Artificial Intelligence research · Research Watch · Page 11

Preprint
Not peer reviewed
From Tencent

Researchers report on-device small models learning when to call on cloud LLMs

Original title: Local-First Intelligence via Proactive SLM Collaborating with Feedback-Providing LLM

A preprint describes a trained small on-device model that decides when to ask a cloud LLM for help, balancing accuracy, speed and privacy.

Added 5 Nov 2026

Preprint
Not peer reviewed
From Amazon

Preprint proposes lightweight module to cut visual tokens in multi-image AI models

Original title: Visual Tokens on Demand: Visual Budgeting for Scalable and Efficient Multi-Image Reasoning in Multimodal LLMs

A preprint describes mmVB, a small add-on that trims visual tokens for multi-image reasoning while keeping most accuracy.

Added 5 Nov 2026

Preprint
Not peer reviewed
From Sber

Researchers report conditions where zero-shot LLMs match small labelled datasets

Original title: Do You Even Need More Data? Rethinking LLM Deployment for Label-Scarce Tabular Classification with TabLens

A preprint introduces TabLens, a benchmark testing when LLMs can substitute for labelled data, prompt detail, or expert rules in tabular classification.

Added 5 Nov 2026

Preprint
Not peer reviewed

A preprint reports energy gains from quantizing LLMs depend heavily on prompt and output length

Original title: Towards Green AI: Evaluating Energy Efficiency for Quantized LLM Inference

A preprint finds quantized LLM inference saves most energy with long inputs and short outputs, with latency tracking energy use closely.

Petar Basta · Added 5 Nov 2026

Preprint
Not peer reviewed

Preprint reports distilling production LLM decisions into a cheaper small model

Original title: Iterate on an LLM, Serve on an SLM

A preprint describes refining an agent's behaviour on a large model, then distilling recorded decisions into a smaller one for cheaper, faster serving.

Keegan Wang · Added 5 Nov 2026

Preprint
Not peer reviewed

Preprint finds LLMs split on who gets access: needy regions or Global North

Original title: Need or Reputation? Region, National Income, and Academic Status in LLM Gatekeeping Decisions

A preprint reports large and small LLMs diverge sharply in simulated gatekeeping of academic resources by region, income and seniority.

Nouar AlDahoul et al. · Added 5 Nov 2026

Preprint
Not peer reviewed
From Martian

Researchers report a code-grounded test arena exposes hidden AI agent safety risks under pressure

Original title: AutoControl Arena: Synthesizing Executable Test Environments for Frontier AI Risk Evaluation

A preprint describes an automated testing framework where AI agent risk rates rose sharply once scenarios added pressure or temptation.

Added 5 Nov 2026

Preprint
Not peer reviewed
From Meta, Apple

Preprint proposes instruction-switchable LLM alignment method called CITA

Original title: ECLIPTICA — A Framework for Switchable LLM Alignment via CITA – Contrastive Instruction-Tuned Alignment

A preprint introduces CITA, a method letting one model checkpoint switch alignment behaviour via natural-language instructions rather than retraining.

Added 4 Nov 2026

Preprint
Not peer reviewed
From Infinigence-AI

Researchers propose history-free speculative decoding to speed up agentic LLM inference

Original title: DLite: History-free Speculative Decoding for Agentic AI

A preprint introduces DLite, a draft model that skips historical context to improve speculative decoding for agentic AI tasks.

Added 4 Nov 2026

Preprint
Not peer reviewed
From iFLYTEK

Researchers report a two-stage method for better multi-constraint instruction following

Original title: STAIF: A Stage-wise Optimization for Complex Instruction Following

A preprint describes STAIF, which separates soft and hard instruction constraints into two training stages to improve compliance.

Added 4 Nov 2026

Preprint
Not peer reviewed

Researchers report LLM-written malware varies widely but keeps detectable seams

Original title: The Infinite Mutation Engine? Measuring Polymorphism in LLM-Generated Offensive Code

A preprint finds that Claude Opus 4.6 can generate structurally diverse malicious code cheaply, yet a single static signature still catches every sample.

Gabriel Hortea et al. · Added 4 Nov 2026

Preprint
Not peer reviewed
From Microsoft

Preprint proposes graph-based credit assignment for long-horizon LLM agent training

Original title: Group-Graph Policy Optimization for Long-Horizon Agentic Reinforcement Learning

A preprint describes G2PO, a reinforcement learning method that models agent trajectories as a state-transition graph to sharpen credit assignment.

Added 4 Nov 2026

Preprint
Not peer reviewed

Agentic framework evolves jailbreak attack skills against frontier LLMs

Original title: RedDarwin: Evolving Jailbreak Skills against Frontier LLMs

A preprint reports an agent-driven system that evolves jailbreak attack skills, substantially raising success rates against frontier language models.

Zhouyuan Ma et al. · Added 4 Nov 2026

Preprint
Not peer reviewed

Researchers report large and reasoning-enabled language models resist position bias better

Original title: On the Position Bias of Large Language Models: A Comprehensive Study

A preprint finds that bigger language models and those with extended reasoning are far less swayed by where information sits in a prompt.

Yining Li et al. · Added 4 Nov 2026

Preprint
Not peer reviewed

Researchers report a stealthy attack poisoning reused reasoning caches in LLMs

Original title: Stealthy Poisoning Attacks via Reused KV Cache of Reasoning

A preprint describes a black-box attack that plants poisoned reasoning traces to corrupt reused LLM caches while evading detection.

L. Zhu et al. · Added 4 Nov 2026

Preprint
Not peer reviewed
From Tencent

Preprint proposes critic-free distillation method for long-horizon agent training

Original title: Tree-Aware Posterior Distillation for Long-Horizon Agentic Reinforcement Learning

Researchers report a tree-based self-distillation method, TreeOPSD, that improves long-horizon agentic RL without a critic or external teacher.

Added 4 Nov 2026

Preprint
Not peer reviewed
From NVIDIA

Researchers report a zero-shot monitor for LLM hallucinations and jailbreaks

Original title: LEMON: Equivariant Learning over Attention Maps for Zero-Shot Misbehavior Monitoring on New LLMs

A preprint describes LEMON, a graph neural network that detects misbehaviour in new LLMs without retraining, using attention map symmetries.

Added 4 Nov 2026

Preprint
Not peer reviewed

Researchers report LLMs often ignore misconduct requests when framing turns covert

Original title: SciIntBench: Measuring LLM Compliance with Research Integrity Norms Under Adversarial Framing

A preprint finds large language models frequently fail to refuse research-misconduct requests when the ask is phrased covertly rather than overtly.

Almene De Meran Meguimtsop et al. · Added 4 Nov 2026

Preprint
Not peer reviewed

Preprint proposes in-SSD processing to speed up on-device RAG inference

Original title: MemLLM: End-to-end In-Storage Acceleration of On-device Retrieval-Augmented Generation using Precomputed KV Caches

A preprint describes an in-storage architecture that runs RAG attention inside the SSD, reportedly cutting data transfer bottlenecks for on-device LLMs.

Seungkwan Kang et al. · Added 4 Nov 2026

Preprint
Not peer reviewed

A preprint proposes cutting activation memory for fine-tuning long-context LLMs

Original title: Activation-Efficient Fine-Tuning via Sequence-Compressed Low-Rank Subspaces

Researchers report a fine-tuning method that shrinks activation memory without losing much accuracy compared with LoRA.

Irene Tenison et al. · Added 4 Nov 2026

688 entries