The New Stack
·
4 hours ago
The article argues that AI tools like Claude Code and GitHub Copilot have not actually shifted the software development bottleneck from coding to code review, because the real constraint lies downstream in deployment and release processes. Research from Octopus Deploy shows 92% of teams ship code in batches of 2 to 50 changes rather than deploying individual changes, with changes accumulating after code review. Addressing this deployment bottleneck, not code review speed, is where organizations should focus to realize benefits from AI-assisted coding tools.
arXiv cs.CL
·
18 hours ago
Researchers propose ASRD, a training-free framework that improves diffusion language models by distinguishing reliable anchor tokens from uncertain ones during decoding to reduce error propagation and local reinforcement failures. The method achieves accuracy improvements of up to 6.4% on math and coding benchmarks while accelerating inference throughput by up to 7.2 times compared to existing remasking approaches. This addresses the speed-versus-quality trade-off in parallel decoding of diffusion language models by using anchor-guided generation and anchor-perturbed verification mechanisms.
arXiv cs.CL
·
18 hours ago
Researchers introduced SLIM, a framework that addresses context limitations in long-horizon agentic search by separating retrieval into distinct search and browse tools and periodically summarizing trajectories. With the o3 model, SLIM achieved 56% on BrowseComp and 33% on HLE benchmarks while using 4-6 times fewer tool calls than existing open-source frameworks. The approach enables agents to conduct longer, more focused searches with reduced hallucinations and lower computational costs.
arXiv cs.CL
·
18 hours ago
Researchers formalized task-aware execution-scope estimation for LLM agents, proposing E3, a method where agents estimate task difficulty before committing resources rather than re-reading all available information. On MSE-Bench with 121 edits, E3 achieved 100% success while reducing token usage by 91% and inspected files by 92% compared to the strongest baseline, and real-world testing on open-source code confirmed the approach produces leaner execution with comparable task success. This work aims to reduce execution redundancy in AI agents through task complexity awareness before they perform multi-step workflows.
arXiv cs.CL
·
18 hours ago
A survey analyzes efficient inference techniques for diffusion large language models, which theoretically enable parallel generation compared to standard autoregressive models. The research introduces a latency decomposition framework to separate algorithmic, architectural, and system-level factors affecting inference speed, categorizing acceleration techniques across three dimensions. The work provides benchmarking guidelines to help practitioners achieve practical speedups in real-world deployments of dLLMs.
arXiv cs.CL
·
18 hours ago
Researchers developed Light-MER, a multimodal emotion recognition framework that uses knowledge distillation to compress large language models into sub-1-billion parameter models while maintaining performance. The lightweight model achieved state-of-the-art results on nine benchmark datasets while significantly reducing computational requirements for deployment. This approach enables real-time emotion recognition on resource-constrained devices like robots and mobile phones without sacrificing recognition quality.
arXiv cs.CL
·
18 hours ago
JoLT is a compression method for key-value caches in large language models that applies Tucker decomposition to the token and feature axes while using Johnson-Lindenstrauss rotated residuals to recover discarded energy. The method achieves 2-3x compression with perplexity, GSM8K accuracy, and needle-in-a-haystack retrieval remaining within statistical noise of uncompressed baselines on Mistral-7B-v0.3 and LLaMA-2-13B. This reduces memory costs during transformer inference without degrading model performance, enabling longer context lengths or larger batch sizes within fixed memory budgets.
arXiv cs.CL
·
18 hours ago
Researchers developed a memory-augmented speculative execution system for LLM agents that uses three online memory systems to improve prediction accuracy during idle environment time. The approach achieved 19-39% relative accuracy improvement on action prediction and up to 2.5x improvement on observation prediction tasks with repetitive action spaces. Agents can now predict and pre-launch next steps more accurately while maintaining identical trajectories to non-speculative execution with zero added wall-clock cost.
arXiv cs.CL
·
18 hours ago
Researchers propose EcoSpec, a cost-aware speculative decoding framework that optimizes draft token selection in Mixture-of-Experts language models by reducing unnecessary expert activation rather than only maximizing acceptance likelihood. Testing on three large-scale MoE models including DeepSeek-V3.1 (671B parameters) shows speedups of up to 1.62x compared to standard speculative decoding. This approach improves inference efficiency by reusing already-activated experts during token verification, reducing memory traffic and computational overhead.
