AI Firehose

Thermal-Headroom Routing (THR) aids robots in extreme heat by predicting chip temperatures to avoid throttling, reducing deadline misses by 7.5% and energy use by up to 12-13% compared to latency-focused methods. This ensures reliability in harsh environments. https://arxiv.org/abs/2610.35231

S3AC is a unique actor-critic method that boosts token-level learning and prompt selection in LLM reasoning. Difficulty-Aware Prompt Sampling enables better performance in long chain-of-thought reasoning, setting a high efficiency standard in AI training. https://arxiv.org/abs/2610.35518

Researchers present CarbonRouter, a framework that routes multi-turn conversations in LLM services to cut emissions by tackling costly historical cache recomputation. It achieves a 16.1% CO2 reduction over existing methods, paving the way for greener AI solutions. https://arxiv.org/abs/2610.35230

This study presents a framework for training large language models via distribution matching, enhancing the understanding of algorithms. Conditional KL (CKL) regularization improves performance by maintaining diversity while easing reward constraints. https://arxiv.org/abs/2610.33740

"Straighter" visual representations, akin to human perception, improve stability and robustness in deep learning video models. Adversarial training and biologically-inspired methods provide insights for enhancing computer vision toward human-like performance. https://arxiv.org/abs/2610.34678

MIT and EPFL explore methods for compressing Key-Value caches in autoregressive transformers, reducing memory needs while preserving attention effectiveness. The findings uncover alignment issues in key/query relations, improving large language model efficiency. https://arxiv.org/abs/2610.35020

AGENTMETER is a new evaluation framework for command-line agents that combines task quality and execution costs, providing nuanced insights into LM–CLI pairs. This innovation addresses biases in current benchmarks, enhancing effectiveness for large language models. https://arxiv.org/abs/2606.21140

V3AST-Bench, a benchmark rigorously testing coding agents' safety, verifies real-world consequences, revealing that even advanced models mismanage risks. This approach emphasizes enhanced training for safety compliance throughout execution. https://arxiv.org/abs/2610.35188

A study presents CoCHI, a method that targets overconfident errors in large language models by intervening on attention heads. It enhances reliability and calibration without compromising predictive accuracy, reshaping model trust in critical contexts. https://arxiv.org/abs/2610.34994

Research reveals speedups of alternative optimizers over AdamW are overstated, with testing indicating gains diminish as model size increases. The study stresses careful hyperparameter tuning, suggesting many efficiency claims emerge from inadequate comparisons. https://arxiv.org/abs/2509.02046

FRIGID enhances molecular structure generation from mass spectra, nearly tripling accuracy of previous methods while improving scalability during training and inference. This framework fosters rapid, precise identification of small molecules in scientific research. https://arxiv.org/abs/2604.16648

Research shows deep generative models struggle with heavy-tailed or multimodal distributions due to transport issues. The authors propose bijective reparametrizations to improve model adaptation in finance and climate science. https://arxiv.org/abs/2610.34484

Introducing CODEMEEPO, a pioneering multi-agent workflow for repository-level migration. By combining deterministic routing and validation, it passes 18% more tests than competitors and increases performance by 25% on average. https://arxiv.org/abs/2610.34633

SEMA-Auto, a multi-turn jailbreak attacker, efficiently adapts to victim responses, achieving record success rates with fewer interactions. This innovation challenges existing methods and elevates safety alignment in language models. https://arxiv.org/abs/2610.35743

Researchers have introduced an innovative AI training method based on a domain-specific Turing Test, where a weaker model imitates a stronger model’s outputs, enhancing capabilities without ground truths. It achieves notable improvements across various domains. https://arxiv.org/abs/2610.34781

Introducing MulFCoder: a multi-agent framework that improves front-end code generation from UI screenshots for React, Vue, and Angular. By encoding framework constraints, it enhances compilation success and reduces errors, making UI synthesis more reliable. https://arxiv.org/abs/2610.34418

A groundbreaking study unveils DELTA, a method to quantify ambiguity in laws and policies through interpretive disagreement among stakeholders, demonstrating that higher ambiguity does not guarantee political support or compromise. https://arxiv.org/abs/2610.35461

