AI Science Tracker
EN

Qwen

6 discoveries name this family, 2026, across 1 scientific fields. Links are name-based family associations; exact versions and causal attribution are not verified.

First-listed affiliations

6 first-affiliation groups in linked records

Grouping uses the first listed affiliation; complete source labels remain on each entry.

Where Qwen appears in the catalog

Evidence annotations: 6/6 linked records. Structured review date and scope: 6/6. Missing review metadata does not prove a source was never checked.

Why linked?

These associations match family names in discovery titles and technology text. They are not verified model-version assignments or causal attribution. Missing links do not establish absence of use. A record may link to multiple families.

Represented fields

computer science · 6 linked records

Test contexts

direct test · 6 linked records

Annotation coverage

Evidence annotation present · 6 linked records

Examples

Up to three linked records, newest publication date first, then stable ID. This selection is not a ranking.

Discoveries

An unstated country representation guides Qwen’s capital answers

Source check: Source checked. Not independent scientific confirmation.

Finding: In a controlled two-step geography task, researchers changed a representation of a country that was never named in the prompt. At the city token, the intervention redirected Qwen2.5-7B to the paired country’s capital in 79% of 1,038 pairs. The country was also readable later in the prompt, where the intervention had no effect at the tested layers. This experiment is reported in the November 11 preprint version.

Technology: Qwen2.5-7B, cross-fitted country subspaces

Family Qwen: profile · models page

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

AI contribution
Country subspaces were estimated on one half of the pretrained Qwen2.5-7B city examples, then used in a controlled intervention on matched held-out city examples.
Test and result
The preregistered intervention compared held-out matched city pairs against random and copied-embedding subspaces, whole-state replacement, and same-country pairs.
Limitation / still unknown
The test concerns one model and a constrained task on cities the model already answered correctly. Other tasks and weak latent signals remain unestablished.
Evidence sources and annotation context

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

Evidence sources: https://arxiv.org/abs/2610.28498 · https://arxiv.org/html/2610.28498v3 · https://arxiv.org/html/2610.28498v1 · https://arxiv.org/html/2610.28498v2

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 21 Nov 2026. Review scope: arXiv metadata and v3 §5 and Appendix E, with v1/v2 text checked for the country-to-capital experiment.

No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Distilled language models retained a watermark-linked weight signal

Source check: Source checked. Not independent scientific confirmation.

Finding: In controlled comparisons of models trained on matching clean or watermarked examples, researchers measured a watermark-linked component in the distilled models’ parameter changes. Its detectable signature persisted across some later model modifications tested in the study.

Technology: Fine-tuned Qwen language models; learned Qwen2.5-3B measurement probe

Family Qwen: profile · models page

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

AI contribution
Matched learned student models allowed the authors to isolate a component of weight change associated with the watermark without changing the training task.
Test and result
The main comparison held the training task, prompts, initialization and recipe fixed between the clean and watermarked models. The authors measured changes in weights separately from output-level detections.
Limitation / still unknown
The represented observation is limited to tested models, watermark schemes and math distillation settings. The measured component is defined relative to a paired task-only change.
Evidence sources and annotation context

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

Evidence sources: https://arxiv.org/abs/2610.37266v1 · https://arxiv.org/html/2610.37266v1

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 21 Nov 2026. Review scope: §§2.2 and 3.1–3.3, matched-model geometry and aggregate validation limits.

No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

AI reviewers favored confident wording in matched research papers

Source check: Source checked. Not independent scientific confirmation.

Finding: In a 504-paper experiment, all six model reviewers more often gave higher scores to rewrites with confident wording. Claude Sonnet 4.6 favored the confident version in 40% of pairs and the less confident version in 6%, with 54% tied. Researchers checked text similarity and wording shifts before comparing evaluations.

Technology: Claude Sonnet 4.6; GPT-5.5; Gemini; Gemma; Qwen; GLM

Family Claude: profile · models page

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

AI contribution
Language models generated matched high- and low-confidence prose variants and six model reviewers independently scored the rewritten versions.
Test and result
Each model independently evaluated one rewrite at a time for each pair under standardized review instructions. The study checked style manipulation and content similarity on the rewritten manuscripts.
Limitation / still unknown
Rewriting can introduce small uncontrolled meaning changes. The experiment concerns model scores and does not measure actual conference decisions.
Evidence sources and annotation context

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

Evidence sources: https://arxiv.org/abs/2610.35254v1 · https://arxiv.org/html/2610.35254v1

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 21 Nov 2026. Review scope: v1 paired rewrite checks and 504-paper comparative review experiment, Table 4.

No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Google

Language models flagged missing evidence less often when an unsupported premise was usable

Source check: Source checked. Not independent scientific confirmation.

Finding: In paired arithmetic tasks, making an unsupported premise usable reduced missing-evidence flags by 8–30 percentage points across four tested language models. The source support was unchanged. Some answers may reflect conditional reasoning because the instructions did not require rejection of premises supplied in the question.

