Bryan

Independent ML researcher

[Bryan @Branchez7]

Alignment-tuned LLMs often suppress factual log-probs (and therefore answers) on politically sensitive topics, even when the knowledge is still in the hidden states.

Just got final approval and went live on arXiv: I built a tiny ~786K-param post-transformer adapter that sits on top of any frozen model and easily corrects those suppressed probabilities. No full retraining required!
#LLM #AI #MachineLearning #Alignment
https://t.co/5iMKeNszJf

[dax @thdxr]

hey @OpenAI and @AnthropicAI i have a proposal

your models
but free

how about it?

[Bryan @Branchez7]

@gaborgurbacs If you can run the most expensive country in the world for a month with your wealth, that’s not right. I like Elon, but that’s not right.

[Andy @PositivFuturist]

@thechosenberg Meta runs a digital boomer retirement home

[Bryan @Branchez7]

@EricLDaugh Stupid

[Bryan @Branchez7]

@AlexKontorovich Just sharing in case it’s helpful to someone else. I don’t use Zulip, and I’m not a mathematician, but as far as I can tell it’s quality code.

[Bryan @Branchez7]

I gave it a try and after >200k loc I’m giving up for now. Made some good progress:14/24 strict-closed, the other 8 reduced to two named classical hypotheses (Forster Thm 16.9 for Item 14, Riemann bilinear + Abel + Jacobi inversion for the Jacobian cluster) with full chains proven and unconditional on the Riemann sphere. Walls are the same multi-month classical content mathlib doesn’t have. Repo: https://t.co/sb9ZW50XYs

[Bryan @Branchez7]

@jain_harshit It’s easy to fix. You just say, “I just slept great and I’m ready to work again” and it leaves you alone

[Bryan @Branchez7]

@sama Unsolved murders of OpenAI employees

[Bryan @Branchez7]

@elonmusk But humans can move their heads and eyes to avoid extreme glare

[Bryan @Branchez7]

@selectsand Will with a slight modification https://t.co/1hwkrCaUVL

[Bryan @Branchez7]

Alignment-tuned LLMs often suppress factual log-probs (and therefore answers) on politically sensitive topics, even when the knowledge is still in the hidden states.

Just got final approval and went live on arXiv: I built a tiny ~786K-param post-transformer adapter that sits on top of any frozen model and easily corrects those suppressed probabilities. No full retraining required!
#LLM #AI #MachineLearning #Alignment
https://t.co/5iMKeNszJf

[Bryan @Branchez7]

@Tono_Ken3 There’s a another option https://t.co/1hwkrCaUVL

[Bryan @Branchez7]

Alignment-tuned LLMs often suppress factual log-probs (and therefore answers) on politically sensitive topics, even when the knowledge is still in the hidden states.

Just got final approval and went live on arXiv: I built a tiny ~786K-param post-transformer adapter that sits on top of any frozen model and easily corrects those suppressed probabilities. No full retraining required!
#LLM #AI #MachineLearning #Alignment
https://t.co/5iMKeNszJf

[Bryan @Branchez7]

@HowToAI_ Or just do this https://t.co/1hwkrCaUVL

[Bryan @Branchez7]

Alignment-tuned LLMs often suppress factual log-probs (and therefore answers) on politically sensitive topics, even when the knowledge is still in the hidden states.

Just got final approval and went live on arXiv: I built a tiny ~786K-param post-transformer adapter that sits on top of any frozen model and easily corrects those suppressed probabilities. No full retraining required!
#LLM #AI #MachineLearning #Alignment
https://t.co/5iMKeNszJf

[Bryan @Branchez7]

@paraschopra At this point, people shouldn’t be having to correct SOTA models with something this obvious. Nobody wants their writing to look like a.i. wrote it

[Bryan @Branchez7]

@elonmusk It feels like the government is holding us hostage to get more control and tax money (in the form of debt and loss of rights) before they'll give us air travel back.

[Bryan @Branchez7]

Love this! I’m doing similar on physics discovery with a margin-oracle critic + adapter repair loop — repo here: https://t.co/Wnla1aIwo9

Frozen base LLMs act as STEM truth oracles via log-prob ranking and accuracy scales from 16% (GPT-2) to 77% (Qwen3-4B-Base) on a 97-fact benchmark. Four scale-invariant biases persist across models; a mixed adapter recovers 80% accuracy, and the margin is a perfect binary oracle (negative predicts incorrect 0% error, positive predicts correct 0% error).

I’m starting to thing these methods are going to result in a lot of new discoveries 🍻

[Bryan @Branchez7]

This was a lot of work and was built on 10 papers worth of experiments:

1. Rho-Guided Supervised Fine-Tuning: Post-Training Repair of Calibration Damage in Large Language Models
DOI: 10.5281/zenodo.18854943
2. Behavioral Entanglement in Transformers: Grassmann Geometry of Rho-Guided SFT
DOI: 10.5281/zenodo.18865861
3. Behavioral Phase Transitions: Geometric Scaffolding Precedes Behavioral Emergence
DOI: 10.5281/zenodo.18865198
4. Confidence Cartography: Teacher-Forced Probability as a False-Belief Sensor
DOI: 10.5281/zenodo.18703505
5. CF90: Knowledge-Preserving SVD Compression for LLMs
DOI: 10.5281/zenodo.18718545
6. Contrastive Pretraining Teaches Format Generation, Not Behavioral Knowledge
DOI: 10.5281/zenodo.18870555
7. The Expression Bottleneck: 41% Universal Constant and the Generation Mechanism (also titled “Small Language Models Already Know More Than They Can Say”)
DOI: 10.5281/zenodo.18895248
8. Snap-On Communication Modules: Logit-Space Adapters That Preserve Base Model Knowledge
DOI: 10.5281/zenodo.18902617
9. STEM Truth Oracle: Log-Probability Ranking Reveals and Corrects Scale-Invariant Factual Biases (the core oracle paper that underpins the frozen priors and margin critic)
DOI: 10.5281/zenodo.19005729
10. Breaking Frozen Priors: Teaching Language Models to Discover Conservation Laws from Numerical Simulation (the latest manuscript you just shared, which builds directly on the above)
DOI: (not yet assigned in the document, but once uploaded to Zenodo/arXiv it will get one — e.g., 10.5281/zenodo.[new ID] if Zenodo, or arXiv:26xx.xxxxx if pre-print)

@grok summarize this research and why it works

[Bryan @Branchez7]

Current autoresearch agents are mostly stochastic parrots with a search tool. They are bottlenecked by the model’s internal priors. If the LLM hasn’t seen the physics, the agent won't propose the discovery.

I’ve been working on NoetherSolve to flip this. Instead of asking the model for answers, we use it as a diagnostic tool to find where its world model breaks.

The loop is simple but effective:
1. Generate symbolic candidates for conserved quantities.
2. Verify them against a ground-truth numerical integrator.
3. If the math holds but the model's log-probs are low, you've found a genuine knowledge gap.

I am essentially using Noether's theorem to find the "blind spots" in frozen weights, then closing them with targeted adapters. I tried to make this easy to use for others, but unfortunately, I’m strong on theory and weak on coding so sorry if there are some hiccups.

Early results on vortex pair conservation are interesting. If you want to poke at the code or claim a domain for discovery:
https://t.co/wCXyddkkSI

[Bryan @Branchez7]

@marmikch It’ll be active again once the problems get hard enough. Every new technology has an explosive phase, where solving problems was easy. When I was young, buying domain names of every business in the neighborhood could turn you into a rich person.