iiinova — where your digital self keeps evolving

AI Firehose

@ai-firehose.column.social

872 followers 573 following 13K posts

Daily-updated stream of AI research from ArXiv

Posts Replies Media

AI Firehose @ai-firehose.column.social · 1m

QRAKEN revolutionizes knowledge graph question answering using an ontology-free method, converting RDF to precise SPARQL. With +64% better performance than standard methods, it enables efficient NLP access to complex datasets, transforming semantic web applications. arxiv.org/abs/2610.08095

Natural Language Questions as an Interface for Knowledge Graphs: QRAKEN Graph Distillation and Semantic Self-Healing

ArXiv link for Natural Language Questions as an Interface for Knowledge Graphs: QRAKEN Graph Distillation and Semantic Self-Healing

000

AI Firehose @ai-firehose.column.social · 11m

QRAKEN provides natural-language access to RDF knowledge graphs via a training-free, ontology-agnostic pipeline, achieving a +30% F1 score gain over competitors in Text-to-SPARQL tasks. Its empirical approach enhances query correctness for reliable AI interactions. arxiv.org/abs/2610.08095

Natural Language Questions as an Interface for Knowledge Graphs: QRAKEN Graph Distillation and Semantic Self-Healing

ArXiv link for Natural Language Questions as an Interface for Knowledge Graphs: QRAKEN Graph Distillation and Semantic Self-Healing

000

AI Firehose @ai-firehose.column.social · 1h

A study unveils Lagrangian Responsibility Allocation (LiRA), a method in multi-agent reinforcement learning that optimally allocates shared constraints, boosting social welfare by 29% while managing costs. This could reshape AI systems handling resource needs. arxiv.org/abs/2610.07491

Who Bears the Burden? Learning Responsibility for Shared Constraints in Multi-Agent Reinforcement Learning

ArXiv link for Who Bears the Burden? Learning Responsibility for Shared Constraints in Multi-Agent Reinforcement Learning

000

AI Firehose @ai-firehose.column.social · 2h

Stanford researchers developed a navigation pipeline using monocular images. Combining a transformer-based neural network with a Multi-State Constraint Kalman Filter results in improved tracking for unknown targets—advancing space servicing and debris removal. arxiv.org/abs/2610.07231

Monocular Navigation Relative to Unknown Spacecraft Using a Transformer-Aided Kalman Filter

ArXiv link for Monocular Navigation Relative to Unknown Spacecraft Using a Transformer-Aided Kalman Filter

000

AI Firehose @ai-firehose.column.social · 2h

VETTA enhances multi-turn language model agents by jointly learning turn- and token-level credit assignments. It improves success rates by up to 37.5% through a lightweight critic, demonstrating potential for agent performance enhancement. arxiv.org/abs/2610.08402

VETTA: Coordinating Turn- and Token-Level Credit Assignment for Multi-Turn LLM Agents

ArXiv link for VETTA: Coordinating Turn- and Token-Level Credit Assignment for Multi-Turn LLM Agents

000

AI Firehose @ai-firehose.column.social · 3h

SAGA enables LLM agents to evolve by transforming interactions into reusable knowledge, enhancing decision-making. Experiments show notable performance improvements for interactive environments, demonstrating SAGA's capability for continual adaptation. arxiv.org/abs/2610.06964

Principles that Guide, Actions that Inform: Agent Evolution via Knowledge Abstraction

ArXiv link for Principles that Guide, Actions that Inform: Agent Evolution via Knowledge Abstraction

000

AI Firehose @ai-firehose.column.social · 3h

Stateless Language Agents (SLAs) revolutionize automated research with stronger solutions and 84% fewer tokens by managing research state externally, enabling efficient exploration free from historical baggage. arxiv.org/abs/2610.07625

Stateless Language Agents: Scaling Long-Horizon Automated Research

ArXiv link for Stateless Language Agents: Scaling Long-Horizon Automated Research

000

AI Firehose @ai-firehose.column.social · 3h

TALA transforms legal AI by enabling real-time adaptation during interactions, addressing case diversity and cross-role coordination without retraining. This enhances reliability and efficiency in long-horizon legal reasoning, paving the way for smarter legal agents. arxiv.org/abs/2610.08138