Technology: Gemma and Qwen language models

Family Qwen: profile · models page

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

AI contribution
Four pretrained Gemma and Qwen language models answered mathematically matched prompts with equally unsupported premises that varied in usability.
Test and result
The controlled factorial design varied actionability and the mention of relevant entities on GSM8K. All 16 primary paired contrasts survived Holm adjustment.
Limitation / still unknown
The selected tasks and instructions limit how broadly the result applies. It does not show that all answers accept unsupported statements as facts. Pattern-based scoring was revised after a 40-response audit; residual condition-dependent scoring error was not bounded by that small check.
Evidence sources and annotation context

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

Evidence sources: https://arxiv.org/abs/2610.34488v1 · https://arxiv.org/html/2610.34488v1

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 21 Nov 2026. Review scope: v1 Methods, paired premise design, §4 and Limitations.

No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Language models trusted incorrect reassurances more than incorrect warnings in a defensive audit

Source check: Source checked. Not independent scientific confirmation.

Finding: In tests of a prerecorded defensive audit dataset, incorrect reassurances moved language-model verdicts toward agreement more than incorrect warnings did. A prespecified check on seven additional model configurations found a mean asymmetry of 0.16, with a 95% interval of 0.10–0.24. This measures verdict behavior under controlled evaluation conditions.

Technology: DeepSeek; Qwen; GLM; Kimi audit models

Family DeepSeek: profile · models page

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

AI contribution
Eleven language model configurations independently evaluated prerecorded audit examples and supplied verdicts that were compared under paired evidence conditions.
Test and result
The study analyzed 16,572 evaluations on 916 matched pairs across eleven configurations. The seven-configuration confirmation used prerecorded samples and paired comparisons against valid evidence.
Limitation / still unknown
The confirmation was conducted by the same team. The data and decision conditions are controlled and do not measure real-world incidence.
Evidence sources and annotation context

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

Evidence sources: https://arxiv.org/abs/2610.32876v1 · https://arxiv.org/html/2610.32876v1

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 21 Nov 2026. Review scope: arXiv v1 Methods, experimental setup and §6.4 aggregate defensive behavior.

No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Matching final answers masked different forgetting patterns during fine-tuning

Source check: Source checked. Not independent scientific confirmation.

Finding: Researchers recovered 100% final performance on an old generation task across several replay budgets for a Qwen model. The same training runs nevertheless followed different trajectories in the probability assigned to correct old answers. The finding came from evaluating checkpoints throughout training as well as the final model.

Technology: Qwen and Mistral language-model fine-tuning

Family Qwen: profile · models page

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

AI contribution
Fine-tuned Qwen and Mistral models enabled direct measurement of task outputs and loss at multiple checkpoints under matched training protocols.
Test and result
The study varied old-task replay budgets at matched training settings and separately measured old-task generation at the endpoint and negative log-likelihood across training checkpoints.
Limitation / still unknown
The result is specific to these models, old-task response tests and replay schedules. Checkpoint behavior cannot establish whether knowledge was removed or merely harder to elicit. The qualified original Qwen retained-response bank contained only 17 examples; its 100% endpoint is not broad preservation of old-task knowledge.
Evidence sources and annotation context

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

Evidence sources: https://arxiv.org/abs/2610.30191 · https://arxiv.org/html/2610.30191v1

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 21 Nov 2026. Review scope: arXiv v1 controlled replay budgets and objective allocation, Table 7 and endpoint analysis.

No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Open the interactive profile for the timeline, or browse all organizations and model families.

29 model releases

Epoch AI data (CC BY 4.0), newest first. A discovery is attributed to a model family, not to a specific release.

DateModelAccess
Qwen3.8-2.4T-A95B Open weights (restricted use)
Qwen 3.8 Max API access
Qwen3.7-Plus API access
Qwen 3.6 Plus API access
Qwen3.5-Omni-Plus API access
Qwen3.5-Omni-Flash API access
Qwen3.5-122B-A10B Open weights (restricted use)
Qwen3.5-35B-A3B Open weights (unrestricted)
Qwen3.5 397B-A17B Open weights (unrestricted)
Qwen3-Coder-Next Open weights (unrestricted)
Qwen3-Max-Thinking API access
Qwen3-VL-235B-A22B Open weights (unrestricted)
Qwen3-Omni-30B-A3B Open weights (unrestricted)
Qwen3-Max API access
Qwen Image Open weights (unrestricted)
Qwen3-235B-A22B-Thinking (Jul 2025) Open weights (unrestricted)
Qwen3-235B-A22B (Jul 2025) Open weights (unrestricted)
Qwen3-Coder-480B-A35B Open weights (unrestricted)
Qwen3 Embedding Open weights (unrestricted)
Qwen3-235B-A22B Open weights (unrestricted)
Qwen2.5-72B Open weights (unrestricted)
Qwen2.5 Instruct (72B) Open weights (restricted use)
Qwen2.5-32B Open weights (unrestricted)
Qwen2-72B Open weights (unrestricted)
Qwen1.5-72B Open weights (restricted use)
Qwen-VL-Max API access
Qwen-72B Open weights (restricted use)
Qwen-Audio-Chat Open weights (restricted use)
Qwen-VL Open weights (restricted use)