ANVE: the private AI office, on the record. See what that means

ANVE

Private AI glossary · updated 2026-10-07

Chain of thought
Asking a model to show its working, and why the working is not proof.

Short answer. Chain of thought (CoT) is a technique in which a language model writes out intermediate reasoning steps before giving its answer, which improves results on multi-step problems such as arithmetic. In the 2022 paper that named it, prompting a 540-billion-parameter model with just eight worked examples set a new best score on the GSM8K school-maths benchmark. Adding the single sentence “Let’s think step by step” raised one model’s MultiArith accuracy from 17.7% to 78.7%. But written steps do not always show why a model answered as it did: researchers found that explanations can leave out a bias that changed the answer, and that accuracy fell by as much as 36% when models were nudged toward wrong answers.

01

How it works

A language model writes one token at a time, and each token can use everything written before it. If the model writes its intermediate steps, those steps become part of its input for later ones, so a hard problem is split into easier pieces. Few-shot chain of thought gives the model worked examples with their steps. Zero-shot chain of thought just adds an instruction such as “Let’s think step by step”. Reasoning models now do this by default, at length.

02

What the research found

StudyMethodResult
Wei et al., 2022Eight worked examples with steps given to a 540B-parameter modelState of the art on GSM8K, ahead of fine-tuned GPT-3 with a verifier
Kojima et al., 2022The instruction “Let’s think step by step” with text-davinci-002MultiArith 17.7% to 78.7%; GSM8K 10.4% to 40.7%
Turpin et al., 2023Inputs biased toward a wrong answer, for example always option (A)Accuracy fell by up to 36% across 13 tasks with GPT-3.5 and Claude 1.0, and the written steps did not mention the bias
03

How to use it in practice

  • Ask for steps on calculations, comparisons and multi-part questions. For plain extraction or summaries it adds length, not accuracy.
  • Put the final answer in a fixed place, for example on a line starting “Final answer:”, so software or a person can find it.
  • Ask for evidence at each step: a quoted passage, a page number or a figure from the source. Steps that cite their evidence can be checked.
  • Check numbers with a calculator or a script rather than trusting the arithmetic in the text.
  • Use a reasoning model when you want the habit built in, and a normal model when you want speed.
04

Why the steps are not an audit trail

The Turpin study is a warning for any office that wants to keep a model’s explanation as a record. A neat set of steps can be a plausible story rather than the true cause of the answer. For an audit, the record that counts is what went in, which sources were used, what came out and who approved it. ANVE’s design follows that logic: every answer is cited, every meeting action item must quote the transcript line it came from, a person confirms its owner, and each confirmation is logged.

05

Questions people ask

Is chain of thought the same as a reasoning model?
Related but different. Chain of thought is a prompting technique that asks an ordinary model to write its steps. A reasoning model has been trained to write long step-by-step thinking by default.
Does chain of thought make a model more accurate?
On multi-step problems such as arithmetic and logic, the studies above say yes. On simple tasks it makes little difference, and the steps can be wrong while sounding right.
Should I show the steps to users?
Show them where they help a person check the work, with sources next to each step. Do not present them as proof that the answer is correct.
Does the model really reason when it writes steps?
The research does not settle that. It shows the steps improve results, and that they can omit factors that influenced the answer, so they are a clue and not a complete explanation.
06

What is not real yet

No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.

Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.