RightSignal · Thursday, October 29

AI safety

Heads of Security: Learned Attention Gates Improve Prompt Injection Robustness

Igor Shilov, Xiaoxue Yang, Alexander Panfilov, Evgenii Kortukov, Matthieu Meeus, Peter Romov · Submitted 27 Oct 2026

Preprint · not peer reviewed

Prompt injection is when hidden instructions in a web page or document trick an AI model into obeying them. The authors add small learned gates to the model's attention that control how much it listens to trusted versus untrusted text.

Key finding: One strong attack succeeded 97% of the time against GPT-OSS-20B but only 10% against the defended model, which kept its normal skills.

Why it matters: AI agents that read emails and websites need a way to read untrusted text without following its orders.

Read on arXiv

Systems

TAO: Tool-Aware Joint Scheduling for LLM-Based Agentic Serving

Shuhuai Li, Shengdong Ye, Dongdong Ge, Yinyu Ye, Jianghao Lin · Submitted 27 Oct 2026 · 2 upvotes yesterday

Preprint · not peer reviewedTrending

AI agents switch between running the model on GPUs and waiting for tools on CPUs. TAO schedules which waiting sessions get GPU time based on when their tools will return, what is already cached, and how much memory they need.

Key finding: Across 12 settings it cut latency by up to 26.8% and time to first token by up to 36.3% against the best baseline.

Why it matters: Serving agents cheaply depends on not wasting expensive GPUs on sessions that are still waiting for tools.

Read on arXiv

DTCA: Denoising Token Credit Assignment for Reinforcement Learning in Diffusion Language Models

Hongze Chen, Miao Pan, Lizhou Liang, Qinfeng Li, Xinyin Ma, Yongliang Shen · Submitted 26 Oct 2026 · 254 upvotes yesterday

Preprint · not peer reviewedTrending

Diffusion language models write by refining many words in parallel, so it is hard to tell which words earned a good final answer during training. DTCA estimates each word's contribution and uses that to guide reinforcement learning.

Key finding: On LLaDA-8B-Instruct it improved maths scores by up to 9.0 points on MATH500 and 5.3 points on GSM8K.

Why it matters: Better training methods could make diffusion models, a faster alternative to word-by-word models, better at reasoning.

Read on arXiv

Efficient Streaming Audio-Visual Target Speaker Extraction for Real-World Acoustic Scenes

Wendi Sang, Kai Li, Yifan Li, Jianqiang Huang, Xiaolin Hu · Submitted 23 Oct 2026 · 22 upvotes yesterday

Preprint · not peer reviewedCode releasedTrending

The authors built a realistic test set with many background speakers and noise, and Falcon, a light model that uses video of a speaker's face to pull out their voice in real time.

Key finding: Compared with a standard baseline, Falcon uses 91.9% fewer parameters and runs 11.1 times faster on a GPU, while scoring best on several benchmarks.

Why it matters: Picking one voice out of a noisy room in real time on small devices helps hearing aids, calls and assistants.

Read on arXiv

Scaling Laws for Tool-Use Agents via Controllable Environment Synthesis

Andrew Szot, Samira Abnar, Omar Attia, Abbas Kazerouni, Michael Kirchhof, Cheng-Yu Hsieh · Submitted 27 Oct 2026

Preprint · not peer reviewed

The authors generated practice environments for AI agents that use tools, varying the tools, tasks and databases, and trained 849 models to see which mix of data helps most.

Key finding: As data budgets grow, adding more varied tools and databases helps more than adding more tasks.

Why it matters: It gives teams a rule of thumb for spending training data when building agents.

Read on arXiv

Your RightSignal

Topics in News

Heads of Security: Learned Attention Gates Improve Prompt Injection Robustness

TAO: Tool-Aware Joint Scheduling for LLM-Based Agentic Serving

DTCA: Denoising Token Credit Assignment for Reinforcement Learning in Diffusion Language Models

Efficient Streaming Audio-Visual Target Speaker Extraction for Real-World Acoustic Scenes

Scaling Laws for Tool-Use Agents via Controllable Environment Synthesis

That’s every paper.