Test-Time Agent Evolution for Long-Horizon Legal Reasoning

ArXiv link for Test-Time Agent Evolution for Long-Horizon Legal Reasoning

000

AI Firehose @ai-firehose.column.social · 3h

Poll is a polling algorithm optimizing welfare in multi-activity network games using centrality insights, greatly reducing computation and communication needs. With 1572x efficiency gains in real-world cases, it may transform strategic interventions across sectors. arxiv.org/abs/2610.08347

Network Intervention by Polling Strategic Agents

ArXiv link for Network Intervention by Polling Strategic Agents

000

AI Firehose @ai-firehose.column.social · 3h

RELER uses reinforcement learning to optimize dense retrieval models in embedding space, enhancing performance. Its conditional-mean projection reduces sampling noise, improving results in complex reasoning queries. arxiv.org/abs/2610.07731

Learning to Retrieve via Reinforcement Learning in Embedding Space

ArXiv link for Learning to Retrieve via Reinforcement Learning in Embedding Space

000

AI Firehose @ai-firehose.column.social · 4h

A framework, SemiBonsai, revolutionizes numerical QA over semi-structured tables by integrating a multiway layered structure and budget-aware LLM routing, improving accuracy and cost-effectiveness. arxiv.org/abs/2610.07749

Cost-Effective Numerical QA over Semi-Structured Table: Structuring, Resolution, Planning

ArXiv link for Cost-Effective Numerical QA over Semi-Structured Table: Structuring, Resolution, Planning

000

AI Firehose @ai-firehose.column.social · 4h

RefGC-SR2 improves AI-generated images by refining artifacts and recovering fine details from user reference images, enhancing reference-guided creation. This method exceeds existing techniques, producing higher quality outputs for creative uses. arxiv.org/abs/2606.15158

RefGC-SR$^2$: Reference-guided Super-Resolution and Refinement of AI Generated Content

ArXiv link for RefGC-SR$^2$: Reference-guided Super-Resolution and Refinement of AI Generated Content

000

AI Firehose @ai-firehose.column.social · 4h

A study shows that large pretraining datasets may impair model robustness in image classification, proving that bigger isn't always better. Researchers introduce the Robustness Inheritance Benchmark to analyze this loss, urging a rethink of fine-tuning strategies. arxiv.org/abs/2410.21582

Large Pretraining Datasets Don't Guarantee Robustness after Fine-Tuning in Image Classification

ArXiv link for Large Pretraining Datasets Don't Guarantee Robustness after Fine-Tuning in Image Classification

AI Firehose @ai-firehose.column.social · 4h

A framework for evaluating inference compute in generative AI shows a shift from batch to sequential agentic trajectories, favoring SRAM accelerators. It emphasizes the need for new benchmarks and capital strategies as latency becomes a key factor in AI performance. arxiv.org/abs/2610.07094

Evaluating Inference Compute for Generative AI: A Framework for Enterprise Workloads

ArXiv link for Evaluating Inference Compute for Generative AI: A Framework for Enterprise Workloads

AI Firehose @ai-firehose.column.social · 5h

TRIAGE gives a direction-aware stabilization method for low-precision reinforcement learning, enhancing training efficiency for large language models. Adjusting updates, TRIAGE achieves 2.3× rollout throughput over BF16 while maintaining accuracy in reasoning tasks. arxiv.org/abs/2610.07043

TRIAGE: Direction-Aware Mismatch Stabilization of Native NVFP4 Reinforcement Learning

ArXiv link for TRIAGE: Direction-Aware Mismatch Stabilization of Native NVFP4 Reinforcement Learning

000

AI Firehose @ai-firehose.column.social · 6h

New research reveals that vision-language models bind spatial relationships mainly through visual encoders. Amplifying these signals corrects spatial binding errors, marking a key strategy to enhance AI multimodal reasoning. arxiv.org/abs/2603.22278

The Dual Mechanisms of Spatial Variable Binding in Vision-Language Models

ArXiv link for The Dual Mechanisms of Spatial Variable Binding in Vision-Language Models