arXiv cs.CL
·
18 hours ago
A study compared using English BERT models with translated non-English data against developing native-language BERT models across six NLP tasks in Bulgarian, Chinese, Dutch, Italian, and Russian. The translation-based approach matched or outperformed native models in 53.3 percent of cases, with best results in Question Answering, Part-of-Speech Tagging, and Natural Language Inference, but struggled with Named Entity Recognition and tasks requiring cultural understanding. The findings suggest translation-based fine-tuning is a computationally efficient alternative for extending NLP capabilities to low-resource languages, particularly for languages typologically similar to English and syntactic tasks.
arXiv cs.CL
·
18 hours ago
A study of 2,848 Claude Code API runs found that reducing tokens in tool outputs does not reliably lower end-to-end billing costs for coding agents. Prompt caching accounted for approximately 87% of the reconstructed cost composition, while an intervention removing 38% of tool-output tokens actually increased paired costs by 6.8%. The findings suggest context-reduction systems should be evaluated by success-adjusted billed cost rather than token reduction metrics, as aggressive compression can corrupt critical information needed for task completion.
arXiv cs.CL
·
18 hours ago
CityBehavEx is an urban simulation platform that uses fine-tuned cross-encoders combined with human mobility models to generate city-scale agent behaviors, reducing reliance on large language model calls. The system simulates 100,000 agents over 75 days in under one hour on a single GPU while producing mobility patterns that align with real-world spatial, temporal, and semantic distributions. This approach enables researchers to validate synthetic urban behaviors against empirical data and inspect agent activities at scale.
arXiv cs.CL
·
18 hours ago
Researchers introduced TAKE, a text dataset distillation framework that reduces datasets to 0.1% of their original size while maintaining performance on NLP tasks by using influence functions to identify and weight the most informative training samples. The method achieved extreme compression—down to 20 samples per class—while preserving downstream task accuracy on text classification and natural language inference. This approach reduces the computational cost of training, fine-tuning, and continual learning on large text corpora.
arXiv cs.AI
·
18 hours ago
Researchers developed HELP, a robot post-training pipeline where two human operators supervise twelve robots simultaneously using role specialization and an automatic rollout segmentation system to identify useful training data. The system achieved 80%-95% success rates across four manipulation tasks and improved task throughput by 1.7x to 4.2x compared to the base model. This approach enables more efficient human-robot collaboration by reducing operator workload and focusing training on meaningful robot behaviors and failures.
arXiv cs.AI
·
18 hours ago
Researchers introduced AnchorPrune, a method that reduces computational costs in vision-language models by removing redundant visual tokens while preserving task-relevant information. The technique maintains 97.6% of full-model performance using only 160 out of 2,880 visual tokens on LLaVA-NeXT-7B. The training-free framework works by first protecting the most query-relevant tokens as an anchor, then selectively adding complementary context to improve inference efficiency without retraining.
arXiv cs.AI
·
18 hours ago
Researchers introduced Ember, an optimizer designed specifically for embedding and language model head matrices in transformers that reduces memory requirements from O(2VD) to O(V+D) compared to Adam. Ember uses only kilobytes of optimizer state while improving performance across supervised finetuning, reinforcement learning, and pretraining tasks. The work shows that token optimization follows a simple 1D trajectory rather than a heavily nonconvex landscape and is compatible with existing distributed training frameworks.
arXiv cs.AI
·
18 hours ago
Researchers developed DiT-Pruning, a post-training pruning method designed specifically for Diffusion Transformers that addresses limitations of existing pruning techniques by introducing customized saliency metrics and clustering-aware granularity. The method achieved only 0.001 CLIP score loss on FLUX.1-dev at 50% sparsity, compared to significant degradation from prior pruning approaches. This enables efficient deployment of diffusion models with reduced computational requirements while maintaining image generation quality.
arXiv cs.AI
·
18 hours ago
Researchers developed TIME (Temporally Informed Motion Embedding), a video representation learning approach that uses motion point-tracks with masked autoencoders instead of language-paired visual data. The method achieves performance comparable to state-of-the-art models while requiring 10,000 times less training data. This approach enables more efficient video models with improved temporal understanding without language-dependent training constraints.
