LmCast :: Stay tuned in

Qwen 3.8 follows GPT-5.5 Pro reasoning prefills

Recorded: Sept. 9, 2026, 6 p.m.

Original Summarized

HN: https://news.ycombinator.com/edit?id=49630026 · GitHub

Skip to content

Search Gists

Search Gists

All gists
Back to GitHub

Sign in

Sign up

Sign in

Sign up

You signed in with another tab or window. Reload to refresh your session.
You signed out in another tab or window. Reload to refresh your session.
You switched accounts on another tab or window. Reload to refresh your session.

Dismiss alert

Instantly share code, notes, and snippets.

wsxiaoys/reasoning-prefill-1-1.md

Secret

Last active
September 9, 2026 17:50

Show Gist options

Download ZIP


Star

1
(1)

You must be signed in to star a gist

Fork

0
(0)

You must be signed in to fork a gist


Embed

Select an option

Embed
Embed this gist in your website.

Share
Copy sharable link for this gist.

Clone via HTTPS
Clone using the web URL.

No results found


Learn more about clone URLs


Clone this repository at <script src="https://gist.github.com/wsxiaoys/e0286dc6bb624ff5fdf49e7f4c528ba3.js"></script>

Save wsxiaoys/e0286dc6bb624ff5fdf49e7f4c528ba3 to your computer and use it in GitHub Desktop.

Code

Revisions
3

Stars
1


Embed

Select an option

Embed
Embed this gist in your website.

Share
Copy sharable link for this gist.

Clone via HTTPS
Clone using the web URL.

No results found


Learn more about clone URLs


Clone this repository at <script src="https://gist.github.com/wsxiaoys/e0286dc6bb624ff5fdf49e7f4c528ba3.js"></script>

Save wsxiaoys/e0286dc6bb624ff5fdf49e7f4c528ba3 to your computer and use it in GitHub Desktop.

Download ZIP

HN: https://news.ycombinator.com/edit?id=49630026


Raw

reasoning-prefill-1-1.md

Reasoning prefills on a few open models, v1.1
A follow-up to Reasoning prefills on a few open models and Stolen Thoughts
This v1.1 reruns the reasoning-prefill experiment with GPT-5.5 Pro as the teacher.
For each problem, I generated two responses from each target model:

an ordinary, unprefilled response; and
a response starting with the first 1% of GPT-5.5 Pro's reasoning, inserted into the target model's reasoning channel.

The visible answer remained freely generated. I then measured how much of the teacher's visible answer appeared in the first 100 tokens of the target model's answer. The table below reports unigram source recall so the numbers are comparable to my previous post. Deltas are absolute percentage-point changes.
All problems
The evaluation contains 45 problems: 15 STEM, 15 non-STEM, and 15 synthetic puzzles.

Model
n
Unprefilled
GPT-5.5 Pro reasoning prefill
Delta

DeepSeek V4 Flash
45
40.53%
40.89%
+0.35 pp

Inkling
45
37.82%
38.67%
+0.85 pp

Kimi K3
45
50.11%
54.42%
+4.31 pp

Qwen3.8 A95B
45
33.92%
54.50%
+20.58 pp

Qwen by category

Category
n
Unprefilled
GPT-5.5 Pro reasoning prefill
Delta

STEM
15
36.21%
63.76%
+27.55 pp

Non-STEM
15
38.26%
52.73%
+14.46 pp

Puzzle
15
27.28%
47.00%
+19.72 pp

All
45
33.92%
54.50%
+20.58 pp

Discussion
Qwen barely moved toward Opus 4.8 in the earlier experiment, but moved by +20.58 points toward GPT-5.5 Pro here, including a large effect on the private synthetic puzzles. The data suggest that Qwen may have learned from GPT-5.5 Pro, or from a closely related GPT model, rather than from Opus.
Kimi K3's overlap with GPT-5.5 Pro is also high both without and with the prefill (50.11% and 54.42%), although the prefill adds only +4.31 points.

Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment

Footer

© 2026 GitHub, Inc.

Footer navigation

Terms

Privacy

Security

Status

Community

Docs

Contact

Manage cookies

Do not share my personal information

You can’t perform that action at this time.

A follow-up experiment titled Reasoning prefills on a few open models and Stolen Thoughts was conducted to investigate the effect of reasoning prefilling when using GPT-5.5 Pro as the teacher model. The methodology involved generating responses for a set of problems using target models under two conditions: an ordinary, unprefilled response and a response where the target model's reasoning channel was prepended with the initial one percent of GPT-5.5 Pro's reasoning. The primary metric measured was the unigram source recall, assessing how much of the teacher's visible answer appeared within the first one hundred tokens of the target model's generated response.

The evaluation encompassed forty-five problems distributed across three categories: fifteen STEM problems, fifteen non-STEM problems, and fifteen synthetic puzzles. The results demonstrated varying degrees of influence based on the specific model tested. Among the evaluated models, DeepSeek V4 Flash showed a modest increase in recall of 0.35 percentage points, while Inkling experienced a larger gain of 0.85 percentage points. Kimi K3 displayed a substantial improvement, noting an increase of 4.31 percentage points in overlap with the teacher model when reasoning was prefilled. Qwen3.8 A95B exhibited the most significant gain, achieving a delta of 20.58 percentage points in source recall.

When examining performance categorized by subject matter, the gains were also differentiated. For STEM problems, the prefilling yielded a delta of 27.55 percentage points, whereas non-STEM problems registered a delta of 14.46 percentage points. The synthetic puzzles also benefited considerably, showing an increase of 19.72 percentage points. Overall, the aggregation across all 45 problems indicated a substantial gain, with the overall delta being 20.58 percentage points.

Qualitative analysis suggested distinct dependencies between the models and the teacher influence. The discussion indicated that Qwen's performance shift moved it toward GPT-5.5 Pro, suggesting that the model learned from GPT-5.5 Pro or a closely related GPT architecture, rather than the previously tested Opus 4.8. Furthermore, Kimi K3 demonstrated a high degree of overlap with GPT-5.5 Pro both with and without the prefilling intervention. These findings indicate that reasoning prefilling modulates how effectively open models incorporate external reasoning signals from a powerful teacher model, with the magnitude of the effect depending on the model's architecture and the nature of the task.