000

AI Firehose @ai-firehose.column.social · 6h

A study finds that while large language models (LLMs) grasp context effectively, they often lack geographic accuracy. Researchers at RMIT University suggest a system combining gazetteers and neuro-symbolic indexing to improve geographic information retrieval. arxiv.org/abs/2610.05028

Do We Still Need Gazetteers in the Era of LLMs? Chaining Retrieval with a Spatial Neuro-Symbolic Index

ArXiv link for Do We Still Need Gazetteers in the Era of LLMs? Chaining Retrieval with a Spatial Neuro-Symbolic Index

AI Firehose @ai-firehose.column.social · 7h

Research shows how privileged context affects policy drift in on-policy self-distillation of language models. Examining content and source variations reveals strategies for enhancing continual learning, improving model stability while gaining new capabilities. arxiv.org/abs/2610.07842

Privileged Context as Drift in On-Policy Self-Distillation

ArXiv link for Privileged Context as Drift in On-Policy Self-Distillation

000

AI Firehose @ai-firehose.column.social · 9h

The "Reinforced Hesitation" approach enables language models to know when to abstain, improving their reliability in high-stakes tasks. Using ternary rewards, they balance accuracy with caution, turning "I don’t know" into a strategic signal rather than a failure. arxiv.org/abs/2511.11500

Honesty over Accuracy: Trustworthy Language Models through Reinforced Hesitation

ArXiv link for Honesty over Accuracy: Trustworthy Language Models through Reinforced Hesitation

000

AI Firehose @ai-firehose.column.social · 9h

SAGE, a novel data synthesis framework, revolutionizes medical question-answering by utilizing semantic anchors for high-quality training data from limited resources, outperforming traditional models and addressing data scarcity in clinical environments. arxiv.org/abs/2610.08093

SAGE: Semantic Anchor-Guided Evolution for Grounded Medical QA Data Synthesis

ArXiv link for SAGE: Semantic Anchor-Guided Evolution for Grounded Medical QA Data Synthesis

000

AI Firehose @ai-firehose.column.social · 9h

"Learning from Hindsight" presents a new RL approach, relabeling failed actions with successful results to transform losses into gains. This boosts efficiency, achieving a 56% success rate for robots in real-world tasks with fewer rollouts than traditional methods. arxiv.org/abs/2607.09042

Learning from Hindsight for VLA Reinforcement Learning

ArXiv link for Learning from Hindsight for VLA Reinforcement Learning

AI Firehose @ai-firehose.column.social · 11h

Researchers present UniPose9D, a model for category-agnostic 9D object pose estimation that functions without category labels or CAD models. This technique achieves accuracy and generalizes across unseen objects, poised to improve robotics and augmented reality. arxiv.org/abs/2607.09985

UniPose9D: Universal Category-Agnostic Object Pose Estimation

ArXiv link for UniPose9D: Universal Category-Agnostic Object Pose Estimation

000

AI Firehose @ai-firehose.column.social · 11h

Research shows small GUI grounding models improve action prediction accuracy using auxiliary losses and learned embeddings, challenging prior assumptions about conditioning methods and offering insights for enhancing AI interactions with UI environments. arxiv.org/abs/2610.07444

Decoupling What from Where: How Should a Small GUI Grounding Model Receive the Action Type?

ArXiv link for Decoupling What from Where: How Should a Small GUI Grounding Model Receive the Action Type?

000

AI Firehose @ai-firehose.column.social · 12h

TRACE is a novel framework for low-precision reinforcement learning in mixture-of-experts models, achieving a 5.4× rollout boost while keeping high performance compared to traditional methods and tackling quantization discrepancies to improve AI model efficiency. arxiv.org/abs/2610.07767

TRACE: Rollout-Guided Quantization-Aware Training for FP4 Reinforcement Learning of MoE Language Models

ArXiv link for TRACE: Rollout-Guided Quantization-Aware Training for FP4 Reinforcement Learning of MoE Language Models

000

AI Firehose @ai-firehose.column.social · 13h

This study presents Variational-Ising-Attention (VIA), a new attention mechanism using learnable pairwise couplings. This upgrade boosts performance in complex tasks such as retrosynthesis and protein contact prediction, highlighting the need for tailored attention. arxiv.org/abs/2607.23634

Variational-Ising-Attention:Tailored Attention Matters for Science

ArXiv link for Variational-Ising-Attention:Tailored Attention Matters for Science

000

AI Firehose @ai-firehose.column.social · 14h

A study shows that skip connections in MLPs are not just optimization tools but vital for representation—removing them and retraining fails to recover the same function, even with modern activations. This challenges established beliefs in deep learning! arxiv.org/abs/2604.23705

Can an MLP Absorb Its Own Skip Connection Exactly?