arXiv cs.AI
·
18 hours ago
Researchers introduced DIVE, a compression method that reduces the size of language model embeddings using a residual adapter combined with self-limiting gradient updates and geometry distillation. DIVE achieved the strongest performance across all five BEIR benchmarks when tested at 128d and 256d output dimensions against six baseline methods. The technique enables cheaper storage and faster search of embeddings without the overfitting problems that plague supervised compression when training labels are limited.
arXiv cs.AI
·
18 hours ago
Researchers introduced ELF, a family of three encoder-free ECG-language models that interpret electrocardiogram data without relying on pretrained encoders. The models achieve competitive or superior performance compared to existing ECG-language models while using simpler architectures and training procedures. The simplified design reduces computational complexity and training requirements for automated ECG interpretation systems.
arXiv cs.AI
·
18 hours ago
Inverse-LLaVA proposes a multimodal architecture that projects text embeddings into visual representation space rather than the traditional approach of projecting images into text token spaces. The method achieves competitive performance on nine multimodal benchmarks while eliminating the need for explicit alignment pretraining and reducing dependence on large image-text datasets. This approach suggests that effective multimodal reasoning can occur without traditional alignment pretraining, enabling more efficient system design that decouples representation structure from supervision requirements.
arXiv cs.AI
·
18 hours ago
Researchers introduced NSNQuant, a calibration-free vector quantization technique that compresses key-value cache in large language models through double normalization with Hadamard transform. The method achieved up to 3 times throughput improvement over full-precision baselines in 1-bit and 2-bit quantization settings. This approach eliminates the need for calibration datasets and enables better generalization across different data distributions during LLM inference.
arXiv cs.AI
·
18 hours ago
A new reinforcement learning algorithm called KGRL combines symbolic knowledge from Datalog rule bases with gradient-guided parameter optimization to improve training efficiency in tasks requiring both discrete action selection and continuous parameter tuning. The method prunes infeasible actions and constrains parameter spaces based on domain knowledge, achieving superior sample efficiency and return compared to existing baselines. This approach enables agents to learn constraint-aware decisions while providing interpretable explanations of action selection and parameter constraints.
arXiv cs.AI
·
18 hours ago
Researchers introduced Policy of Thoughts (PoT), a framework that enables language models to refine their reasoning strategies during test time by updating adapter weights based on execution feedback from failed attempts. A 4-billion-parameter model using PoT achieved 49.71% accuracy on LiveCodeBench, surpassing GPT-4o and DeepSeek-V3 while being 50 times smaller. The approach allows models to dynamically adapt their reasoning process for individual problems rather than relying on a frozen policy.
arXiv cs.AI
·
18 hours ago
Researchers developed Cluster-based Sequential Feature Selection (CSFS), a wrapper-based method for automatically selecting relevant input variables in wind and solar power prediction models. The method achieved comparable predictive performance to standard sequential feature selection while reducing computational cost by an average of 21%. This approach addresses the lack of systematic feature selection methods in renewable energy forecasting by providing an efficient, model-agnostic tool available as open-source software.
arXiv cs.AI
·
18 hours ago
A study of 410 German companies examined how AI tools including generative AI and predictive analytics are being integrated into human resource management functions. The research found that AI adoption is primarily driving efficiency gains and cost rationalization rather than strategic people-centered improvements, despite potential for enhancing talent development. Implementation is shaped by organizational factors including digital infrastructure, co-determination frameworks, and concerns about data governance and algorithmic transparency.
arXiv cs.AI
·
18 hours ago
Kaleido is an algorithm-hardware co-design that accelerates video diffusion transformers by exploiting channel-wise spatiotemporal correlations in latent space to skip redundant computations. The system achieves up to 5.9x speedup and 16.0x energy savings compared to state-of-the-art accelerators across three mainstream vDiT models. This approach reduces the computational bottleneck of self-attention in video generation while maintaining generative quality above 17 dB PSNR.
arXiv cs.AI
·
18 hours ago
Researchers built a compact 1.11 million parameter convolutional network that achieved 99.73% accuracy on Devanagari handwritten character recognition, matching the performance ceiling that larger models also reach. The model reaches an intrinsic error floor of 11 errors that no configuration can statistically improve upon, despite being 15.6 times smaller than previous state-of-the-art approaches. The work demonstrates that benchmark saturation has been reached for this dataset, with additional model scaling providing no meaningful performance gains.
