One-Line Summary

Today’s papers probe model reliability and real-world readiness: probability self-consistency checks, pretraining data poisoning pathways, embodied planning that couples text and vision, a blind-assistance video benchmark, and a faster agent runtime.

Research Papers

LLMs show probability inconsistencies when aggregating subgroup answers

This paper asks whether large language models (LLMs) give answers that “add up” when you split a population into subgroups, get estimates for each, and then recombine them—a basic check tied to the law of total probability. The authors propose a simple protocol: partition with a binary tree of personas, prompt for each subpopulation’s estimate, aggregate back to the overall population, and compare across different partitions. ¹

Across domains and frontier models, they find violations of these self-consistency identities. Notably, they report a “macro fallacy”: aggregates reconstructed from fine‑grained subgroup answers often align better with human reference data than the model’s direct population‑level answer. ¹

The effect persists across tree structures and estimation tasks, and can be partially recovered through implicit prompting—suggesting models hold relevant subpopulation knowledge that they fail to reliably propagate into aggregate estimates. The authors position statistical self‑consistency as an unsaturated, reference‑free evaluation criterion for LLMs. ¹

Poisoning pretraining via public forums and measuring its reach

This work shows that attackers can inject harmful content into web‑scale pretraining corpora through public discussion interfaces, not only via centralized sources like Wikipedia. To estimate whether adversarial content survives crawling and curation, the authors introduce HalfLife—an analysis that tracks inclusion of poisoned content into language model training data. ²

By modeling the interaction with real data pipelines, they highlight third‑party webpage content as a viable vector for poisoning and underscore the importance of measuring whether injections actually enter datasets. For practitioners, the focus shifts from static filters to auditing end‑to‑end data flows. ²

RxBrain fuses language reasoning with visual imagination for embodied plans

RxBrain presents an embodied cognition foundation model that writes plans as a single sequence combining language reasoning with visual imagination of world states. Rather than only describing steps in text or only predicting future frames, it couples task decomposition and decision logic with predicted intermediate and final physical states. ³

Technically, it uses a unified multimodal Mixture‑of‑Transformers architecture to understand and generate language, images, and video, trained via an automatic pipeline that converts embodied videos into aligned text‑visual planning steps. The authors introduce RxBrain‑Bench and report promising continuous robot action generation without large‑scale action‑data pretraining. ³

VIABench tests multimodal models on real blind assistance

VIABench is a video benchmark evaluating how well multimodal models assist blind users using first‑person videos recorded or shared by visually impaired individuals. It defines three tasks: Proactive Reminder (anticipate and narrate navigation‑critical events), Visual Question Answering (VQA, answer user questions about the scene), and Vision‑Guided Interaction (perform context‑aware interactions). ⁴

The pipeline supports both online (real‑time) and offline evaluation. Initial results show today’s Multimodal Large Language Models (MLLMs) struggle to provide comprehensive support, especially on proactive reminders that require accurate anticipation and responsiveness. Code and data are slated for release on GitHub to spur tailored research. ⁴

Open Source & Repos

AgentOS runs coding agents with millisecond cold starts and lower cost

AgentOS is an open‑source runtime for coding agents that emphasizes fast start‑up and strong isolation, pitched as a faster, lighter, cheaper alternative to traditional sandboxes. The project reports near‑zero cold starts of about 6 ms and up to 32× lower cost, alongside deny‑by‑default permissions for filesystem, network, and process access. ⁵

It embeds directly into backend services as an npm package so agents can call application functions through bindings without extra network hops, and it highlights browser‑grade isolation. The repo includes built‑in agents like Pi, Claude Code, and OpenCode, plus orchestration features. ⁵

Why It Matters

Today’s set pushes AI toward accountable, real‑world behavior. Self‑consistency turns probability identities into an everyday sanity check for decision‑relevant prompts; data‑poisoning work reframes safety as an end‑to‑end pipeline question; embodied planning and blind‑assistance benchmarks steer models toward grounded perception‑action loops rather than abstract text answers. ¹

On the tooling side, faster, cheaper agent runtimes lower the barrier to experimenting with automated workflows under tighter security controls. Together, these threads point to a maturing stack: measure what matters, secure what you train on, ground plans in sensory reality, and ship agents that start fast and stay contained. ⁵

This Week, Try It

1. Partition–Prompt–Aggregate with your favorite chatbot: ask for a population estimate, then ask by subgroups and average—compare the two and note gaps. ¹
2. Skim AgentOS on GitHub and its README to see how agent isolation and cold starts are handled in practice. ⁵