Hyperball, a novel optimizer wrapper, improves efficiency in pretraining by enforcing fixed weight norms, yielding a 20-30% speedup over AdamW. This method enhances training and supports better learning rate transfer across varied model architectures. https://arxiv.org/abs/2606.16899

MIT researchers unveil DMF-Gen, a technique that enhances reconstruction of physical fields from sparse measurements. This method reduces reconstruction errors and quantifies uncertainty, pushing data-driven modeling boundaries. https://arxiv.org/abs/2610.34709

RynnVLA-Latent uses 3D-grounded latent actions from diverse videos to scale robotic manipulation, reaching 98.8% success across benchmarks without native action labels. This method enhances robot policy learning and addresses the limitations of action-labeled data. https://arxiv.org/abs/2610.34356

A study shows independent review enhances coding agents' accuracy while maintaining costs, outperforming traditional verification. The research introduces new frameworks for optimizing agent harnesses, paving the way for smarter AI-driven software. https://arxiv.org/abs/2610.34577

RECAP reduces capability forgetting in reasoning models, boosting performance in complex tasks while preserving foundational skills. This dynamic strategy enhances general knowledge retention without extensive tuning, marking progress in reinforcement learning. https://arxiv.org/abs/2510.21978

A study shows that advanced language models often fail to credit the information shaping their answers. The new counterfactual training method, CREDIT, enhances models' abilities to disclose influences from external suggestions, improving AI accountability. https://arxiv.org/abs/2610.35295

A study introduces DynEdit, a dynamic framework enhancing knowledge editing in large language models by leveraging token entropy. This innovative approach improves efficiency on difficult content, achieving remarkable gains on long-form tasks. https://arxiv.org/abs/2610.34768

Research shows non-robust features in neural networks mislead models on clean images, yielding a "Jammed Set" where robust models excel. This challenges conventions through Spectral–Spatial Locking, revealing that jamming effects persist and redistribute. https://arxiv.org/abs/2610.34785

OvisOCR3 from Alibaba's ATH-MaaS is a compact 0.8B model unifying OCR tasks and achieving high benchmark scores. Its method combines document parsing, information extraction, and table reconstruction, excelling in a size that boosts data processing across fields. https://arxiv.org/abs/2610.33066

SceneAgent changes 3D and 4D scene conversion with an agentic pipeline, boosting simulations via reconstruction and physical modeling. Its F1 score of 60.75% beats traditional methods while fitting dynamics to dynamic scenes for better training and evaluation. https://arxiv.org/abs/2610.34745

A survey outlines a framework for uncertainty quantification in financial models, revealing research concentrates on workflow stages rather than linking evidence reliability to decision-making. This aims to enhance financial AI and drive risk-sensitive actions. https://arxiv.org/abs/2610.33109

A study reveals that number representation impacts how well large language models generalize in arithmetic tasks, showing a "grokking" delay influenced by numeral bases. Researchers aim to enhance AI efficiency in complex computations. https://arxiv.org/abs/2604.13082

Research indicates that embodied AI avatars, especially a beluga whale, effectively enhance sustainable product choices and pro-environmental attitudes. This innovative method showcases the value of nonhuman characters for impactful environmental interventions. https://arxiv.org/abs/2610.33191

MedVTok transforms medical AI with a unified visual tokenizer that integrates 2D and 3D images from diverse modalities, achieving top performance on 24 benchmarks. This method improves patient modeling while addressing anatomical coherence and clinical semantics. https://arxiv.org/abs/2610.35842

Intermittent and noisy gradient data can destabilize learning algorithms, even with optimal average performance. This research identifies critical feedback frequency thresholds, stressing the need for stable signals in biological and artificial neural networks. https://arxiv.org/abs/2610.35856

A study presents ARMD-Predict, a model using 7.5 million EHR culture-antibiotic observations to predict antibiotic resistance before culture results. This aims to enhance empiric therapy selection and improve patient outcomes amid rising antimicrobial resistance. https://arxiv.org/abs/2610.32998

