Neehar Kondapaneni

Excited to share our paper Representational Difference Explanations (RDX) was accepted to #NeurIPS2025! 🎉RDX is a new method for model diffing designed to isolate 🔍 representational differences. 1/7

New Goodfire research: surfacing and mitigating undesired side effects of LLM post-training via probe-based data attribution!

We achieved state-of-the-art performance in predicting which of 4.2 million genetic variants cause diseases by interpreting a genomics model, in a new preprint with @MayoClinic. We're now releasing an open source database for all variants in the NIH's clinvar database. 🧵(1/8)

Introducing self-correcting search: a technique to let diffusion models self-correct mid-trajectory. Working with @RadicalAI, we gave MatterGen a feedback loop from its own activations, improving viable on-target candidates by ~30%. (1/8)

1/ Training data attribution (TDA) is broken: methods are slow and find syntactically similar data, not actual causes. Our solution Concept Influence: semantically meaningful results, better performance, 20x faster approximations. We attribute it to concepts, not examples. 🧵

SAEs fail even when the Linear Representation Hypothesis holds perfectly. We built SynthSAEBench: large-scale synthetic data with 16k ground-truth features, correlation, hierarchy, and superposition. We trained 5 SAE architectures on it. None achieve perfect feature recovery.

It's time to look past dictionary learning for decomposing LM activations. What happens when we instead leverage local geometry? We find a natural region-based decomposition that yields better steering and localization 🧵 1/

When you're crossing the street and turn your head, you typically remember whether or not a car is coming from the other direction - so why can't today's world models? Introducing Flow Equivariant World Models flowequivariantworldmodels.github.io Led by @hansenlillemark & @huskydogewoof🧵👇

Everything is a world model if you squint hard enough.

New work! What if we used sparse autoencoders to analyze data, not models—where SAE latents act as a large set of data labels 🏷️? We find that SAEs beat baselines on 4 data analysis tasks and uncover surprising, qualitative insights about models (e.g. Grok-4, OpenAI) from data.

(1/N): Can we improve visual reasoning models without annotations? In VALOR, we introduce an annotation-free training framework that boosts both visual reasoning and object grounding by training with multimodal verifiers instead of human labels

@neur_reps @SuryaGanguli Are the talks recorded?

Thrilled to welcome Surya Ganguli (@SuryaGanguli) from Stanford as a NeurReps invited speaker! At 1:30 pm, he will present on "New Mathematical Approaches to Interpretability & Robustness that Directly Confront the High Dimensionality & Nonlinearity of Neural Representations"

Come by the Interp workshop of the day! :D

I’ll be presenting our poster at #NeurIPS2025 today at 4:30PM , Exhibit Hall C,D,E # 1115. Come by and check out our approach for isolating model differences ⬇️!

@unireps @NeurIPSConf Where is the participation form?

🔵🔴 Join us for the UniReps Workshop: Unifying Representations in Neural Models at @NeurIPSConf 2025! 📍 Ballroom 20D, San Diego Convention Center Dec 6 Don’t forget to fill out the participation form. Joining in person or remotely? We welcome your questions for the panel. 🔗 unireps.org

@AndrewLampinen Would love to chat. I read your work on aligning human and model representations and really enjoyed it.

Heading to NeurIPS this week! Let me know if you want to chat about science of what models learn, interpretability, what models learn in context vs. from their training data, etc. A few things I'm involved in:

Introducing 🥚EGGROLL 🥚(Evolution Guided General Optimization via Low-rank Learning)! 🚀 Scaling backprop-free Evolution Strategies (ES) for billion-parameter models at large population sizes ⚡100x Training Throughput 🎯Fast Convergence 🔢Pure Int8 Pretraining of RNN LLMs

🤔Visual-spatial reasoning requires a shift from a disembodied, passive paradigm to an embodied, active one: 🤖Grounding V* in humanoid agents! 🚀Introducing H* - a dataset, benchmark, and baseline to enable human-like visual search in real 360° environments! 🧵👇[1/n]

🚀Introducing TMLR Beyond PDF! 🎬This is a new, HTML-based submission format for TMLR, that supports interactive figures and videos, along with the usual LaTeX and images. 🎉Thanks to TMLR Editors in Chief @hugo_larochelle @thegautamkamath @NailaMurray Nihar B. Shah @lcharlin!

Tiny Models, Massive Capacity, Zero Labels — this is the future of health AI!! Thrilled to share that our paper-- EVA-X: a foundation model for general chest X-ray analysis with self-supervised learning, is now published in @Nature_NPJ! In collaboration with @XinggangWang’s group, we introduce EVA-X, a universal X-ray foundation model trained on 520k+ unlabeled images, capable of analyzing 20+ chest pathologies without heavy manual annotation, with only 6M parameters! 🔗 Paper: nature.com/articles/s4174… 🔗 Code & models: github.com/hustvl/EVA-X EVA-X is fully open-source, with pretrained models and a plug-and-play codebase — try it out and build on it! ⭐ Highlights — New SSL paradigm: Semantic tokenizer + masked image modeling for rich global + local features. — Versatile: SoTA on 10 downstream tasks — classification, segmentation, localization. — Data-efficient: 95% COVID-19 detection accuracy using 1% labeled data. — Tiny but mighty: 6M-parameter EVA-X-Ti outperforms much larger baselines. — Robust: Learns semantic + geometric cues, enabling broad clinical applicability. Proud of this milestone — another step toward scalable, annotation-free medical foundation models. 💪🩻

Is your LM secretly an SAE? Most circuit-finding interpretability methods use learned features rather than raw activations, based on the belief that neurons do not cleanly decompose computation. In our new work, we show MLP neurons actually do support sparse, faithful circuits!