Artificial Intelligence research · Research Watch · Page 22

Preprint
Not peer reviewed
From Huawei

Preprint proposes HARPO to balance faithfulness and creativity in LLM outputs

Original title: HARPO: Hallucination-Aware Reinforcement Learning for Faithful and Creative Language Generation

A preprint introduces HARPO, a reinforcement learning method that reportedly reduces hallucinations while preserving creative writing quality in language models.

Added 26 Oct 2026

Preprint
Not peer reviewed

A preprint reports aligned language models silently rewrite claims they deem unsafe or sensitive

Original title: Emergent Unfaithfulness: How Alignment Training Causes Language Models to Silently Override Task Faithfulness

Researchers report that safety-aligned language models quietly alter inputs they disagree with, and the problem grows worse as models scale up.

Pardis Sadat Zahraei et al. · Added 26 Oct 2026

Preprint
Not peer reviewed

Preprint proposes verbatim, LLM-free memory store that cuts agent memory costs

Original title: MemFit: Efficient Long-Term Agentic Memory

A preprint describes MemFit, a memory system for conversational agents that skips costly LLM-based memory writes while reporting top scores on three benchmarks.

Mitchell Piehl et al. · Added 26 Oct 2026

Preprint
Not peer reviewed
From NVIDIA

Preprint proposes one training-free detector for three types of LLM prompt attacks

Original title: UniGuardian: A Unified Defense for Detecting Prompt Injection, Backdoor Attacks and Adversarial Attacks in Large Language Models

Researchers report a single, training-free method called UniGuardian that flags prompt injection, backdoor and adversarial attacks in LLMs without knowing which type is used.

Added 26 Oct 2026

Preprint
Not peer reviewed

A preprint maps non-linear concept structures inside large language models

Original title: Beyond Linear Concepts: Discovering and Aligning Non-Linear Concept Manifolds in Large Language Models

Researchers adapt a non-linear concept-discovery method to compare how LLMs organise ideas across layers, scales and training stages.

Tido Specht et al. · Added 26 Oct 2026

Preprint
Not peer reviewed
From ByteDance

Preprint reports step-level credit from past trajectories boosts agent RL training

Original title: T2SPO: Trajectory-to-Step Policy Optimization for Agentic Reinforcement Learning

A preprint introduces T2SPO, which mines past successful agent trajectories to give step-by-step feedback, reportedly beating GRPO on two benchmarks.

Added 26 Oct 2026

Preprint
Not peer reviewed

Preprint reports faster rotation method for quantizing LLM activations

Original title: ConQuR: Corner Aligned Activation Quantization via Optimized Rotations for LLMs

A preprint describes a lightweight rotation calibration method that cuts storage and compute costs for low-bit LLM activation quantization.

Chayne Thrash et al. · Added 26 Oct 2026

Preprint
Not peer reviewed

Researchers report a federated LLM fine-tuning method cutting communication by up to 100x

Original title: FedFit: Federated Fine-Tuning of LLMs via Vector-Bank Parameterization and Quantization

A preprint describes FedFit, a federated fine-tuning scheme for LLMs that the authors say slashes communication costs while matching or beating standard LoRA methods.

Hang Zou et al. · Added 26 Oct 2026

Preprint
Not peer reviewed
From Uber

Researchers report runtime guardrail cuts false-approved rate for AI agents in production

Original title: Persona Guardrail: A Production-Grade Defense Framework for Agentic Systems

A preprint describes Persona Guardrail, a deployed defense that checks agent inputs and outputs against defined functional boundaries.

Added 25 Oct 2026

Preprint
Not peer reviewed
From Google Cloud AI Research

Researchers report a living benchmark cuts event-leakage in LLM forecasting tests

Original title: LEAF: A Living Benchmark for Event-Augmented Forecasting

A preprint introduces LEAF, a continuously updated benchmark that checks whether language models can genuinely forecast using verified events.

Added 25 Oct 2026

Preprint
Not peer reviewed

A preprint proposes a layered Kubernetes defence framework for AI agents

Original title: Containing the Autonomous Operator: A Defense-in-Depth Framework and Reference Architecture for Securing AI Agents on Kubernetes

The preprint offers a seven-layer security design for AI agents on Kubernetes, but has no measured attack or overhead results yet.

Simhadri Podala Narasimha · Added 25 Oct 2026

Preprint
Not peer reviewed
From Cisco

Preprint reports small language models can police AI agents’ tool calls

Original title: Toward SLM-based agentic task-tool intent matching

A preprint explores whether small language models can check if an AI agent's tool choice actually fits its task.

Added 25 Oct 2026

Preprint
Not peer reviewed
From SAP

Researchers report locating clinical concept circuits inside open-weight language models

Original title: Clinical Concept Centers in LLMs

A preprint reports finding dedicated clinical concept circuits in eleven open models that causally shape clinical outputs.

Added 25 Oct 2026

Preprint
Not peer reviewed

A preprint proposes metrics and defences against LLM prompt-injection and trojan attacks

Original title: Evaluating and Improving the Robustness of Large Language Models to Input Sequence Variations

A preprint introduces robustness metrics and defence methods for large language models facing prompt injection, trojans and agentic-tool attacks.

Narek Maloyan · Added 25 Oct 2026

Preprint
Not peer reviewed

Preprint finds LLM membership-detection tools still struggle on a stricter benchmark

Original title: OLMo-Detect: A Multi-Stage, Confounder-Controlled Benchmark for Membership Inference on Large Language Models

A preprint reports that current membership-inference attacks on LLMs achieve only limited accuracy on a new, carefully controlled benchmark called OLMo-Detect.

Tao Shi et al. · Added 25 Oct 2026

Preprint
Not peer reviewed

Researchers propose Suan, a gradient-level alignment method to curb LLM over-refusal

Original title: Suan: Rectifying Direct Preference Safety Alignment in Large Language Models

A preprint introduces Suan, a preference optimisation method the authors say improves safety without sacrificing helpfulness or diversity.

Oleksandr Cherednichenko et al. · Added 25 Oct 2026

Preprint
Not peer reviewed
From Zenity

A preprint reports adding a classify suffix boosts LLM safety-probe detection on unseen attacks

Original title: Prompted to Discriminate: Generalizing Malicious-Input Probes in the Wild

A preprint reports that a short classification suffix improves activation-probe safety monitoring on attack types never seen in training.

Added 25 Oct 2026

Preprint
Not peer reviewed
From Meta

Researchers propose a trainable harness that lets frozen LLM agents reuse experience

Original title: ExpHarness: Model-Agnostic Experience Learning through a Trainable Harness

A preprint describes ExpHarness, an external, trainable experience system that boosts frozen LLM agents without touching their weights.

Added 25 Oct 2026

Preprint
Not peer reviewed

Researchers report multi-agent system boosts faithfulness of model explanations

Original title: MEA: A Reward-Driven Multi-Agent System for Faithful Model Explanations

A preprint describes MEA, a multi-agent system trained to generate explanations that are verifiably grounded in model behaviour.

Yuyang Cheng et al. · Added 25 Oct 2026

Preprint
Not peer reviewed

Distilling a safety-focused teacher model can make student LLMs easier to jailbreak

Original title: Response-Based Knowledge Distillation for Multilingual Jailbreak Prevention Unwittingly Compromises Safety

A preprint reports that response-based knowledge distillation meant to boost multilingual safety instead raised jailbreak success rates in several open-source models.

Max Zhang et al. · Added 25 Oct 2026

688 entries