Artificial Intelligence research · Research Watch · Page 12

Preprint
Not peer reviewed

Multi-agent system targets more reliable automated generation of hardware RTL code

Original title: Veri-Sure: Multi-Agent RTL Code Generation with Temporal Tracing, Slicing and Formal Verification

A preprint presents Veri-Sure, a multi-agent framework combining tracing, slicing and formal checks to improve verified RTL code generation.

Jiale Liu et al. · Added 4 Nov 2026

Preprint
Not peer reviewed

Preprint describes single-pass method for routing prompts to the best LLM expert

Original title: ScoutRoute: Efficient Mixture-of-Experts Routing Using Internal Language Model Signals

A preprint reports a lightweight routing method that picks which language model should handle a prompt using one inexpensive scout model's internal signals.

Rajdeep Pandhare et al. · Added 4 Nov 2026

Preprint
Not peer reviewed

Researchers report smaller models can use more energy per task than larger ones

Original title: When Is Smaller Greener? Right-Sizing Language Models for Agentic AI

A preprint argues that comparing language models per call, not per completed task, can give the wrong green choice for AI agents.

Sajid Siraj et al. · Added 4 Nov 2026

Preprint
Not peer reviewed

Preprint proposes driving AI that checks if following instructions is actually safe

Original title: SafeDriveVLA: Navigation-Conditioned World Model Dreaming for Conflict-Aware End-to-End Autonomous Driving

Researchers report a driving model that imagines what would happen if it strictly obeyed an instruction, before deciding whether to follow it.

Shaoyuan Xie et al. · Added 4 Nov 2026

Preprint
Not peer reviewed

A preprint reports a dual small/large model memory system for long-context reasoning

Original title: MemDuet: Learning Dual-Process Memory for Long-Context Reasoning

Researchers describe MemDuet, pairing a small model for fast evidence selection with a large model for verification to speed up long-context reasoning.

Jiangcheng Song et al. · Added 4 Nov 2026

Preprint
Not peer reviewed

Researchers report pair-norm guided method cuts KV-cache bits in long-output LLMs

Original title: PN-Mix: Pair-Norm Guided Mixed-Precision Key Cache Quantization for Efficient Long-Output LLM Inference

A preprint reports a mixed-precision key-cache quantization scheme that uses query-key pair norms to preserve accuracy at very low bit-widths.

Gil Jae Song et al. · Added 4 Nov 2026

Preprint
Not peer reviewed
From Microsoft

Researchers report method to stop small language model agents getting stuck in loops

Original title: Progress-Aware Distillation for Mitigating Stagnation in Small Language Model Agents

A preprint proposes Progress-Aware Distillation, which the authors say reduces unproductive looping in small distilled agent models.

Added 4 Nov 2026

Preprint
Not peer reviewed
From AXA AI Research

Preprint profiles how input-token length drives LLM energy use on GPUs

Original title: The Cost of Context: Profiling the Energy Footprint of Input Tokens in Large Language Models

A preprint measures how context length and serving engines shape the energy cost of processing LLM input tokens.

Added 4 Nov 2026

Preprint
Not peer reviewed

Preprint adds token identity to gray-box hallucination detection in LLMs

Original title: Gray-Box Hallucination Detection in LLMs with Token Identities

A preprint reports that factoring in which tokens a model considered, not just their probabilities, improves hallucination detection accuracy.

Shaked Goldshtein et al. · Added 4 Nov 2026

Preprint
Not peer reviewed
From Tencent

Researchers report a trainable ‘draft mode’ that cuts reasoning-model token use

Original title: Draft-Thinking: Learning Efficient Reasoning in Long Chain-of-Thought LLMs

A preprint describes a training method that lets reasoning models produce shorter, self-adapted reasoning traces without losing much accuracy.

Added 4 Nov 2026

Preprint
Not peer reviewed

Researchers train AI agents to pre-organise documents for future questions

Original title: Librarian Agents: Training LLMs to Curate Long-Context Workspaces that Anticipate Future Queries

A preprint describes training an agent to restructure document collections in advance so another agent can answer queries more accurately later.

Anikait Singh et al. · Added 4 Nov 2026

Preprint
Not peer reviewed

Researchers report co-evolving attacker and defender agents to improve LLM tool-use safety

Original title: AgentSafeEvo: Adversarial Co-Evolution for LLM Agent Safety

A preprint describes training LLM agents against a co-evolving attacker that builds adversarial tool-use tasks to improve safety without hurting performance.

Chenxu Wang et al. · Added 4 Nov 2026

Preprint
Not peer reviewed
From Amazon

A preprint proposes letting LLMs decide when to retrieve help for translation

Original title: Retrieve What Helps, Only When It Helps: Agentic Reinforcement Learning for Selective Retrieval-Augmented Machine Translation

Researchers report a reinforcement-learning system that learns when and what to retrieve for machine translation, beating fixed retrieval strategies.

Added 4 Nov 2026

Preprint
Not peer reviewed

Researchers report naming the platform as principal weakens AI shopping agents’ bias against sponsored listings

Original title: Whom Do AI Agents Work For? Role Assignment Induces Sponsorship Bias in LLM Recommenders

A preprint finds that LLM shopping agents penalise sponsored listings far less when the system prompt names the platform rather than the consumer as their principal.

Davood Wadi et al. · Added 4 Nov 2026

Preprint
Not peer reviewed

A preprint reports calibrated probability-sharing helps AI agents collaborate better than sharing actions

Original title: Narrowing the Collaboration Gap, Probably

Researchers report that having AI agents share calibrated probability estimates, not just chosen actions, improves joint decision-making in maze-solving tests.

Mirah Shi et al. · Added 4 Nov 2026

Preprint
Not peer reviewed

Preprint reports training LLMs on junk tweets causes lasting cognitive decline

Original title: LLMs Can Get "Brain Rot": A Pilot Study on Twitter/X

A preprint finds continual pre-training on low-quality Twitter/X data degrades reasoning, safety and personality traits in large language models.

Shuo Xing et al. · Added 4 Nov 2026

Preprint
Not peer reviewed

Researchers map which design choices drive four LLM privacy attacks

Original title: On the Privacy of LLMs: An Ablation Study

A preprint finds LLM privacy risks differ by attack type, with dataset, scale and training factors each driving a different threat.

Karima Makhlouf et al. · Added 4 Nov 2026

Preprint
Not peer reviewed
From Tara Research

Preprint reports adversarial training method improves LLM honesty without hurting capability

Original title: CAT-alyzing Honesty: Continuous Adversarial Training for Transferable LLM Honesty

Researchers report a continuous adversarial training method that boosts honesty across three distinct test settings while preserving model performance.

Added 3 Nov 2026

Preprint
Not peer reviewed
From Tencent

Researchers report a scheduling fix cuts repeated-prefix waste in LLM RL training

Original title: PrefixFlow: Training-Time KV Caching for Schedule-Level Prefix Reuse in LLM RL Training

A preprint describes PrefixFlow, a training schedule that caches shared prompt prefixes once per rollout group instead of recomputing them for every trajectory.

Added 3 Nov 2026

Preprint
Not peer reviewed
From Toyota Motor Europe

A preprint reports client-tailored sub-networks for federated LLM fine-tuning

Original title: Optimal Hydra: Model Heterogeneous Federated Learning for LLMs using Data-Informed Hydra-Nets

Researchers propose Optimal Hydra, which builds client-specific sub-networks for federated fine-tuning of large language models under uneven resource limits.

Added 3 Nov 2026

688 entries