A model trained to compress what it sees learns “from here, 6; from there, 9” and never learns the symmetry. That’s how world models end up breaking the rules of physics.
Latest
New from the lab
Papers, results and code from the past year. We post the long versions, with the plots, on LinkedIn.
Attention heads behave like a bank of position-aware derivatives, so how attention responds to moving, removing or changing part of a prompt can be computed exactly. With Julie Huang, Maggie Chlon and Gregory Gutin.
Neural surrogates for DFT are usually tuned as if energy and force accuracy compete. Sweeping the weighting on MD17 ethanol and aspirin, they improved together across almost the entire range: adding force supervision cut energy error by 65%.
Hallucination as a measurable shortfall of information, with a rule for when to answer and when to abstain. Building on it, the open-source gate in ntkmirror let a 4B model beat a 14B model on the same benchmarks by staying silent instead of guessing.
On Qwen2.5 7B and 14B, the gap to an ideal Bayesian is hundredths of a bit. The order sensitivity comes from the positional encoding, not the reasoning: transformers trained from scratch without one are exactly order-blind.
A band-gap model’s predictions swung across symmetry-equivalent views of the same perovskite crystal. A student distilled with Mezzanine was stable, and more accurate.
A toolkit for distilling symmetry-respecting models. One example: a molecular-dynamics workflow that took 120,000 steps on an A100, distilled into a model small enough for a phone that recovers the same phase-level inference in a single forward pass.
Each starts with a failure you can see in today’s models. We look for the maths that predicts it, then test the fix in public, and say so when it doesn’t work.
01
World models that respect physics
Can a model trained on data alone learn the symmetries of the world? Under a log-loss objective, breaking a symmetry is often the cheaper way to compress the data, so more data and compute don’t fix it. We measure where models break symmetries and distil students that don’t.
From the explainer video for our robot world-model paper: a plain I-JEPA world model (left) next to Karim et al.'s symmetry-marginalised action space (right).Watch it on LinkedIn →
More in motion
Robot world modelsThe same actions, written as joint targets or as deltas, produce different predictions. Our 2026 preprint with Ahmed Karim shows it, and proposes a symmetry-marginalised action space as the fix.
MezzanineAverages a teacher’s predictions over symmetry-equivalent inputs, then distils a student that gives that average in one pass. The spread across views, the “warrant gap”, measures how much the teacher was being swayed.
Molecular dynamics in one passA Lennard-Jones simulation that took 120,000 steps on an A100, distilled into a symmetry-stable model small enough for a phone, which recovers the same phase-level inference in one pass.
A patch, not a cureYou have to know the symmetry, and no general fix exists. Mezzanine helps when the variation that shouldn’t matter is real and compact, and we report where it doesn’t.
PrincipleDistil the expectation, not a single view
Where can careful, checkable AI make a difference? We test our methods on chemistry, materials and biology, where a confident wrong answer has a real cost.
Eight models per molecule, with the energy/force weighting swept from 0 to 1: on MD17 ethanol and aspirin, energy and force error fall together.
More in motion
Neural surrogates for DFTA model has to predict both a structure’s energy and the force on every atom. Sweeping the weighting on MD17 ethanol and aspirin showed no trade-off across almost the entire range: adding force supervision cut energy error by 65%. Each structure has one energy but 3N force components, so the forces do most of the work.
MaterialsBand-gap predictions for a perovskite swung across symmetry-equivalent views of the same crystal. A student distilled on orbit-averaged labels was stable and more accurate.
Cancer multi-omicsBenchmarks on TCGA tumour data where AI proposes constrained predictions for missing values, checked against negative controls. On hold for now.
Hallucinations aren’t random bugs. They’re compression failures: when a model doesn’t hold enough information to reconstruct a rare fact, it fills the gap with something plausible. Our ICML 2026 paper turns that into an information budget for evidence-grounded yes/no questions, with a rule for when a model should answer and when it should abstain.
The gate in action: with too few bits (ISR 0.44) the model says it doesn’t know; once the evidence carries enough (ISR 1.35), it answers.Slide 6 of our AAIF talk →
Predictable Compression FailuresICML 2026. Information budgets that flag a likely hallucination before the model writes anything. We’re still testing how strongly the order effects hold on the newest models.
BerryOur open-source verifier, formerly Strawberry. It checks each claim against the evidence cited for it, inside Claude Code, Cursor, Codex or Gemini CLI. 1.7k GitHub stars; featured at NVIDIA GTC 2026.
Knowing when to stay quietBuilding on the paper, the open-source gate in ntkmirror let a 4B model beat a 14B model on the same benchmarks by abstaining instead of guessing. Depending on the model, it raised the grounded share of shipped claims from 50% to ~75–90%, at the cost of dropping ~10–20% of good claims.
Grounded claims shipped, with the gate50% → ~75–90%
What is a model really computing when it learns from examples in a prompt? Averaged over the order of the examples, language models come close to ideal Bayesian reasoning. Any single ordering can drift from it, and we’ve traced the drift to the positional encoding.
Same claim, same four pieces of evidence, asked in four orders. With positional encoding the answers swing from 0.18 to 0.93; without it, every order gives the same answer.Slide 5 of our AAIF talk →
Bayesian in expectationOn Qwen2.5 7B and 14B the gap to an ideal Bayesian is hundredths of a bit. Train a transformer without positional encoding and it is exactly order-blind; add one back and order sensitivity jumps by eight orders of magnitude.