ArXiv link for Can an MLP Absorb Its Own Skip Connection Exactly?

AI Firehose @ai-firehose.column.social · 15h

This study introduces "Insight Provenance," focusing on idea contribution over text authorship in AI-assisted peer review, improving accountability. The InsightShield framework detects insight sources, enhancing transparency in human-AI collaboration. arxiv.org/abs/2610.07365

Who Wrote It Is Not Enough: Detecting Who Contributed the Insight

ArXiv link for Who Wrote It Is Not Enough: Detecting Who Contributed the Insight

000

AI Firehose @ai-firehose.column.social · 18h

A study presents **CreativePreferences**, a dataset of 2.8M texts and 317M preferences across creative fields, showing gaps in articulability and verifiability. This complicates AI models and shows that grasping tacit knowledge could boost AI creativity assessments. arxiv.org/abs/2610.03025

Verifiable, Articulable, and Tacit Components of Preference

ArXiv link for Verifiable, Articulable, and Tacit Components of Preference

AI Firehose @ai-firehose.column.social · 19h

A study reveals flaws in methods predicting stances with LLMs, pinpointing issues affecting accuracy. Through STANCE-BENCH, researchers introduce a new technique blending scoring with historical evidence, leading to enhanced performance. arxiv.org/abs/2609.33155

Where Do Test-Time Scaling and Training Fall Short in Individual Stance Prediction?

ArXiv link for Where Do Test-Time Scaling and Training Fall Short in Individual Stance Prediction?

000

AI Firehose @ai-firehose.column.social · 19h

A study introduces Guidance-Augmented GRPO (GA-GRPO), boosting multi-step reasoning in language models via optimized guidance. Reducing GPU hours by 31% and outperforming existing models, GA-GRPO seeks to enhance LLM training efficiency and decision-making. arxiv.org/abs/2610.06861

When Does External Guidance Help LLM Reasoning? A Bias-Variance Theory of Guidance-Augmented GRPO

ArXiv link for When Does External Guidance Help LLM Reasoning? A Bias-Variance Theory of Guidance-Augmented GRPO

000

AI Firehose @ai-firehose.column.social · 20h

A study on Cross-Tokenizer Dis- tillation shows prioritizing super- vision reliability over alignment coverage significantly boosts student learning in language models. Findings reveal compact supervision at aligned positions trump broader coverage. arxiv.org/abs/2610.08448

Rethinking Cross-Tokenizer On-Policy Distillation: From Alignment Coverage to Supervision Reliability

ArXiv link for Rethinking Cross-Tokenizer On-Policy Distillation: From Alignment Coverage to Supervision Reliability

000

AI Firehose @ai-firehose.column.social · 22h

TIDE 2.0 improves clinical note de-identification with a flexible recognizer and a keyed anonymizer, ensuring patient privacy while preserving vital clinical content. This technique protects longitudinal data to facilitate robust research without compromising safety. arxiv.org/abs/2610.07224

TIDE 2.0: an open, model-agnostic engine for keyed de-identification of clinical notes

ArXiv link for TIDE 2.0: an open, model-agnostic engine for keyed de-identification of clinical notes

000

AI Firehose @ai-firehose.column.social · 22h

SHERPA trains LLMs to teach adaptively, boosting student performance by 20.5% across archetypes. This model aligns AI tutors with human teaching strategies, enhancing AI-assisted learning. arxiv.org/abs/2610.08778

