Thomson Reuters–Imperial
Frontier AI Research Lab

Calls for SovereignAI — an institution's capacity to build, deploy and govern AI independently — rarely come with concrete advice on how to get there. This report argues the route is Continual Learning on readily available open-weight models, and demonstrates it: a full mid- and post-training stack applied to Qwen3.5-397B and Qwen3.6-35B, with safeguards for plasticity and stability at every stage and the minimum number of high-impact parameter interventions. The result is a distinctive π-shaped improvement pattern — sharp gains across targeted and untargeted capabilities alike, while nearly eliminating the forgetting that dogs narrow domain adaptation. Read less as a model launch than as a blueprint for a model factory.

The papers listed below are the components this report is built from: the constitution supplies its alignment target, the model-steering paper its value-alignment stage, CapTrack its forgetting measurement, and Scales++ the pruning that makes that measurement cheap enough to run at every checkpoint.

Primary authors Shengzhuang Chen · Jerrod Parker · Yejin Bang · Andrew M. Bean · Nabeel Seedat
Core contributors Stefan Winzeck · Daniil Glazko · Jannik Zgraggen · Fangyi Yu · Scott Arnott · Dietrich Trautmann · Luca Ciuffreda · Guglielmo Bonifazi · Davide Romano · Bradley Bell · Kirsty Fielding
Contributors Daniele Giofrè · Tom Zielund · Ipshita Chatterjee · Sneha Murthy Ghantasala · Manpreet Nanreh · John Scoville · Maciej Sakowicz · Wassim Seifeddine · Lukas Thede
Project lead Jonathan Richard Schwarz

In partnership with Imperial College London · DatologyAI · Lambda

A written constitution, grounded in the Universal Declaration of Human Rights, setting out the values a model is intended to reflect, the reasoning behind them, and the standard against which its outputs can be judged. It requires the model to be broadly safe, broadly ethical, unbiased and impartial, compliant with relevant guidelines, and genuinely helpful — with safety and ethics taking precedence. A deliberate departure from proprietary commercial constitutions: the normative foundations of a model should be a shared public resource, open to contribution and debate. Serves as the alignment target for the Snowdon and Thomson series.

A rank-one weight edit routed through a diagonal Fisher metric, so the correction lands where the model's predictive distribution is locally least sensitive. Constitutional DPO then consolidates the edit over preference pairs drawn from expert-curated and synthetic data. Capability retention is treated as an objective rather than a diagnostic: a re-alignment that trades away general competence is of no practical use. The edit merges into the weights — no added parameters, no added latency. Also introduces Perspective Bench, the geopolitical-contestation evaluation used to score re-alignment. Produces the Snowdon checkpoints and supplies the value-alignment stage of the Thomson pipeline.

Lay users seeking legal guidance omit the jurisdiction, dates, party relationships and procedural posture that determine the answer. InsufficiencyBench evaluates what LLMs do when a query is missing what it needs: whether they identify the gap and ask, or quietly assume values and answer anyway.

Contract scrubbing — the final pass over a transactional agreement for errors and inconsistencies — is routine, painstaking, and apparently well matched to long-context reasoning, consistency checking and NER. Nobody had formally evaluated it. ContractScrub is the first benchmark that does: 3,014 annotated tasks across 44 CUAD contracts, hand-seeded by experienced lawyers with defined-term misuse, broken cross-references, capitalisation errors and inconsistent language. Across nine frontier and open-weight models, only one clears 0.75 macro-average recall, despite strong scores on the general benchmarks that supposedly cover these skills.

The accuracy-centric view of forgetting suited early continual-learning work in vision, where a model was optimised for a single metric. Modern LLMs are interactive systems whose usefulness rests on reasoning, robustness, default preferences and protocol adherence, so CapTrack redefines forgetting as systematic model drift. Rather than propose another standalone benchmark, it organises established benchmarks into a capability taxonomy across three dimensions — CAN (latent competence), WILL (default behavioural preferences) and HOW (protocol compliance) — and reports relative change from the out-of-the-box checkpoint rather than absolute scores. The empirical study covers legal and medical domains across seven LLaMA, Qwen and Gemma models up to 80B parameters: instruction fine-tuning induces the strongest drift, preference optimisation is more conservative and can partially recover lost capability, differences between model families persist, and no universal mitigation emerges.

Affiliations span Thomson Reuters Foundational Research, Tübingen AI Center, Helmholtz Munich, MCML/TU Munich and Imperial College London.

