🔍 Read the full analysis: Weight-Space AI Editing: How We Achieved Zero Legal Score At Home on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
A home experiment used directional ablation to reduce looping in the local model Ornith 1.5 35B-A3B. On the same setup, the edited model answered all 10 published Lexassay tasks, up from 4, but its legal score remained zero; the authors report 55 wrong or invented citations.
A home experiment reported that a direct edit to a local AI model’s weights stopped it from getting stuck in loops often enough to answer all 10 published Lexassay tasks, compared with four on the same setup before the edit. But the model’s legal score stayed at zero, and the answers included 55 wrong or invented citations, according to the experiment’s report.
The experiment tested Ornith 1.5 35B-A3B, a model its publisher designed for coding and agent tasks rather than legal work. The report says the model had scored 1 out of 100 on the broader benchmark and that two of its responses to the 10 published tasks did not finish, running into a 62,000-token output limit while repeating itself.
The editing method was directional ablation: the team ran the model, recorded internal signals, identified a direction associated with looping, then removed that direction at half strength from the output weights of 27 of the model’s 40 layers. The report says this involved no training, no new legal data and no AI reading the weights as text. That differs from the MIT researchers’ “weight-space programming” approach, in which a model reasons directly about another model’s weights.
On the 10 published tasks, the edited model answered 10 of 10, but met just 7 of 50 checked legal criteria, up from 3. The report counts 55 incorrect or fabricated citations among the edited answers. It also says that using the model card’s recommended sampling settings without editing the weights raised the unchanged model’s completion count to 9 of 10 on its original setup. Those results are not a direct comparison across identical setups.
Infographic 01 · What we changed
One loop direction, weakened at half strength
Inspired by the MIT post, but a simpler, classic technique (directional ablation): code written by a Claude Opus 5.5 session measured one loop direction by running the model and projected it out of the weights. No training, and no AI reading weights as text.
The six steps
- Write test material that is not the benchmark: 24 legal development questions plus 8 general controls, 0 overlap with benchmark prompts
- Find the true loop onset at token level: from there, at least 90% of the next 1,500 positions repeat earlier 32-token sequences
- Run the model and measure one loop direction per layer: average internal signal inside the loop minus the reasoning just before it
- Weaken it in the weights:
W ← W − α·r·(rᵀW)on the MLP output matrices of all 256 experts and the shared expert - Bake an edited copy; the original files stay untouched and the edited weights stay internal
- Benchmark original and edit on the same local setup, graded blind by three Claude agents; figures cover the 10 published tasks
Quick test: share of pure copying
Restart 20 tokens before a known loop onset, 8 cases, 3,000 tokens. Lower is better.
The cost of the edit
- Full strength: 23 of 24 legal questions finished, but control answers broke: tokens dropped out of short formulas, and month counting went wrong
- Half strength: 23 of 24 finished, 9 of 9 checked control facts right, and one German control question was answered in Chinese once
- Before the edit: 8 of 24 legal development questions did not finish; all 8 controls did
One run per setting. The loop detector was revised three times during the experiment, so the development numbers come from a moving yardstick. The direction is a statistical description, not a single “loop neuron”.
Sources: our experiment records, 4–5 Nov 2026 (executing Claude Opus 5.5 session; review by GPT-6 Astra in Codex); Ornith-1.5-35B-A3B model card.
Powered by Thorsten Meyer AI
Finishing Answers Is Not Legal Accuracy
The result illustrates a distinction that matters to anyone evaluating AI for legal work: completion is not correctness. A model that reaches the end of a response may still give unreliable legal analysis or cite authorities that do not support its claims. Here, the intervention improved the reported rate of completed answers on one setup, while the benchmark’s legal score remained zero.
The report’s benchmark checks citations against primary legal sources and covers questions involving German, U.S. and EU law. That makes citation errors more than a presentation flaw: users could mistake fluent, complete responses for legally grounded answers. The findings do not establish how the technique would perform on other models, tasks or hidden benchmark questions, but they show why a change in model behavior should be evaluated against the actual task, not just whether the system stops looping.