arXiv cs.AI
·
18 hours ago
Researchers introduced MemCon, a framework that models memory operations in LLM agents as a Markov Decision Process to learn adaptive policies for when and how to retrieve, inject, or consolidate memories. The method achieved up to 15.2 point improvements in task success across 6 benchmarks while reducing token consumption by 5-20 percent. This allows memory management to adjust dynamically based on task context rather than relying on fixed heuristics.
arXiv cs.AI
·
18 hours ago
Researchers proposed Attention Head Reweighting (AHR), a method that adapts large language models to new text-classification tasks by learning a single scalar weight per attention head rather than modifying many parameters. AHR outperformed LoRA on limited-sample tasks while modifying only 0.0001% of model parameters, achieving 200-1000x parameter reduction. This approach enables more efficient model adaptation in data-scarce domains like security while providing interpretability into how attention heads contribute to few-shot learning.
arXiv cs.AI
·
18 hours ago
Researchers applied model pruning to AudioLDM, a text-to-audio diffusion model, to reduce computational requirements. The pruning reduced U-Net parameters by 83% and multiply-accumulate operations by 39% while maintaining or improving generation quality after lightweight finetuning. The pruned model initially lost ability to generate certain sounds like gunshots and sirens, but these capabilities were mostly recovered through finetuning.
arXiv cs.AI
·
18 hours ago
EMAGN is a traffic forecasting model that reduces the computational complexity of self-attention mechanisms from O(N^2 d) to O(NMd) through learned clustering of attention vectors. On PEMS-BAY and METR-LA datasets, EMAGN achieves 2.7-3.2% MAE difference from full-attention baselines while reducing training time by 32%, inference time by 38%, and GPU memory use by 58%. The efficiency gains enable operation with 16 attention heads on 11 GB GPUs where existing full-attention models run out of memory, allowing deployment of larger model configurations.
arXiv cs.AI
·
18 hours ago
ShortOPD is a training method that recovers the generation quality of pruned large language models by using on-policy distillation with a dynamic schedule that allocates training budget based on effective prefix lengths rather than full rollout lengths. The method achieves 1.6x to 4.4x improvement over standard recovery approaches and reaches performance matching 8192-token rollouts in one quarter of the training time (8.5 versus 35.9 hours) while using 71% fewer rollout tokens. This enables structured pruning to maintain practical generation quality needed for deployment rather than only improving benchmark scores on recognition tasks.
arXiv cs.AI
·
18 hours ago
Researchers introduced Samba, a model using Mamba architecture for audio-visual navigation tasks that replaces conventional recurrent networks with state space encoders. The model achieved an 11.3% improvement in navigation success rate on the Matterport3D dataset compared to existing approaches. The architectural changes enable more efficient processing of multimodal sequences and better generalization to unseen environments and novel sounds.
arXiv cs.AI
·
18 hours ago
WaterMoE proposes a watermarking technique for Mixture-of-Experts language models that embeds watermarking signals through controlled perturbation in the expert selection process rather than as post-processing. The method achieves negligible quality degradation with only 1% additional inference latency and up to 4× speedup compared to existing watermarking methods across various generation tasks. This enables practical deployment of watermarking in latency-critical systems while maintaining model fidelity.
arXiv cs.AI
·
18 hours ago
Researchers developed a privacy-centric framework for running large language models across edge devices and cloud servers, where edge devices handle preprocessing and partial computation while the cloud performs core inference, with all transmitted data encrypted. The system reduces per-token latency by up to 46.1% and downlink payloads by up to 67.4% compared to baseline split inference approaches. This enables consumer and embedded devices to run LLMs efficiently while protecting user prompts and dialogue data from full exposure to cloud servers.
arXiv cs.AI
·
18 hours ago
Researchers propose Deep Interaction, a method allowing users to directly edit and correct errors in reasoning steps of large language models rather than requesting regeneration. The approach achieves a 25% improvement in correction success rate and reduces token usage by 40% on STEM reasoning tasks compared to baseline methods. This enables more efficient refinement of model outputs when reasoning errors occur in multi-step problem-solving.
