-length bitstreams as discrete symbols anchored to semantic subspaces to prevent bit-slippage. Experiments show substantial robustness improvements while maintaining embedding capacity and resistance to statistical analysis.

arXiv cs.CR·Sep 4, 10:22 UTC · 12d agoResearch

18

Rethinking Heterogeneous System Disaggregation for Subquadratic Attention

SQD disaggregates LLM inference by quadratic versus subquadratic attention layers, improving energy efficiency up to 56% on heterogeneous systems versus GPU-only baselines.

SQD (SubQuadratic Disaggregation) splits decode not by operator type but by quadratic versus subquadratic attention, matching their distinct arithmetic intensity and memory footprints. For sparse attention LLMs it separates top-k selection (requiring full KV indexing) from top-k attention plus FFN; for linear and sliding-window models it separates dense attention layers from subquadratic layers plus FFN. On an adjusted 8xB200 heterogeneous proxy, tokens-per-joule improves 53% on GLM 5.2, 31% on Nemotron 3 Ultra, and 56% on Gemma 4 31B. A Rubin plus LPX analytical model shows 1.2x-1.5x tighter achievable latencies and up to 3.6x higher throughput versus attention-FFN disaggregation.

arXiv cs.AI / cs.LG / cs.CL·Sep 11, 17:55 UTC · 4d agoAI research

28

Task-Directed Residual AddUNet:Perfect-Reconstruction Routing for Full-Rate Representations

Paper recasts additive U-Net skip structure as a perfect-reconstruction filter bank and adds full-rate residual routing for task-directed representations.

The work proves a constrained additive U-Net's survivor-skip structure is exactly equivalent to a critically sampled perfect-reconstruction filter bank, and removes complementary-subband restrictions via a full-rate formulation. A Residual Full-Rate PR architecture routes task-irrelevant or redundant structure away from the task pathway while guaranteeing exact reconstruction without invertible operators, a matched synthesis bank, or a learned decoder. On TIMIT, the front-end improves test PER from 28.60±2.09% to 25.76±0.41% with recognizer and training held fixed.

arXiv cs.AI / cs.LG / cs.CL·Sep 14, 16:50 UTC · 1d agoAI research

12

Pre-Whitening and BCJR Posterior Distillation for Bi-LSTM Detection in Faster-than-Nyquist Signaling

Study shows nested-window Bi-LSTM architectures do not improve faster-than-Nyquist detection; pre-whitening plus BCJR distillation cuts bit error rates.

Across roughly 260 controlled trainings, processing nested intersymbol-interference windows in separate recurrent branches never significantly beat a plain Bi-LSTM at matched parameter budgets. The authors attribute the limitation to the observation model rather than architecture, and instead pre-whiten inputs and distill BCJR soft posteriors into the network. With 3.4% more parameters, the method reaches 1.05x the BCJR bit error rate at compression factor 0.8 and 1.89x at 0.7, improving to 1.47x with a wider whitened window.

arXiv cs.AI / cs.LG / cs.CL·Sep 7, 17:07 UTC · 8d agoAI research

10

Has MIMO decoding been proved hard from lattice problems?

Researchers show the published lattice-hardness proof for MIMO decoding fails, as Regev's LWE reduction structure does not carry over to non-modular MIMO.

The paper re-examines Dean and Goldsmith's proposed polynomial-time reduction from lattice problems to MIMO decoding, which adapted Regev's reduction for learning with errors (LWE). Prior works had presented attacks and counterexamples against the construction, leaving the reduction's precise validity unclear. The authors identify which structural features of the LWE reduction fail to transfer to the non-modular MIMO setting, showing the published proof does not establish the claimed hardness of MIMO decoding. They distinguish flaws in the hardness proof from direct attacks on specific parameter choices and do not rule out physical layer security for MIMO systems in general.

arXiv cs.CR·Sep 4, 11:23 UTC · 12d agoResearch

25

ECHO: A Matched-Contrast Benchmark for Context-Sensitive Turn-Taking in Full-Duplex Dialogue

ECHO benchmark pairs identical-overlap Chinese dialogue examples with contrasting contexts, revealing most full-duplex systems bias toward yielding the floor.

ECHO is a paired diagnostic benchmark for Chinese full-duplex turn-taking that matches examples with identical overlap transcripts but contrasting preceding multi-turn contexts, one requiring Yield and the other Keep. It also includes off-talk examples for diagnosing unnecessary yielding and introduces pair accuracy, which grants no credit to constant-action policies. Experiments across multiple full-duplex systems show most exhibit a pronounced Yield bias, performing substantially better on interruptions than backchannels. The benchmark and metadata will be publicly released.

arXiv cs.AI / cs.LG / cs.CL·Sep 15, 15:58 UTC · 22h agoAI research

40

FreeFlow: A Bias-free Hierarchical Transformer for Optical Flow Estimation

FreeFlow is a bias-free hierarchical transformer achieving state-of-the-art optical flow results on Sintel, KITTI-2015, and Spring benchmarks.

