“Finding a way to get LLMs to reliably use external tools is far from trivial.”
Our reading, the sentence that was checked:
Getting LLMs to reliably use external tools is far from trivial.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Getting GPT to work with external tools is harder than you think, paragraph 22.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-getting-gpt-to-work-with-external-19406b-a1
We recorded 16 claims from this post. See every claim from this post.
The wording above is the version this site observed on 2026-10-29. The source may have been edited after 2023-10-23. This page does not claim the wording was captured then.
What the claim commits to
- Kind
- Observation
- Deadline, as the speaker wrote it
- None stated. This site does not add one.
- Deadline state when assessed
- No horizon
- Deadline state on 2026-10-29
- No horizon (the release’s reference date)
From the speaker
- Finding a way
- LLMs to reliably use external tools
- far from trivial
Our interpretation
- The ordinary reading is that dependable tool use across varied tasks required substantial engineering or research beyond simply connecting an LLM to tools.
- Reliability concerns correct task completion, including selecting actions and handling tool results, not merely producing a syntactically valid call.
- This is a contemporary difficulty assessment, not a forecast that the problem is impossible.
- For this ongoing technical observation, I use the 2023 evaluation cycle through publication; no exact observation window is specified.
Counts as supported if: Contemporary primary evaluations show recurring failures after tool integration, alongside substantive methods developed to improve planning, tool selection, or recovery.
Counts as contradicted if: Reproducible contemporary evidence shows that straightforward integration achieves dependable completion across varied tools and tasks without substantial additional methods.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Was it true for the period described?
Outcome: Supported
The dated evidence matches what the claim committed to, under the criteria frozen before grading.
Scored on a plain reading. The speaker named no measure for this claim, so it was judged on the reading an informed reader would take, stated below under our interpretation. Both reviews had to find that reading fair.
- Evidence as of
- Deadline state when assessed
- No horizon
- Assessment recorded
- Policy version
- 2026-10-28.7
Rationale
WebArena's original results show that tool-enabled agents still frequently failed complete tasks. ToolLLM documents substantial additional methods that improved performance. Together they support the ordinary reading that dependable, varied tool use was a significant engineering problem in 2023. ToolLLM's successes also show that progress was possible; the claim does not require impossibility.
Limitations
- The conclusion uses an ordinary qualitative reading of difficulty.
- Simple, narrowly constrained tool calls may have been reliable.
- These historical results do not assess systems available in 2026.
Evidence (2 items)
Supports the claim: arXiv; Shuyan Zhou and colleagues,
10.59%
Why it matters here: The July benchmark reports low completion rates for agents operating realistic websites with supporting tools, despite incorporating reasoning before acting. This supports the inference that connecting models to an actionable environment did not itself deliver dependable performance.
- Publisher
- arXiv; Shuyan Zhou and colleagues
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Supports the claim: arXiv; Yujia Qin and colleagues,
significantly enhances
Why it matters here: The researchers developed search over alternative action sequences, tool-specific training data, and API retrieval to improve complex tool use. Their successful results qualify pessimistic readings while documenting substantive engineering beyond a simple connection.
- Publisher
- arXiv; Yujia Qin and colleagues
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Getting GPT to work with external tools is harder than you think
Finding a way to get LLMs to reliably use external tools is far from trivial.
- Where
- paragraph 22
- Published
- 2023-10-23
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-getting-gpt-to-work-with-external-19406b-a1: Current. Outcome: Supported. Evidence as of 2026-10-28. Public date not recorded.
Recorded events
Assessment publishedBy this site
First assessment published after both reviews and the final-text check passed.
About assessment marcus-getting-gpt-to-work-with-external-19406b-a1.