Rewrites a research paper so a curious 12–14-year-old can read it: the same paper, in the authors' own voice, keeping every result, number and uncertainty, in plain words. It rewrites, it does not summarise. Fine-tuned from Qwen/Qwen3.5-4B. See also the 9B version: maximerivest/qwen3.5-9b-paper-rewriter.
An Opus judge scored each rewrite from 0 to 10 on four aspects, on the same 3 held-out papers (15 parts) for every model. The large models used our refined prompts.
| Model | Faithful | Clear | Pleasant | Structure | Average |
|---|---|---|---|---|---|
| This model | 6.1 | 7.0 | 7.2 | 8.5 | 7.2 |
| Claude Opus 5.5 | 8.3 | 8.5 | 8.1 | 8.9 | 8.5 |
| GPT-6 Astra | 8.5 | 8.1 | 7.4 | 8.7 | 8.2 |
| GPT-6 Luna | 7.2 | 6.6 | 7.7 | 8.5 | 7.5 |
| Qwen3.5-4B, untrained | 3.8 |
Bold: best in each column. Faithful = faithful and exact; Clear = understandable; Structure = structure and voice.
Part by part, against each model (better, same, or worse on each of the 15 parts):
Serious problems found by the judge: about 11.3 per paper (Opus 0.7, Astra 2.0, Luna 9.0).
Read these numbers with care: three papers is a small test, so gaps under about half a point may be luck; and the judge is itself a frontier model, so it may favour the style of the synthetic rewrites this model learned from.
One paper (~4,500 words) sent alone, vLLM 0.29, bf16:
| GPU | Seconds per paper | With the draft head (MTP) | Papers per hour, GPU kept full |
|---|---|---|---|
| H100 80GB | 20 | 9 | 1,064 |
| RTX PRO 6000 | 30 | 21 | 502 |
About $3.6 per 1,000 typical papers (~7,100 words) on an RTX PRO 6000 rented at $1.80/h and kept busy.
The model was trained on two calls per paper, and expects exactly that layout.
rewrite.py in this repository does both calls with the standard library only.
vllm serve maximerivest/qwen3.5-4b-paper-rewriter --served-model-name rewriter --max-model-len 65536 \
--default-chat-template-kwargs '{"enable_thinking": false}' \
--speculative-config '{"method": "mtp", "num_speculative_tokens": 2}' # optional, ~2x faster for one reader
python rewrite.py paper.json > rewrite.md
paper.json holds the paper's parts as plain text: title, abstract,
introduction_first (first paragraph of the introduction), introduction_rest,
methods, results, discussion, conclusion, and an optional glossary.
Use greedy decoding (temperature 0) and turn thinking off. The writer field picks
a writing style; rewrite.py uses writer_a, the main style of the training data
(for the 9B, the judge scored it the same as the label used in our benchmark). The exact system and user
messages are in rewrite.py. rewrite.py reproduces the training prompts character for
character; we checked this on 15 prompts.
Each request carries a reference_glossary: definitions of the passage's technical
terms, one per entry, as - term (long form): definition [source]. We built ours from an
offline copy of Wikipedia. Almost all training examples had one (62 of 21,069 had it
empty), so the model works best with a glossary. Without one it still runs, but that
setting was rarely seen in training. Give each request only the entries whose term
appears in its text (rewrite.py does this).
mtp.safetensors is the multi-token-prediction head of the original Qwen/Qwen3.5-4B.
Our fine-tuning did not keep this head, so we added the base model's back. vLLM uses it for speculative decoding with the
--speculative-config line above. With greedy decoding, every drafted word is checked
by the model; the judge scored our 9B the same with and without it. Transformers ignores
these weights.
rewrite.py caps each reply at about
twice the section's length.Apache 2.0, like the base model (see LICENSE). Rewrites of CC BY papers should credit
the original authors and say the text was adapted.
Rewrites a research paper so a curious 12–14-year-old can read it: the same paper, in the authors' own voice, keeping every result, number and uncertainty, in plain words. It rewrites, it does not summarise. Fine-tuned from Qwen/Qwen3.5-4B. See also the 9B version: maximerivest/qwen3.5-9b-paper-rewriter.
An Opus judge scored each rewrite from 0 to 10 on four aspects, on the same 3 held-out papers (15 parts) for every model. The large models used our refined prompts.
| Model | Faithful | Clear | Pleasant | Structure | Average |
|---|---|---|---|---|---|
| This model | 6.1 | 7.0 | 7.2 | 8.5 | 7.2 |
| Claude Opus 5.5 | 8.3 | 8.5 | 8.1 | 8.9 | 8.5 |
| GPT-6 Astra | 8.5 | 8.1 | 7.4 | 8.7 | 8.2 |
| GPT-6 Luna | 7.2 | 6.6 | 7.7 | 8.5 | 7.5 |
| Qwen3.5-4B, untrained | 3.8 |
Bold: best in each column. Faithful = faithful and exact; Clear = understandable; Structure = structure and voice.
Part by part, against each model (better, same, or worse on each of the 15 parts):
Serious problems found by the judge: about 11.3 per paper (Opus 0.7, Astra 2.0, Luna 9.0).
Read these numbers with care: three papers is a small test, so gaps under about half a point may be luck; and the judge is itself a frontier model, so it may favour the style of the synthetic rewrites this model learned from.
One paper (~4,500 words) sent alone, vLLM 0.29, bf16:
| GPU | Seconds per paper | With the draft head (MTP) | Papers per hour, GPU kept full |
|---|---|---|---|
| H100 80GB | 20 | 9 | 1,064 |
| RTX PRO 6000 | 30 | 21 | 502 |
About $3.6 per 1,000 typical papers (~7,100 words) on an RTX PRO 6000 rented at $1.80/h and kept busy.
The model was trained on two calls per paper, and expects exactly that layout.
rewrite.py in this repository does both calls with the standard library only.
vllm serve maximerivest/qwen3.5-4b-paper-rewriter --served-model-name rewriter --max-model-len 65536 \
--default-chat-template-kwargs '{"enable_thinking": false}' \
--speculative-config '{"method": "mtp", "num_speculative_tokens": 2}' # optional, ~2x faster for one reader
python rewrite.py paper.json > rewrite.md
paper.json holds the paper's parts as plain text: title, abstract,
introduction_first (first paragraph of the introduction), introduction_rest,
methods, results, discussion, conclusion, and an optional glossary.
Use greedy decoding (temperature 0) and turn thinking off. The writer field picks
a writing style; rewrite.py uses writer_a, the main style of the training data
(for the 9B, the judge scored it the same as the label used in our benchmark). The exact system and user
messages are in rewrite.py. rewrite.py reproduces the training prompts character for
character; we checked this on 15 prompts.
Each request carries a reference_glossary: definitions of the passage's technical
terms, one per entry, as - term (long form): definition [source]. We built ours from an
offline copy of Wikipedia. Almost all training examples had one (62 of 21,069 had it
empty), so the model works best with a glossary. Without one it still runs, but that
setting was rarely seen in training. Give each request only the entries whose term
appears in its text (rewrite.py does this).
mtp.safetensors is the multi-token-prediction head of the original Qwen/Qwen3.5-4B.
Our fine-tuning did not keep this head, so we added the base model's back. vLLM uses it for speculative decoding with the
--speculative-config line above. With greedy decoding, every drafted word is checked
by the model; the judge scored our 9B the same with and without it. Transformers ignores
these weights.
rewrite.py caps each reply at about
twice the section's length.Apache 2.0, like the base model (see LICENSE). Rewrites of CC BY papers should credit
the original authors and say the text was adapted.