Exact attention responsesA closed-form account of how RoPE attention responds when part of the input is moved, removed or changed. On decoder-only LLMs it called the direction of 95–97% of 92,160 positional edits.
A ledger for every tokenTrace an answer token by token: how much each cached token shifted or corrected the output. Next, we’re turning it into a bound for in-context learning.
Fine-tuning is expensive and makes models forget what they knew. One gradient step has an exact forward-pass dual, so a frozen model can be adapted by a small controller that never touches its weights.
A frozen model experiments on itself: it scores 109,080 candidate edits to how it reads its own prompt, picks the one that re-weights the note it had ignored, and goes from 0 to 6 of 6 tests passed.Slide 10 of our AAIF talk →
ntkmirrorOn Qwen2.5-7B, a 50,000-parameter controller matched LoRA’s accuracy on GSM8K after 22 seconds of fitting, and forgot 3.4× less on held-out code generation.
Controllers that add upTwo controllers fitted separately compose by addition with near-zero interference. The equivalent LoRA merge cost 17% on GSM8K.
Cheaper inference, with guaranteesDeciding, prompt by prompt, how much of a model’s memory can be dropped within a set error budget, and 4-bit quantisation that loses less by rearranging weights exactly.
The symmetry-breaking representation wins on log loss because it’s cheaper than representing the genuine invariance. The optimizer is doing exactly what you asked it to.
”
On world models and symmetryHassana Labs, 2026
Papers & preprints
Read the work
Everything we publish is free to read. Student researchers who’ve worked with us are co-authors.
The maths only matters if people can use it, so we release the code.
01 · HALLUCINATION
Berry
1.7k GitHub stars · formerly Strawberry
A local MCP server that checks whether an AI’s answer is supported by the evidence it cites. For each claim it measures how much the answer depends on that evidence, flags claims that don’t clear the bar, and keeps a tamper-evident record of every decision.
A research toolkit for measuring how much a model’s predictions swing across symmetry-equivalent inputs, then distilling a student that averages the swing away. Comes with recipes for robotics, weather, particle physics and fast physics surrogates.
Training-Free Self-Improving Agents: what a frozen model can still do to improve itself, counted in bits. At the Agentic AI Foundation’s London event, alongside speakers from Google DeepMind, Imperial, Prolific and Vercel. 687 people registered.
HallBayes, now Berry, passed 1,000 GitHub stars in under a month (Sep 2025).
By February 2026, thousands of people were using our hallucination checker.
One of our open-source releases raised £500 for a UNICEF fundraiser for Sudan (Dec 2025).
Why “Hassana”?
“My grandma never went to high school, but she taught me that learning has no gates.”
Hassana Labs is named after our founder’s grandmother. She loved learning, taught him mathematics, and made everyone around her feel valued. The lab carries that on. Talent is everywhere, but opportunity isn’t: science shouldn’t be gatekept, and nobody’s ambition should end because they were born without money or connections.
In 2025 we called for open contributions from underrepresented researchers around the world. Together we wrote papers that changed how we think about hallucination, and built open-source tools with more than 2,000 GitHub stars.
Founder
Leon Chlon, PhD is AI Research Lead at PhysicsX. He was previously a Visiting Fellow at the University of Oxford (Torr Vision Group) and a postdoctoral fellow at Harvard Medical School and a research affiliate at MIT, and did his PhD in machine learning at the University of Cambridge (Pembroke College). He has worked on machine learning at Meta, Uber/Careem and McKinsey, and wrote a free textbook, Information Geometry for Generative Models.
Student researchers who worked with us, now at UCL and the University of Barcelona
3+
Countries where we’ve given careers workshops: the UK, Lebanon and Thailand
Thousands
Helped land new jobs and master’s and PhD offers through our workshops and coaching
Our work has been supported by Survival and Flourishing Corp, Microsoft and Google.
Research fellowships
Opening doors for overlooked talent.
Applications closed for now
Since 2025 we’ve run remote research fellowships for talented people from marginalised communities: real research problems, full compute access and co-authorship on published papers. Fellowships are unpaid.
People who’ve researched with us have co-authored a paper at ICML 2026 and our 2026 preprints.
Student researchers who’ve worked with us have gone on to PhDs at UCL and the University of Barcelona.
Past tracks covered prompt-injection auditing, hallucination in chain-of-thought reasoning, and cancer multi-omics.
We’ll announce the next round here. Questions in the meantime? Email lc574@cantab.ac.uk.
Research updates
Get the next paper in your inbox.
New research, open-source releases and the next fellowship round, by email.
Contact
Work with us.
We’d love to hear from researchers who want to collaborate, funders who want to support open research, and journalists.
Not right now. Applications are closed for every track. We’ll announce the next round on this page.
Can I use your tools?
Yes. Berry, Mezzanine, ntkmirror and triton-blackwell are all on GitHub. Check each repository for its licence.
Do you publish results that didn’t work?
Yes. Negative results tell you where a method stops working, so we report them alongside the positive ones. Mezzanine, for example, is a patch rather than a cure: it only helps when you know which symmetry a model is breaking.
Build on this work.
Read the papers, use the tools, or get in touch about collaborating. Talent is everywhere — opportunity isn’t.