The authors also disclose that a Claude session performed the edit and Claude agents graded the answers. They say OpenAI’s GPT-6 Astra, running in Codex, independently reviewed code, figures and claims and identified faults. That review adds a separate check, but it does not make the experiment a second test or replace replication.
Popular right now
Philips 3300 Series Fully Automatic Espresso Machine with Milk Frother, Grinder, Hot and Iced Coffee Presets, SilentBrew, 15-Bar Pressure (E
Bose QuietComfort True Wireless Earbuds with Active Noise Cancellation, Bluetooth 5.3, USB-C Charging, IPX4, Black
Dyson Airwrap i.d. Multi-Styler for Straight and Wavy Hair
As an affiliate, we earn on qualifying purchases.
A Home Test of MIT’s Editing Idea
The experiment was prompted by a report from four MIT CSAIL researchers: Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab. Their work described “weight-space programming,” including a Qwen case in which GPT-6 Astra reportedly identified a direction linked to looping and edited the model’s weights. The MIT report said the Qwen model’s mathematics score rose from 7.9% to 48.1% after the intervention. Those results belong to the MIT experiment, not this home test.
The home experiment aimed to examine a related prediction: removing a behavior that prevents a model from completing a task may expose an existing capability, but should not supply knowledge the model lacks. Its method instead used model activations gathered by running Ornith and a computed direction removed from selected weights. The report characterizes this as a simpler, established approach, rather than a replication of the MIT method.
Lexassay asks models 15 questions about German, U.S. and EU law, without internet access or tools. The article’s figures cover only the 10 published tasks; five other tasks remain hidden. The authors say the runs, answers, grades and citation checks are public on Lexassay. Its updated November 9 figures refer to those 10 published tasks.
Limits of the Ten-Task Test
The report describes one run per setting, a small development set and a loop detector that changed during the experiment. It also says three graders came from the same model family. The independent GPT-6 Astra review found faults in the work, according to the report, but the article does not describe a repeat experiment that would establish whether the results are reproducible.
The unchanged model’s completion results also varied by setup: it answered 8 of 10 in a published run but 4 of 10 on the setup used for comparison with the edit. The report says the two local setups are not directly comparable. It is therefore unclear how much of the change in completed answers came from the weight edit, how much depended on the setup and sampling choices, and whether the result would persist across repeated runs. The article reports no improvement in the legal score.
Public Runs and Further Testing
The immediate next step is scrutiny of the public Lexassay runs, including the model’s answers, grades and citation checks, which the report says are available alongside a method page. Readers can compare the edited run with the unchanged-model runs while keeping the setup differences in view.
Further testing would need repeated runs under matched settings, a stable loop-detection method and evaluation beyond the 10 published tasks. The report does not announce a scheduled follow-up or a new model release. Until such evidence is available, the findings remain a single experiment about response completion, not evidence that weight editing makes the model reliable for legal advice.
Key Questions
What changed after the weight edit?
On the comparison setup, Ornith 1.5 answered all 10 published tasks, compared with four before the edit. The report says the legal score stayed at zero.
Did the edit improve the model’s legal knowledge?
The reported benchmark score did not improve: it remained zero. The edit was aimed at reducing looping, not adding legal information.
How was the model edited?
The experiment used directional ablation. The team ran the model to identify a direction associated with looping, then removed that direction at half strength from selected output weights. The report says it did not use new training or legal data.
Is this the same technique used by MIT researchers?
No. The report says the MIT work used a model to reason directly about another network’s weights. The home experiment used activations gathered by running the model and a computed direction removed from its weights.
Can the result show that the model is safe for legal use?
No. The report says the edited answers included 55 wrong or invented citations, and its legal score remained zero. It also describes a small test set and one run per setting, so the result does not establish legal reliability.
Source: ThorstenMeyerAI.com