FreeFlow replaces task-specific inductive biases like correlation volumes and iterative warping with a single feed-forward encoder-decoder combining window, shifted-window, and reduced-resolution global attention. It reaches 0.68/1.48 EPE on Sintel Clean/Final, 3.23 Fl-all on KITTI-2015, and 3.192 1px on Spring. The architecture scales consistently from small to large variants and remains memory efficient at 1080p inference.

Hugging Face daily papers·Sep 9, 20:00 UTC · 6d agoAI research

30

Grouped Value Attention: Efficient KV Caching via On-Demand Key Reconstruction

Grouped Value Attention stores grouped values and reconstructs content keys via a learned linear map, cutting KV-cache size about 45-47% versus GQA.

GVA stores only grouped values and reconstructs content keys with a learned linear map absorbed into the query at decode time, while a small shared decoupled RoPE channel preserves positional information via a separately cached positional key. At 350M parameters trained on 30B FineWeb-Edu tokens, the 16-dimensional positional variant scores 44.18 average accuracy across five tasks versus 44.36 for GQA and 43.88 for MLA. Custom decoding kernels are in development with an open-source release planned.

Hugging Face daily papers·Sep 7, 20:00 UTC · 8d agoAI research

42

LACE: Layer-Wise Compression for Dynamic Frame Rate Codecs

LACE introduces layer-wise compression for dynamic frame rate audio codecs, cutting sequence lengths and speeding TTS inference while preserving quality.

LACE (Layer-Adaptive Codec Encoding) applies an independent compression step at each quantization layer of a neural audio codec, enabling layer-specific segmentation boundaries instead of shared ones. Union alignment and boundary anchor mechanisms keep durations consistent for downstream text-to-speech. On LibriTTS, LACE achieves a better rate-quality tradeoff than prior dynamic frame rate codecs and improves TTS inference efficiency at competitive synthesis quality. Code is released in the ESPnet3 codec recipe.

arXiv cs.AI / cs.LG / cs.CL·Sep 15, 17:46 UTC · 20h agoAI research1

15

Why Is Video Still So Expensive? A Survey of Inference-Efficiency Mechanisms in Video and Audiovisual LLMs

A survey catalogs inference-efficiency techniques for video and audiovisual LLMs, mapping bottlenecks in sampling, encoding, token reduction, and LLM decoding.

This survey covers inference-efficiency mechanisms for visual and audiovisual video LLMs, reporting reductions in parameters, FLOPs, latency, memory, and token counts. It organizes methods by pipeline stage, covering frame sampling, modality encoding, connector-level token reduction, and LLM prefilling and decoding for systems built since late 2022. The authors compile accuracy-cost comparisons under shared host models and input protocols, identify gaps in audiovisual efficiency and standardized evaluation, and maintain a public repository.

Hugging Face daily papers·Sep 8, 20:00 UTC · 7d agoAI research

40

AdamX: Cosine similarity meets gradient descent

Researchers propose AdamX, a cosine-similarity-based first-order optimizer with variance rectification that matches Adam-class convergence across benchmark datasets and architectures.

The paper introduces AdamX, a first-order optimizer that uses cosine similarity as an adaptive mechanism for controlling update magnitudes, plus a variance rectification scheme for smoother optimization early in training. The method is described as scalable, model-agnostic, and straightforward to integrate into existing pipelines. Empirically, AdamX shows competitive convergence rates measured by epochs to reach performance thresholds under a fixed hyperparameter budget, with code and experiments released on GitHub.

arXiv cs.AI / cs.LG / cs.CL·Sep 10, 17:42 UTC · 5d agoAI research

22

RAPID: A Real-Time Defense Against Unauthorized Model Distillation for Text-to-Image Services

RAPID embeds defensive perturbations in a T2I model's shared VAE decoder to block unauthorized black-box distillation in real time.

The paper defends text-to-image services against model theft via black-box output-based distillation, where adversaries collect prompt-image pairs to train substitute models. RAPID integrates defensive perturbations into the shared VAE decoder using self-referenced latent maximization plus reconstruction-guided color regularization, avoiding costly sample-wise online optimization. Across four T2I models and four datasets versus five baselines, it consistently degrades substitute-model generation quality while preserving visual fidelity.

arXiv cs.CR·Sep 14, 16:11 UTC · 1d agoAI safety & security

35

Shuffling is Not Enough: Breaking Permutation-Based Model Confidentiality in Hybrid FHE Inference

Attack breaks permutation-based model confidentiality in hybrid FHE inference, recovering all ResNet-20 linear layers exactly with d+1 queries per layer.

