When we announced Thomson’s benchmark results, we said the model was competitive with the strongest frontier models at a fraction of their size and cost. That post was about what Thomson is capable of as of today. This is the story of how we got it there.
Thomson began as an internal project, built to solve a problem we had ourselves.
Thomson Reuters holds 175 years of authoritative data across legal, news, tax and accounting: Westlaw, Practical Law, Checkpoint and Reuters. We also employ thousands of subject-matter experts whose working lives are spent deciding what is correct. For three years we watched general-purpose models improve rapidly while both assets sat outside the training loop. We also faced the questions our customers were asking us: what dependency are we accepting on someone else’s architecture and pricing, and what do we do when the capability we need most is on nobody’s roadmap?
Our answer to that was the Thomson LLM, and it worked well enough that we now want to share it with the rest of the world wrestling with these same questions.
Where the argument came from
The team that built Thomson did not arrive at Thomson Reuters with just a view about legal AI, but with a view about reliability.
Safe Sign Technologies was founded in 2022 by lawyers and researchers whose background was in model safety, robustness and reliability, several coming out of applied AI in medicine and law, from Harvard and Cambridge. Medicine and law share a property most application domains do not: being nearly right is still wrong, and the cost of a confident error is borne by someone other than the person who made it. Both have long and demanding traditions of rigorous verification as a result, and those shaped how we approached the problem.
The argument we made from that starting point was, at the time and until recently, unfashionable. In 2022 and 2023 the field was watching capability curves. The consensus was that frontier models would absorb professional work as a by-product of getting cleverer and that any attempt to keep pace with the “scaling laws” of AI was futile. On that view the sensible move for a small company was to build a layer on top and wait.
We believed the binding constraint was different. Capability, we argued, would become abundant; it was the object of enormous and well-funded competition, and there was no reason to expect it to stay scarce. What would remain scarce was trust and reliability: being right in a way that can be checked, in a domain where someone whose career depends on it. Trust is not a by-product of capability. It is a separate research problem requiring different evidence, and nobody was going to solve it for law as a side effect of solving it for everything.
Very few people agreed. Making that case repeatedly, to investors and to ourselves, through pivots and long stretches with nothing to point at, was most of the job.
Thomson Reuters acquired Safe Sign in August 2024, in the company’s first pre-revenue acquisition. The team became Thomson Reuters’ Foundational Research team, and crucially, the research posture that pre-dated the acquisition survived the transition. We continued to treat the work as a research problem rather than solely a product problem, which is why so much of the effort below went into measurement.
Starting from open weights
I said previously that our starting hypothesis was that capability would be abundant. The rate of progress of open-source AI has continued to prove this thesis over the last several years. Thomson benefits from this directly, with a leading open-weight foundation model as its starting point. We’ve changed the root model of Thomson many times over the last several years, and will continue to do so as the frontier of open-weight models evolves. This is a tide that Thomson moves with, not one that washes it away.
At the time of writing, the base model for Thomson is the Imperial College London Snowdon model. This model was developed by the FAIR Lab at Imperial, which Thomson Reuters and Imperial founded jointly, as an academic by-product of the acquisition of Safe Sign.
That choice is usually framed as a trade-off, and there is something to it: open-weight models can lag the closed frontier, and published analyses generally put that lag at a few months [1]. The conventional choice is, therefore, between capability and control.
We did not think this was an acceptable dilemma for professional work. The frontier is measured on general capability, but our customers are judged on something narrower: whether a citation holds up, whether an answer is complete, whether the reasoning survives a partner’s review. There is no rule that a model strong on the second must concede the first. As we reported at launch, Thomson performs competitively with the strongest frontier models on the market, including Claude Opus 4.8, and ahead of GPT-5.5, Claude Sonnet 5 and Gemini 3.1 Pro. It also leads them on the measure this post is concerned with: whether the citations in a research report survive being checked.
Turning the archive into training data
Thomson Reuters content is the deepest asset in this field and the reason a model of this kind was possible at all. It is also, as any archive of this scale would be, material that has to be prepared before a model can learn from it well.
Content has to be found, which in an organisation of this breadth and history is a substantial exercise in itself. It must be assessed for rights, selected for measurable impact on model performance rather than relevance in the abstract cleaned, structured, deduplicated, and finally deployed into a data mixture, which is where the most consequential decisions are made.
To date we have used less than ten per cent of Thomson Reuters content in continued pre-training. Westlaw, Practical Law, Checkpoint and Reuters News have been drawn on selectively. The areas where the model is not yet best in class are not ceilings we have reached, but areas where the relevant content has not yet been brought to bear.
The specialisation problem
Data mixture matters so much because specialising a model can damage it. Fine-tuning on domain-specific data can cause catastrophic forgetting: the model overwrites capabilities acquired during pre-training and its general performance degrades [2]. The effect is well documented, and mitigations exist, including replay of general data and regularisation of parameter updates. None fully solve it.
