Perplexity’s R1 1776: A Comparison of Responses and Benchmarks

要約

30 秒でわかる

In February 2025, Matthew Berman reviews Perplexity’s R1 1776, a post-trained version of DeepSeek-R1 intended to reduce refusals and evasions on politically sensitive Chinese topics. He discusses an example and company-reported evaluations. Those results are limited to the tested prompts, judgments and benchmarks and do not establish that any model is universally uncensored, unbiased or accurate.

1A post-trained R1 variantPerplexity released R1 1776 to change behavior on politically sensitive topics; Berman describes it as uncensored.
2One examplePerplexity contrasts responses to a hypothetical question about Taiwan and NVIDIA. The example alone cannot establish accuracy or lack of bias.
3Company-reported evaluationThe chart reports an above-80% rate for DeepSeek-R1 and near-zero for R1 1776 on Perplexity’s particular test, not across all possible questions.
4Selected benchmark resultsThe displayed scores are close on several tests, but that does not show every capability is unchanged.

ポイント

R1 1776 is a post-trained DeepSeek-R1 variant released by Perplexity in 2025.
Perplexity reported a much lower rate of responses it categorized as evasive or overly sanitized on its own sensitive-topic evaluation.
A single geopolitical and stock-market example does not establish neutrality or predict an actual stock movement.
The displayed math and reasoning results were similar on selected tests, not proof of identical performance across all tasks.

トピックFoundation Models 評価とオブザーバビリティ AI Alignment

全文

チャプターごとに整えた全文を読む

元の文字起こしと対照

文字起こしのみ

R1 1776 and an example

Berman introduces Perplexity’s post-trained R1 model and shows the company’s response comparison. A single example cannot prove neutrality.

Perplexity released R1 1776, a post-trained version of DeepSeek-R1 intended to reduce evasions on Chinese politically sensitive topics. I describe it as removing censorship, but the change concerns the behavior Perplexity targeted and evaluated; it does not prove all refusals, restrictions or biases have disappeared.

The company shows a question about how a hypothetical change in Taiwan’s status might affect NVIDIA’s share price. In the displayed example, the original R1 response moves away from the financial question, while R1 1776 discusses possible economic and supply-chain effects. I call the latter unbiased and thoughtful, but one example cannot demonstrate neutrality, factual accuracy or predict a future stock price.

Interpreting the evaluations

Perplexity reports fewer evasive responses on a selected test and similar scores on several benchmarks. The findings have specific testing boundaries.

スライド 1 / 2Perplexity’s sensitive-topic evaluationThe chart compares company-classified evasive or overly sanitized responses on a particular test set, not a universal measure of censorship.

Perplexity’s chart places DeepSeek-R1 above 80% and R1 1776 near zero for the responses it categorized as censored in its test. In its [own description](https://www.linkedin.com/posts/perplexity-ai_today-were-open-sourcing-r1-1776a-version-activity-7297684543937859584-nVZY), the company says it assembled more than 1,000 multilingual sensitive-topic examples and used human annotators and language-model judges to identify evasive or overly sanitized answers. The chart represents that evaluation and its criteria, not a universal zero-censorship rate.

スライド 2 / 2Perplexity’s selected benchmark comparisonThe reported math and reasoning scores are close on the listed evaluations, with differences between individual scores.

Perplexity also reports similar results on the displayed internal, MMLU, DROP, MATH-500 and AIME 2024 tests. For example, the table shows 97.2 versus 97.3 on MATH-500 for R1 1776 and DeepSeek-R1. These selected results suggest comparable performance on those measurements; they do not establish that all capabilities or all kinds of bias are unchanged.