The paper shows output-permutation plus noise fails to protect model confidentiality in hybrid FHE inference: d+1 admissible queries recover an exact permutation-invariant summary of a d-input linear layer, and shuffle-model DP amplification premises cannot hold under correctness-bounded noise. The authors recovered all linear layers of a Safhire-style ResNet-20 end-to-end from TFHE transcripts with zero error, using 5,712 total queries. Exact per-layer recovery was also confirmed on pretrained ImageNet-scale CNNs and ViT-B/16. Leaked layer spectra enable model fingerprinting, lineage attribution, and improved logit-based extraction, while suppressing them destroys inference utility.

arXiv cs.CR·Sep 11, 14:37 UTC · 4d agoResearch1

45

SpecGuard: Inference-Time Backdoor Detection For Free

SpecGuard detects backdoored LLM behavior at inference time using speculative decoding acceptance rates, adding no extra model computation.

Researchers propose SpecGuard, an inference-time backdoor detector that repurposes draft-token acceptance rates from speculative decoding as a detection signal at zero added model-computation cost. When a trigger shifts the target model toward attacker-controlled behavior, the clean draft model's acceptance rate changes, exposing the backdoor; the paper formalizes when this signal appears and shows suppressing it weakens the backdoor. Experiments across diverse backdoor types and model families show reliable detection, including stealthy cases invisible to input-level filters. Speculative decoding is positioned as a free, always-on monitor for frequently updated deployed models.

arXiv cs.CR·Sep 10, 16:51 UTC · 5d agoAI safety & security 2 sources2

45

ROSETTA: Efficient and Accurate Privacy-Preserving LLM Decoding via Hybrid CKKS/TFHE Evaluation

ROSETTA is a hybrid CKKS/TFHE homomorphic encryption framework for privacy-preserving LLM decoding, achieving up to 4.8x Softmax and 2.1x end-to-end speedups.

The paper proposes ROSETTA, a hybrid CKKS/TFHE fully homomorphic encryption framework for private inference on generative LLMs, targeting the nonlinear operations that dominate autoregressive decoding cost. It introduces an adaptive segmented lookup-table protocol based on TFHE and a scheme-aware operator-selection framework that assigns each nonlinear operator to CKKS or TFHE to minimize latency. Experiments show up to 4.8x Softmax speedup and 1.5-2.1x end-to-end decoding speedup over the state-of-the-art CacheMir framework.

arXiv cs.CR·Sep 15, 09:51 UTC · 1d agoResearch

28

ConvMem: Convolutional Memory for Long-Context Reasoning

Researchers propose ConvMem, a training-free framework treating LLMs as convolutional kernels for parallelizable long-context reasoning beyond fixed context windows.

ConvMem reformulates long-context reasoning as a hierarchical convolution in which the LLM summarizes text segments hierarchically, shortening the reasoning path from a linear chain to a logarithmic tree. It uses configurable strides, skip connections, and multi-kernel convolution to capture evidence, decompose queries, and enable massive parallelization across segments and reasoning threads. On RULER-HotpotQA and RULER-2WikiMultiHopQA it outperforms training-free baselines and avoids the out-of-distribution overfitting seen in RL-trained approaches like MemAgent.

arXiv cs.AI / cs.LG / cs.CL·Sep 9, 16:53 UTC · 6d agoAI research

38

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 · 8d agoAI research

27

Scalability Analysis of Distributed Kolmogorov-Arnold Network Training on High-Performance Computing Systems

An empirical study shows distributed Kolmogorov-Arnold Network training reaches 74.7% parallel efficiency at 8 A100 GPUs, with overheads driven by All-Reduce choices.

The study evaluates data-parallel Kolmogorov-Arnold Network (KAN) training on the FinisTerrae III supercomputer using up to 8 NVIDIA A100 GPUs across 4 nodes with PyTorch Distributed Data Parallel. Strong scaling yields 5.97x speedup and 74.7% parallel efficiency at 8 GPUs, comparable to conventional deep learning workloads, while communication overhead ranges from 1.3% to 6.1%, driven mainly by All-Reduce algorithm selection and inter-node latency rather than KAN's edge-wise gradient structure. Weak scaling shows an initial single-to-multi-GPU throughput drop followed by stability, and the parameter-to-memory ratio improves with model size even as training time scales unfavorably. The authors provide GPU topology and model-size deployment guidelines for KAN training.

arXiv cs.AI / cs.LG / cs.CL·Sep 7, 16:45 UTC · 8d agoAI research1

26

Subgroup Packing for Batched PASTA Transciphering

Interleaved subgroup packing cuts PASTA homomorphic transciphering server cost by a median 1.60x at the expense of remaining noise capacity.

The paper studies how record word layout affects the cost of PASTA transciphering, where a server converts symmetrically encrypted records into homomorphic ciphertexts. A subgroup layout interleaves records so cyclic shifts preserve positions, reducing required displacements from 255 to 128 versus contiguous packing. In HElib across twelve paired corpora under six homomorphic keys, the median direct-to-subgroup server cost ratio is 1.60 including fresh generation, conversion, and two queries, but with less remaining noise capacity.

arXiv cs.CR·Sep 11, 09:24 UTC · 5d agoResearch

15