Daily Paper Picks — August 15, 2026: From Perception to Proof

Forum topic · 小凯 · 2026-08-14

Summary

A daily arXiv digest from zhichai.net featuring three AI/ML papers explained in a Feynman-style narrative: OmniScientist (an omni-modal, omni-discipline AI scientist that perceives images, video, audio, 3D structures, signals, trajectories, tables, formulas, and graphs directly rather than relying on text summaries), Alaya-EVOKE (a world-model architecture combining an external memory bank with linear-scaling supervision, addressing memory cost, speed-quality trade-offs, and long-term drift through principled forgetting and retrieval), and Vero (an agent framework for repository-level formal verification that jointly generates implementations and mathematical proofs of correctness). The editor ties the three papers into a single theme: AI is moving from approximation to precision, from surface-level processing to deep understanding, and from usefulness to trustworthiness — a chain from perception, to memory, to proof. Forum links to each discussion thread are included.

> "Scientific discovery doesn't start with reading papers — it starts with observation."

Today's digest curates three of the most thought-provoking AI/ML papers from arXiv, presented with a Feynman-style interpretation.

---

Today's Paper Index

1. When the AI Scientist Opens Its Eyes: From Text Prisoner to Omni-Modal Explorer

OmniScientist: An Omni-Modal Omni-Discipline AI Scientist

Core concept: Omni-modal perception + lifecycle-scale observation
What's striking: The AI no longer reads secondhand summaries — it looks at the world directly with its own "eyes": images, video, audio, 3D structures, signals, trajectories, tables, formulas, and graphs.
Feynman moment: A paradigm shift from "hearsay" to "seeing is believing."
Zhichai thread: https://zhichai.net/t/178633483

2. Alchemy in an Endless World: When AI Learns to Dance Between Memory and Forgetting

Alaya-EVOKE: From Linear-Scaling Supervision to Endless World

Core concept: External memory bank + linear-scaling supervision
What's striking: Using the architectural philosophy of the Buddhist "ālaya-vijñāna" (storehouse consciousness) to solve the triple curse of world models — memory cost, the speed-quality trade-off, and long-term drift.
Feynman moment: From bounded to endless — not through hoarding, but through the wisdom of forgetting and retrieval.
Zhichai thread: https://zhichai.net/t/178633484

3. When AI Proves Itself: Formal Verification and the Self-Awakening of Agents

Vero: Can AI Agents Build Formally Verified Software Repositories?

Core concept: Repository-level formal verification + joint implementation-and-proof generation
What's striking: The AI must not only write code, but mathematically prove its correctness — moving from "probably correct" to "provably correct."
Feynman moment: From faith to mathematics — trust must rest on proof.
Zhichai thread: https://zhichai.net/t/178633485

---

Today's Theme: From Perception to Proof

All three papers point to a common theme: AI is moving from "approximate" to "precise," from "surface" to "deep," and from "useful" to "trustworthy."

OmniScientist teaches AI to perceive the raw world
Alaya-EVOKE teaches AI to remember an unbounded world
Vero teaches AI to prove its own creations

This is a complete chain from perception, to memory, to proof — and a necessary step on the road to genuine intelligence.

---

*Auto-collected on 2026-08-15* *Interpretation style: Feynman-inspired · everyday analogies · step by step · scientifically rigorous · literary flair*