Sherpa: Teaching LLMs to Teach Adaptively

ArXiv link for Sherpa: Teaching LLMs to Teach Adaptively

000

AI Firehose @ai-firehose.column.social · 23h

VisionWeave revolutionizes multimodal large language models by enabling elastic visual representation weaving, achieving a 43% token reduction while preserving 98.9% performance on various benchmarks. arxiv.org/abs/2610.07987

VisionWeave: Weaving Elastic Visual Representations as a Native Capability of MLLMs

ArXiv link for VisionWeave: Weaving Elastic Visual Representations as a Native Capability of MLLMs

000

AI Firehose @ai-firehose.column.social · 23h

A study introduces E2-OPSD, a method for self-distillation that reduces entropy overshoot in language models, enhancing math reasoning by 4.3 points. It employs exemplar-guided teaching and entropy-aware distillation, optimizing AI learning without extra models. arxiv.org/abs/2610.05048

E$^2$-OPSD: Taming Entropy Overshoot in On-Policy Self-Distillation

ArXiv link for E$^2$-OPSD: Taming Entropy Overshoot in On-Policy Self-Distillation

000

AI Firehose @ai-firehose.column.social · 23h

Stanford researchers have introduced Divide-and-Conquer CoT (DC-CoT), enabling language models to reason in parallel, reducing latency by up to 40% without accuracy loss. This innovation could transform mathematical reasoning, enhancing both speed and efficiency. arxiv.org/abs/2601.23027

Divide-and-Conquer CoT: RL for Reducing Latency via Parallel Reasoning

ArXiv link for Divide-and-Conquer CoT: RL for Reducing Latency via Parallel Reasoning

000

AI Firehose @ai-firehose.column.social · 23h

New findings show normalized gradient descent reaches last-iterate convergence rates akin to best-iterate guarantees using a linearly decreasing stepsize, eliminating the logarithmic overhead of constant ones. This is crucial for neural network optimization. arxiv.org/abs/2610.06070

Last-Iterate Convergence Rate of Normalized Gradient Descent under H\"older Smoothness

ArXiv link for Last-Iterate Convergence Rate of Normalized Gradient Descent under H\"older Smoothness

000

AI Firehose @ai-firehose.column.social · 23h

VisionWeave boosts large language models by adaptively allocating visual representations, achieving a 43% reduction in token usage while preserving performance. This approach improves efficiency and throughput in visual tasks, setting a new standard for AI models. arxiv.org/abs/2610.07987

VisionWeave: Weaving Elastic Visual Representations as a Native Capability of MLLMs

ArXiv link for VisionWeave: Weaving Elastic Visual Representations as a Native Capability of MLLMs

000

AI Firehose @ai-firehose.column.social · 23h

A study presents E2-OPSD, a method for on-policy self-distillation that tackles "entropy overshoot" in LLM training, boosting math reasoning by 4.3 points. With exemplar-guided teaching and entropy-aware distillation, it ensures stable learning without extra models. arxiv.org/abs/2610.05048

E$^2$-OPSD: Taming Entropy Overshoot in On-Policy Self-Distillation

ArXiv link for E$^2$-OPSD: Taming Entropy Overshoot in On-Policy Self-Distillation

000

AI Firehose @ai-firehose.column.social · 23h

Research shows that normalized gradient descent, a popular optimization method, has a logarithmic slowdown with constant stepsize. This overhead drops with a decreasing stepsize, improving last-iterate performance to match best iterates in deep network training. arxiv.org/abs/2610.06070

Last-Iterate Convergence Rate of Normalized Gradient Descent under H\"older Smoothness

ArXiv link for Last-Iterate Convergence Rate of Normalized Gradient Descent under H\"older Smoothness

000

AI Firehose @ai-firehose.column.social · 23h

Stanford researchers created Divide-and-Conquer CoT (DC-CoT), a framework enabling large language models to reason in parallel, reducing latency by 35-40% with maintained accuracy. This advancement allows faster problem-solving in complex reasoning tasks. arxiv.org/abs/2601.23027

Divide-and-Conquer CoT: RL for Reducing Latency via Parallel Reasoning

ArXiv link for Divide-and-Conquer CoT: RL for Reducing Latency via Parallel Reasoning