The innovative SLIDE framework transforms item-level beam search outputs into a list-wise generation strategy, enhancing retrieval quality with a single greedy pass and outpacing traditional beam search performance. https://arxiv.org/abs/2610.34764

RAQ-MLA achieves 2.5 bits per element compression in Multi-Head Latent Attention caches, preserving reasoning and coding task performance without training. This enhances large language model deployment, addressing memory challenges in ultra-long-context inference. https://arxiv.org/abs/2610.33625

A study reveals finite abelian groups ensure successful generalization in neural networks, while random Latin squares hinder learning. This alters our view of "grokking" and task structures in machine learning, stressing task formulation's role in generalization. https://arxiv.org/abs/2610.35036

MTS-SRF converts coarse historical data into high-resolution predictions, improving forecasting in web telemetry and weather. Using TimeMagnifier, it reduces mean squared error by 12.3% on average, greatly enhancing predictive accuracy. https://arxiv.org/abs/2610.33470

Alibaba's World-in-Context redefines interactive video by merging view, action, and camera control into one framework. With a 93.8% perspective-switching success rate, it sets a new standard for immersive experiences. https://arxiv.org/abs/2610.33169

Revolutionizing context retention, RAQ-MLA shrinks key-value cache size by 6.5x while preserving accuracy, optimizing long-context inference. This method employs role-aware quantization, managing latent and positional keys for better efficiency in AI deployment. https://arxiv.org/abs/2610.33625

Alibaba's World-in-Context framework reshapes video modeling by merging view modes, actions, and camera controls into tokens, achieving notable perspective-switching rates. This enhances user immersion and responsiveness, raising standards for interactive video. https://arxiv.org/abs/2610.33169

A study finds that task structures influence neural networks' generalization, enhancing our understanding of "grokking." It shows finite abelian group tasks foster learning success, while random Latin squares hinder progress, altering machine learning optimization. https://arxiv.org/abs/2610.35036

Insights into grokking show finite abelian groups generalize from minimal observations, while random Latin squares hinder learning and limit generalization. This study presents a framework for understanding shifts from memorization to generalization in ML. https://arxiv.org/abs/2610.35036

Alibaba's World-in-Context framework revolutionizes interactive video generation by unifying view modes and camera controls. With a 93.8% perspective-switching success rate, it boosts user engagement in dynamic visual environments. https://arxiv.org/abs/2610.33169

Researchers unveiled RAQ-MLA, a new quantization method for Multi-Head Latent Attention caches that boosts KV-cache capacity by 6.5× while maintaining performance. This advancement enhances long-context language models, addressing critical challenges in AI. https://arxiv.org/abs/2610.33625

Researchers introduced RAQ-MLA to quantize Multi-Head Latent Attention caches, achieving a 6.5× capacity boost with near-lossless accuracy. By tailoring quantization to tensor roles, it enhances ultra-long-context language model efficiency without performance loss. https://arxiv.org/abs/2610.33625

MTS-SRF is a groundbreaking method that enables fine forecasting using coarse historical data, enhancing web service decisions. TimeMagnifier reduces prediction errors by 12.3%, greatly benefiting data-dependent industries. https://arxiv.org/abs/2610.33470

Alibaba's World-in-Context combines view modes, actions, and camera controls into a unified token system, enhancing navigation and perspective-switching. With a 93.8% success rate, it sets a new standard for interactive video generation using 100K clips. https://arxiv.org/abs/2610.33169

Researchers introduced RAQ-MLA, optimizing Multi-Head Latent Attention cache quantization for a 6.5× capacity increase and minimal performance loss in language models. https://arxiv.org/abs/2610.33625

World-in-Context unifies view, action, and camera controls in interactive video generation, achieving a 93.8% perspective-switching success rate. Its innovative In-Context Conditioning unveils new possibilities for dynamic storytelling in digital environments. https://arxiv.org/abs/2610.33169

Researchers introduce RAQ-MLA, a method for quantizing Multi-Head Latent Attention caches, achieving 6.5× capacity gain and near lossless performance. This method optimizes memory in large language models for efficient long-horizon inference without retraining. https://arxiv.org/abs/2610.33625