Tiny benchmarks are usually built model-centrically: keep the items that discriminated between existing models. That carries a large upfront cost, cannot handle a new benchmark cold, and rests on the fragile assumption that future models will fail like their predecessors. Scales++ inverts this to an item-centric criterion — select on the intrinsic cognitive demands of each sample. Upfront selection cost falls by over 18×, and a 0.5% subset of the Open LLM Leaderboard predicts full scores to 2.9% mean absolute error. This is what lets CapTrack run repeatedly across checkpoints on a modest budget.

Affiliations span Thomson Reuters Foundational Research, University of Oxford and Imperial College London.

@article{chen2026thomson,
  title   = {Thomson: Continual Learning of Frontier Models for SovereignAI},
  author  = {Chen, Shengzhuang and Parker, Jerrod and Bang, Yejin and
             Bean, Andrew M. and Seedat, Nabeel and Winzeck, Stefan and
             Glazko, Daniil and Zgraggen, Jannik and Yu, Fangyi and
             Arnott, Scott and Trautmann, Dietrich and Ciuffreda, Luca and
             Bonifazi, Guglielmo and Romano, Davide and Bell, Bradley and
             Fielding, Kirsty and Giofr{\`e}, Daniele and Zielund, Tom and
             Chatterjee, Ipshita and Ghantasala, Sneha Murthy and
             Nanreh, Manpreet and Scoville, John and Sakowicz, Maciej and
             Seifeddine, Wassim and Thede, Lukas and Schwarz, Jonathan Richard},
  journal = {arXiv preprint arXiv:2608.27147},
  year    = {2026},
  url     = {https://arxiv.org/abs/2608.27147}
}

@techreport{publicaiconstitution2026,
  title       = {The Public AI Constitution Project},
  author      = {Patriniche, Luca and Bell, Bradley and Trautmann, Dietrich and
                 Vucekovich, Nikola and Callinan, Zoe and Soni, Priyanka and
                 Williams, Isabel and Coyne, Aidan and Nanreh, Mira and
                 Fielding, Kirsty and Seifeddine, Wassim and Simon, Felix M. and
                 Bang, Yejin and Schwarz, Jonathan Richard},
  institution = {Imperial College London},
  year        = {2026},
  month       = {8},
  url         = {https://huggingface.co/spaces/tri-fair-lab/publications/blob/main/Public_AI_Constitution.pdf}
}

@techreport{snowdon2026,
  title       = {Cheap and Effective Re-Alignment of Frontier Models
                 through Capability-Preserving Model Steering},
  institution = {Imperial College London},
  year        = {2026},
  url         = {https://huggingface.co/spaces/tri-fair-lab/publications/blob/main/Frontier_Model_Realignment.pdf}
}

@inproceedings{vincent2026insufficiencybench,
  title     = {InsufficiencyBench: Evaluating LLM legal advice on
               underspecified user queries},
  author    = {Vincent, Samuel J. and Calloway, Daniel and Yu, Fangyi and
               Bean, Andrew M. and Seedat, Nabeel},
  booktitle = {Proceedings of the AI for Law Workshop at the International
               Conference on Machine Learning (ICML)},
  year      = {2026},
  note      = {Best Paper Honorable Mention. arXiv:2608.20220},
  url       = {https://arxiv.org/abs/2608.20220}
}

@inproceedings{bang2026contractscrub,
  title     = {ContractScrub: A benchmark for final review of legal contracts},
  author    = {Bang, Yejin and Fielding, Kirsty and Oliver, Brandan and
               Birke, Brian and Seedat, Nabeel and Bean, Andrew M.},
  booktitle = {Proceedings of the AI for Law Workshop at the International
               Conference on Machine Learning (ICML)},
  year      = {2026},
  note      = {arXiv:2608.20204},
  url       = {https://arxiv.org/abs/2608.20204}
}

@article{thede2026captrack,
  title   = {CapTrack: Multifaceted Evaluation of Forgetting in LLM Post-Training},
  author  = {Thede, Lukas and Winzeck, Stefan and Akata, Zeynep and
             Schwarz, Jonathan Richard},
  journal = {arXiv preprint arXiv:2603.06610},
  year    = {2026},
  url     = {https://arxiv.org/abs/2603.06610}
}

@article{bean2025scales,
  title   = {Scales++: Compute Efficient Evaluation Subset Selection with
             Cognitive Scales Embeddings},
  author  = {Bean, Andrew M. and Seedat, Nabeel and Chen, Shengzhuang and
             Schwarz, Jonathan Richard},
  journal = {arXiv preprint arXiv:2510.26384},
  year    = {2025},
  url     = {https://www.arxiv.org/abs/2510.26384}
}