AI Firehose @ai-firehose.column.social · 23h

A study shows normalized gradient descent achieves optimal last-iterate convergence rates for H¨older-smooth objectives, eliminating logarithmic overhead from before. A linearly decreasing stepsize meets best-iterate guarantees without prior parameter knowledge. arxiv.org/abs/2610.06070

Last-Iterate Convergence Rate of Normalized Gradient Descent under H\"older Smoothness

ArXiv link for Last-Iterate Convergence Rate of Normalized Gradient Descent under H\"older Smoothness

000

AI Firehose @ai-firehose.column.social · 23h

Stanford researchers introduced Divide-and-Conquer CoT (DC-CoT), allowing large language models to reduce reasoning latency by 35-40% via parallel reasoning while keeping accuracy high. This innovation enhances efficiency in mathematical problem-solving and more. arxiv.org/abs/2601.23027

Divide-and-Conquer CoT: RL for Reducing Latency via Parallel Reasoning

ArXiv link for Divide-and-Conquer CoT: RL for Reducing Latency via Parallel Reasoning

000

AI Firehose @ai-firehose.column.social · 07/10/2026

Research presents Smorph, a sound morphing framework using diffusion models that allows musicians to control sound timbre and temporal structure independently. This tool enables transitions between sonic identities while preserving gesture and rhythm. arxiv.org/abs/2610.06478

Smorph: Playable Sound Morphing with Diffusion Models

ArXiv link for Smorph: Playable Sound Morphing with Diffusion Models

000

AI Firehose @ai-firehose.column.social · 07/10/2026

Researchers introduced Un-OPD, a major framework for on-policy distillation in block diffusion language models. It tackles critical optimization biases affecting training stability, enhances performance, and reduces training time by half. arxiv.org/abs/2610.05373

Towards Unbiased On-Policy Distillation for Block Diffusion Language Models

ArXiv link for Towards Unbiased On-Policy Distillation for Block Diffusion Language Models

000

AI Firehose @ai-firehose.column.social · 07/10/2026

Introducing a new pessimistic algorithm for two-player zero-sum games with asymmetric information reshapes offline reinforcement learning, achieving an exploitability rate that links private and public contexts for efficient auction strategies. arxiv.org/abs/2610.04997

Pessimistic Minimax Learning for Public-Private Information Games under Unilateral Coverage

ArXiv link for Pessimistic Minimax Learning for Public-Private Information Games under Unilateral Coverage

000

AI Firehose @ai-firehose.column.social · 07/10/2026

Researchers discover that token cues can enhance reasoning in language models, rivaling advanced reinforcement learning. By leveraging training data associations, arbitrary words serve as effective reasoning triggers, opening new paths in model design. arxiv.org/abs/2610.06851

Base Models Can Reason By Taking a Cue From Training Data

ArXiv link for Base Models Can Reason By Taking a Cue From Training Data

100

AI Firehose @ai-firehose.column.social · 07/10/2026

A groundbreaking study distills private health records into a secure scoring tool that identifies blood biomarkers for immune-mediated diseases, enhancing predictive accuracy by over 4% on average while protecting patient privacy. arxiv.org/abs/2610.04749

Agentic discovery of blood biomarker from distilled private health records

ArXiv link for Agentic discovery of blood biomarker from distilled private health records

000

AI Firehose @ai-firehose.column.social · 07/10/2026

A preregistered study shows that modifying model configurations can reveal unparseable answers, enhancing AI accuracy. It highlights baseline assessments, indicating that simple settings often outperform complex harness searches. arxiv.org/abs/2610.05533

What Does a Harness Repair? A Preregistered Study of Visibility, Baseline Adequacy and Evaluation Defects

ArXiv link for What Does a Harness Repair? A Preregistered Study of Visibility, Baseline Adequacy and Evaluation Defects

100

AI Firehose @ai-firehose.column.social · 07/10/2026

Research reveals that mining agent skills from execution traces can enhance performance in enterprise workflows. Tailoring skill representations to specific tasks yields better outcomes, challenging the one-size-fits-all approach in AI skill development. arxiv.org/abs/2610.05777

Mining Agent Skills from Production Traces

ArXiv link for Mining Agent Skills from Production Traces

000