One finding matters more than the others for our purposes. Kotha, Springer and Raghunathan’s 2024 study [3] examining what degrades during domain fine-tuning identified instruction-following as the principal contributor to forgetting: what erodes first is not the model’s knowledge of the world but its ability to do as it is told. For professional work that is close to a worst case, since real legal work is never only legal reasoning but legal reasoning while adhering to a format, a jurisdiction, a house style, an exclusion, a client’s standing preference. Output that requires reworking has not saved anyone any time.
We therefore treated general capability retention as a first-class training objective rather than an acceptable loss. Instruction following is among the capabilities specialisation is most likely to erode, and it is one of the categories in our published benchmark results where Thomson stands up best against the frontier models, scoring 0.914 ahead of Claude Opus 4.8, Gemini 3.1 Pro and GPT-5.5. That is the clearest evidence we have that the model was specialised without being narrowed.
Where the expertise actually comes from
Many organisations claim their AI systems are “trained with expert input”. The phrase carries little meaning without an answer to the real question: how does a lawyer’s judgement become a training signal? Experts do not produce training data, but a standard. We have had to work to ensure the collective edge in expertise held by Thomson Reuters domain experts is realised in the quality of our training data. This is how we did it.
Rubrics at maximum complexity. Partner-level practitioners worked full-time for months constructing evaluation rubrics for the hardest legal research tasks we could specify: the kind of multi-jurisdictional question where a good answer has fifteen necessary components and a plausible-looking one has nine. Each rubric enumerates what a correct response must contain. This is slow, expensive, and cannot be crowdsourced or synthesised.
Commercial judgement, not only legal judgement. We required lawyers fresh out of commercial practice to ground the training data in what clients actually care about, which is frequently not what a textbook would emphasise. Take an indemnity. In most commercial agreements, it is heavily negotiated and often enforced, and treating it as significant is correct. But in an NDA it is usually neither, and almost never the crux. A model trained only on doctrine cannot tell those situations apart, and one that flags an NDA indemnity as urgently as the confidentiality carve-outs has identified a legal issue and wasted a lawyer’s attention. Teaching that distinction requires people who have sat on the other side of the negotiation.
Preference data at scale. Thousands of hours of qualified lawyer time selecting between model outputs against complex criteria. Not “which is better” but which better serves a client with a particular posture, in a particular jurisdiction, at a particular stage of a matter.
Thomson Reuters employs around 1,500 attorney-editors whose day job is producing the analytical content lawyers rely on. The obvious move is to train on their published output. The harder and more valuable move is to capture what happens between the first draft and the published article: the judgement calls, the discarded framings, the reasons a proposition was narrowed, the authority considered and rejected. That intermediate work is where so much expertise lives, and it is almost never written down.
This problem generalises directly to our customers. A firm’s advantage is not simply its precedent bank: precedents circulate, deals become public, documents get shared. The advantage is what years of doing the work have built in the minds of its lawyers, who eventually retire or move. Capturing the reasoning rather than the artefact is the same problem, and we have worked on it at scale on our own corpus first.
Internal deployment as a research instrument. Thomson has been deployed widely inside Thomson Reuters, with thousands of domain experts using it on their hardest problems, which gives us failure modes reported by people qualified to diagnose them. Our teams are not incentivised to use Thomson for Thomson’s sake; if they use it, it is because they have decided it can do something others can’t.
The consistency problem
Expertise does not straightforwardly produce consistency. In some respects it produces the opposite: the more experienced the practitioner, the more nuanced their judgement, which is exactly what you want in a partner and exactly what creates noise in a training set. Two excellent lawyers can disagree on a scoring decision not because either is wrong but because each applies a refined intuition the other does not share.
The literature bears this out. On the LEXam legal reasoning benchmark [4], three legal experts independently scoring the same answers on a ten-point scale reached a quadratic weighted kappa of 0.49, with a mean absolute deviation approaching two points. Work on implicit legal citations [5] reports similar or worse agreement. More troubling, Rehag’s survey of legal machine learning datasets found they systematically removed all traces of disagreement rather than treating conflicting expert annotations as informative [6].
Take expert output at face value and train on it, and you teach the model an averaged version of several incompatible standards: vaguely acceptable to everyone rather than correct according to anyone.
A large share of our effort therefore went into data quality: calibrating annotators against worked examples, measuring agreement continuously and treating drops as signals about the task specification rather than the annotator, and structuring rubrics tightly enough that disagreement surfaces as genuine ambiguity rather than noise. Where it persists, the question is usually contested, which is itself something the model should learn.
