Skip to main content

READ / QUESTION / REPRODUCE

Read with a question.

Foundational papers selected for AI engineering and embodied intelligence. These short guides are based on the linked abstracts; they are not full-paper reviews or personal experimental results.

← Research · Saved reading

3 papers

RAG · 2020 · Abstract guide

Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks

How can generation use external knowledge?

Method
Combine a sequence generator with dense retrieval; compare shared passages per sequence with passages that can change per token.
Reported evidence
The authors report improvements on open-domain QA and generation tasks.
Reading boundary
Results concern the paper's tasks and baselines, not every knowledge-base integration.
Questions to investigate

Reading questions: how do retrieval failures affect output, and can evidence be traced?

Source abstract and paper ↗ · Planned experiment →

ReAct · 2022 · Abstract guide

ReAct: Synergizing Reasoning and Acting in Language Models

How can a language model combine reasoning with interaction?

Method
Interleave reasoning traces and task actions, using external information to update plans.
Reported evidence
The authors report evaluations on QA, fact verification and interactive decision-making benchmarks.
Reading boundary
Benchmark performance does not establish reliable real-world tool use; explanations are not proof of correctness.
Questions to investigate

Reading questions: how are failed actions handled, and are no-tool baselines comparable?

Source abstract and paper ↗ · Planned experiment →

RT-2 · 2023 · Abstract guide

RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control

How can visual-language knowledge support robotic control?

Method
Represent actions as text tokens and co-fine-tune on robot trajectories and vision-language tasks.
Reported evidence
The authors report generalization involving new objects and instructions in robotic evaluations.
Reading boundary
This does not imply generalization to arbitrary hardware or environments. No reproduction has been performed here.
Questions to investigate

Reading questions: what roles do action representation, data distribution and execution play?

Source abstract and paper ↗ · Planned experiment →