Artificial Intelligence research · Research Watch · Page 20

Preprint
Not peer reviewed

Preprint finds multi-agent AI debate rarely beats budget-matched sampling

Original title: Beyond Symmetric Agents: Cognitive Diversity and Multi-Agent Debate in Small Language Models

A preprint reports that debate among small language models offers little advantage over simple budget-matched sampling methods.

Leonardo Ferreira et al. · Added 28 Oct 2026

Preprint
Not peer reviewed

Researchers report extracting memorised API keys from black-box LLMs via output-only access

Original title: Practical Secrets Extraction against Black-box LLMs

A preprint describes a black-box method that recovers memorised API keys from commercial LLMs using only their text outputs.

Shiqian Zhao et al. · Added 28 Oct 2026

Preprint
Not peer reviewed
From Google DeepMind

Researchers report benign audio fine-tuning can break safety guardrails in Audio LLMs

Original title: Benign Fine-Tuning Breaks Safety Alignment in Audio LLMs

A preprint reports that fine-tuning Audio LLMs on seemingly harmless audio can sharply raise jailbreak success rates, depending on model architecture.

Added 28 Oct 2026

Preprint
Not peer reviewed

Preprint finds tiny parameter subsets can trigger LLM collapse

Original title: Weight-Adjusted Gradients Reveal Parameter Importance and Failure Modes in LLMs

A preprint describes a metric combining weight and gradient signals that pinpoints a minuscule fraction of parameters capable of causing severe LLM breakdowns.

Shrestha Datta et al. · Added 28 Oct 2026

Preprint
Not peer reviewed

Preprint reports AI agents can piece together hidden prompt injections from fragments

Original title: Divide and Inject: Can Agents Reconstruct an Indirect Prompt Injection from Fragments?

A preprint reports an attack that splits malicious instructions into fragments, letting agents reconstruct and act on them with higher success than prior methods.

Michael Lee et al. · Added 28 Oct 2026

Preprint
Not peer reviewed

Researchers report advantage-clipping method speeds up and improves RL training for LLM reasoning

Original title: Towards Better Training Signal: Advantage Clipped Policy Optimization

A preprint reports a new policy-optimisation method, ACPO, that boosts accuracy and training speed over PPO and GRPO on maths reasoning benchmarks.

Ruichuan Huang et al. · Added 28 Oct 2026

Preprint
Not peer reviewed

Preprint ties LLM reasoning rewards to formally verified proof steps

Original title: Learning to Prove, Not Just to Answer: Reinforcement Learning from Formal Verification for Natural-Language Logical Reasoning

A preprint reports that rewarding formally verified proof steps, not just correct answers, improves both accuracy and proof validity in logical reasoning LLMs.

Qili Zhang et al. · Added 28 Oct 2026

Preprint
Not peer reviewed
From Meituan

Researchers report mixed-precision method keeps accuracy while shrinking recurrent-state memory

Original title: DAMP: Decay-Aware Mixed-Precision Recurrent-State Quantization

A preprint presents DAMP, a method that compresses recurrent states in hybrid attention models while keeping accuracy close to full precision.

Added 28 Oct 2026

Preprint
Not peer reviewed

Researchers report video scenarios can be used to jailbreak Video-MLLMs

Original title: SceneJail: Exploiting Video Scenario Context to Jailbreak Multimodal LLMs

A preprint describes SceneJail, an attack that embeds harmful queries in matched video scenarios to bypass Video-MLLM safety checks.

Wenyu Chen et al. · Added 28 Oct 2026

Preprint
Not peer reviewed

MiniRep system aims to make multi-agent LLM debate more resistant to manipulation

Original title: MiniRep: Robust Reputation-Based Aggregation for Multi-Agent Debate

A preprint presents MiniRep, a reputation-based aggregation method for multi-agent LLM debate designed to resist malicious agents.

Jiaming Zhang et al. · Added 28 Oct 2026

Preprint
Not peer reviewed
From Horizon Robotics

A preprint reports a ranking-based method for offline-to-online robot RL

Original title: RankQ: Offline-to-Online Reinforcement Learning via Self-Supervised Action Ranking

Researchers propose RankQ, which orders actions by quality rather than penalising unseen ones, reporting gains in simulated and real robot tasks.

Added 28 Oct 2026

Preprint
Not peer reviewed

Researchers propose weight-level method to suppress unwanted LLM personas

Original title: LLM Persona Unlearning

A preprint introduces a benchmark and method aimed at making unwanted LLM personas harder to trigger without ruining general usefulness.

Kemou Li et al. · Added 28 Oct 2026

Preprint
Not peer reviewed
From Meta

New benchmark tests whether coding agents can build whole repositories from scratch

Original title: E2E-SWE: Benchmarking LLMs on Building Working Codebases from Scratch

A preprint introduces E2E-SWE, a benchmark checking if LLM agents can build complete, working repositories from a specification alone.

Added 28 Oct 2026

Preprint
Not peer reviewed

A preprint reports synthetic images can sharpen VLM visual perception skills

Original title: VisionFoundry: Teaching VLMs Visual Perception with Synthetic Images

A preprint describes VisionFoundry, a pipeline that builds synthetic image-QA data to boost vision-language models' spatial and perceptual accuracy.

Guanyu Zhou et al. · Added 28 Oct 2026

Preprint
Not peer reviewed

Researchers propose a trusted action-card defence for autonomous browser agents

Original title: The Verifiable Action Card: Trustworthy Human-in-the-Loop Control for Secure Autonomous Agents

A preprint describes a human-in-the-loop safeguard that grounds browser-agent approvals in the real pending action, cutting attack success across tested models.

Hasnain Irshad et al. · Added 28 Oct 2026

Preprint
Not peer reviewed

Researchers report GPU side channel leaking LLM privacy via sparse attention patterns

Original title: SparLeak: Privacy Leakage from Sparse Attention in LLM Inference on Shared GPUs

A preprint describes a GPU side-channel attack that infers private query details and model outputs from sparse attention memory access patterns.

Fahao Chen et al. · Added 28 Oct 2026

Preprint
Not peer reviewed

Researchers propose Bayesian method to cut cost of repeated LLM evaluation runs

Original title: Sequential Bayesian Evaluation of Large Language Model Behavior

A preprint presents a Bayesian sequential sampling method that reduces the number of generations needed to evaluate stochastic LLM behaviour.

Saatvik Kher et al. · Added 28 Oct 2026

Preprint
Not peer reviewed
From Ant Group

Researchers report clinical AI agents falter when patient records lack supporting evidence

Original title: EHR-RobustGym: Benchmarking and Training Agents for Robust Clinical Reasoning

A preprint introduces a benchmark showing clinical LLM agents struggle to detect when hospital records lack evidence to answer a query.

Added 28 Oct 2026

Preprint
Not peer reviewed
From FuriosaAI

Researchers simulate flash-backed memory to speed up LLM serving

Original title: Characterizing High Bandwidth Flash for LLM Serving

A preprint reports that adding high-bandwidth flash with smarter scheduling can cut LLM serving time and energy use in simulation.

Added 28 Oct 2026

Preprint
Not peer reviewed
From Amazon, Microsoft

A preprint reports training on aligned data can still cause narrow misalignment elsewhere

Original title: Aligned Data Can Induce Misalignment via Context Confusion

Researchers describe context confusion, where fine-tuning on data that is aligned in one domain makes models misaligned in a different, related domain.

Added 28 Oct 2026

688 entries