Safety, values and red-teaming
A dedicated team of lawyers worked on bias, political neutrality and toxic behaviour, with extensive human and automated red-teaming. We treat these as training objectives rather than output filters: a filter catches a bad answer on the way out, an objective changes what the model is disposed to produce.
Political neutrality deserves particular mention given that Reuters sits inside this company. Realignment towards factuality and pluralism was an explicit part of the training programme rather than a compliance exercise appended to it, and it is measured rather than asserted: Thomson performs strongly against the frontier models on our internal neutrality evaluation, with detail to follow in the technical report. For a company that publishes news as well as legal analysis, that is not peripheral.
Beyond legal data
Not all of the training data is legal. We drew on domains rich in explicit chain-of-thought reasoning, where the reasoning must be set out rather than left implicit, on the view that a model reasoning well in structured non-legal settings reasons better in legal ones. Checkpoint and Reuters give depth in tax, accounting and world events, because legal work is rarely purely legal.
Rigour and factuality
Our own lawyers publish at leading AI conferences [7] the people building the evaluation apparatus treat it as research rather than quality assurance, which is what makes it rigorous enough to train against.
We think that rigour produces the result we care most about. In our published deep research evaluation, Thomson working over Westlaw and Practical Law scored 0.83 on factuality against 0.65 and 0.68 for leading frontier models given unrestricted access to the open web. Completeness was close between all three. Factuality was not.
Completeness is a capability measure: it asks whether the system covered the ground, and on it the three were nearly level, because frontier models with the open web and enough time will generally find the material. Factuality is a reliability measure: it asks whether the system can be checked and survive it, and on that they were not close at all.
The metric is not a measure of whether an answer sounds authoritative or whether the conclusions are broadly sound. Every claim is extracted and matched against the source cited for it, and the score is the proportion of assertions whose own citations hold up when checked.
That is the failure that has kept general-purpose AI in the assistant’s chair. A system reliably right about its own sources is a different category of instrument from one merely fluent about them. It is the difference between something an associate uses and something a partner signs.
We call this Fiduciary-Grade AI: a standard for AI used where accuracy, accountability and trust are not optional, for professionals working under duties of care and regulatory oversight. Thomson demonstrates that it can be pursued at the model layer rather than bolted on above it.
Capability and sovereignty are not mutually exclusive
The lesson is not just about law. Any organisation holding a deep proprietary corpus and real domain expertise has been told it must choose: either rent frontier capability and accept the dependency or own an open model and accept some distance from the frontier. The choice is false, provided you are willing to do the unglamorous work of preparing the data, converting expert judgement into consistent signal, and building the evaluations before you try to move the numbers. The reward is a model you own rather than rent, pointed at the problems you choose, improving on your schedule rather than somebody else’s.
The compute is not the binding constraint. The corpus and the people who know what correct looks like within it are, and those have never been concentrated in the frontier laboratories. They sit inside institutions that spent a century accumulating them without thinking of themselves as AI companies.
We are one of those institutions. Thomson is what happened when a research team that had spent three years arguing trust and reliability were the scarce input finally got access to the data and the experts to prove it.
Thomson enters production this month powering CoCounsel skills including high-volume structured document review, with integration across the legal and tax portfolio to follow. A full technical report is forthcoming. Thomson was built in collaboration with DatologyAI, Lambda, Together AI, Imperial College London and the Thomson Reuters–Imperial Frontier AI Research Lab.
Sources for external claims
[1] Epoch AI, open-weight capability lag analyses (October 2025; May 2026); Stanford AI Index 2026.
[2] Luo et al., “An Empirical Study of Catastrophic Forgetting in Large Language Models” (2023); Song et al., arXiv:2501.13669.
[3] Kotha, Springer and Raghunathan, arXiv:2406.12227.
[4] LEXam, arXiv:2505.12864.
[5] “Where Experts Disagree, Models Fail”, arXiv:2603.22973 (2026).
[6] Rehaag, “I beg to differ”, Artificial Intelligence and Law (Springer, 2023).
[7] Yejin Bang, Kirsty Fielding, Brandan Oliver, Brian Birke, Nabeel Seedat & Andrew M. Bean, ContractScrub: A Benchmark for Final Review of Legal Contracts (Thomson Reuters Foundational Research, 2026) (in Proceedings of the AI for Law Workshop at the International Conference on Machine Learning (ICML 2026), available at https://openreview.net/forum?id=Zw7vAZusSu); and Samuel J. Vincent, Daniel Calloway, Fangyi Yu, Andrew M. Bean & Nabeel Seedat, InsufficiencyBench: Evaluating LLM Legal Advice on Underspecified User Queries (Thomson Reuters Foundational Research, 2026) (in Proceedings of the AI for Law Workshop at the International Conference on Machine Learning (ICML 2026), available at https://openreview.net).
