🌏 中文版
Units 15–18 place language models in a longer decision loop: Embodied Perception, Inference-time Compute, Agent Reasoning, and Embodied Agents. Inputs may include visual or environmental state; outputs may be tool calls or actions rather than text.
Inference-time compute is resource allocation
More test-time computation can mean more candidates, longer reasoning traces, search, verification, or revision. The useful comparison asks whether added compute improves verifiable outcomes under a fixed task and budget. Longer output alone is not deeper reasoning.
Choose one primary budget—tokens, wall-clock time, model calls, or money—and retain each observation, action, tool result, and final answer. Without traces, model, tool, and environment failures collapse into one label.
Reasoning agents add state and tools
Single-turn QA evaluates an answer. Agent evaluation also needs task completion, wasted steps, tool failures, cost, and safety constraints. Hold tool availability, permissions, initial state, and stopping conditions constant before comparing agents.
Embodiment propagates errors
Embodied perception constructs state from observations; an embodied agent converts a language-level plan into action. Perception errors propagate through planning and execution. Preserve replayable observation-action logs and place human gates before high-impact actions.
The official schedule also includes guest sessions on computer-use agent safety, memory, continual learning, and speech. Course Info limits recordings to enrolled students and Cal-affiliated auditors, while the schedule provides no anonymous slides for some sessions. This series records their place in the curriculum without reconstructing them.
Finish with a staged project
The Course Project requires an abstract, midpoint report, presentation, and final report. A self-study project should keep the same checkpoints: define the question and baseline, submit a midpoint failure analysis, and only then consolidate results. Team matching, staff feedback, and course compute credits are not public, so arrange peer review and a compute ceiling yourself.
Same two model calls, different outcomes
This is an invented teaching scenario, not a classroom experiment. The task is to report an order’s delivery date using at most two model calls. Both agents have identical tools, permissions, and initial data. Success means matching the order tool’s returned date.
Agent A guesses Friday on its first call and revises its own text to Saturday on its second, without checking the order. Agent B requests the order lookup on its first call; the tool returns Monday, and B answers Monday on its second. In this scenario A fails the criterion and B passes. The difference is access to relevant evidence, not output length.
Check: If B’s tool returns an error and the model treats the error text as a date, where should diagnosis start? Inspect the tool result and the model’s error-handling step; the final answer alone does not show a general reasoning failure. Only model-call count is fixed here. A latency or monetary comparison must also include tool costs.
Changelog
- 2026-10-11: Added an invented teaching example, worked reasoning, and a comprehension check with an answer.
References
Loading...