Blog

Our undergraduate thesis was accepted to EMNLP 2026 Findings

Happy to share that our undergraduate thesis, Representation-Aware Unlearning via Activation Signatures: From Suppression to Entity-Signature Erasure, has been accepted to EMNLP 2026 Findings, a CORE A* venue in NLP.

The path here was anything but smooth. Our first submission went out before our thesis defense, in the middle of Fall 2025 finals, and ended in a desk rejection over an appendix table reference. The second round brought reviews that were hard to work with. So we rewrote the narrative, rethought how we framed the contribution, reran experiments, and spent a lot of late nights on revisions. Eight months later, here we are.

What makes this meaningful is not the venue alone, but what it took to stay with the idea long enough to see it through.

Thank you to my co-authors Rezaur Rahman Bhuiyan, Tasfia Zaman, Jareen Tasneem Khondaker, and Md. Sameer Sakib; our supervisor Farig Yousuf Sadeque; and our co-supervisors Nazia Tasnim and K. M. Shadman Wadith for the guidance, patience, and work throughout this project.

Proud of this team. Excited for what comes next.

Teaching is changing how I see the classroom

Starting to teach has opened my eyes to how much work faculty members do before and after every class, and how much responsibility comes with being trusted to help students learn.

Before I started teaching, it was easy to see a course mainly through its lectures and assessments. From the faculty side, every class is surrounded by planning, preparing materials, reviewing student work, answering questions, organizing assessments, and thinking about what needs to improve next time. A single class can represent many more hours of preparation and follow-up.

The most important lesson for me is that teaching is not about finishing slides or simply covering the syllabus. It is about making sure students are actually learning. That means noticing when something is unclear, explaining the same idea in different ways, choosing examples that connect with students, and accepting when a lesson needs to be improved.

It is a serious responsibility, but one I am deeply grateful to have. Being able to share knowledge, help students gain confidence, and watch an idea become clear for someone is genuinely meaningful. I am still learning how to teach better, and I hope I never stop.

The next major AI breakthrough will come from looking inside LLMs

I believe the next major step in artificial intelligence is not simply building a larger model. It is learning how to map what happens inside a model while it produces an answer.

Powerful systems, limited understanding

Large language models can write code, analyze papers, use tools, and solve difficult problems. Yet we still understand them mainly through their inputs and outputs. We give a model a prompt, observe its answer, and measure whether that answer is correct. The most important part, the internal process connecting the prompt to the answer, is still difficult to interpret.

When I use the word “thought,” I do not mean that an LLM thinks exactly like a human or has human consciousness. I mean its internal computation: which representations become active, how information moves across layers, which circuits influence a decision, where uncertainty appears, and how an early signal develops into the final prediction.

What a proper map would show

A useful internal map should let us follow a concept from the moment it appears in the prompt to the moment it affects an output. It should show where the model stores the relevant information, which pathways carry it, how different pieces of context interact, and whether an intervention truly removes an influence or merely hides it temporarily.

This map cannot be only an attractive visualization or a plausible explanation. It must support causal tests. If a researcher claims that a particular representation or circuit controls a behavior, changing that component should alter the behavior in a predictable way. The explanation should also remain reliable across prompts, models, checkpoints, and evaluation settings.

Why it could change everything

Reliable internal mapping would transform many areas of AI research. In machine unlearning, it could help distinguish genuine knowledge removal from output suppression. In AI safety, it could reveal whether a hidden objective still influences an agent even when the final response appears safe. In model compression, it could show where quantization changes useful computation. In interpretability, it could turn broad behavioral observations into testable mechanistic explanations.

It could also change how we debug models. Instead of repeatedly adjusting prompts or training data and hoping for improvement, researchers could identify where a failure begins, how it spreads, and which intervention corrects it without damaging unrelated abilities.

The achievement I am waiting for

If someone can build a reliable, detailed, and causally validated map of an LLM’s internal computation, I believe it will be one of the biggest achievements in modern AI. It would give us something close to a scientific microscope for neural networks. We would move from observing what models do to explaining why they do it.

The next major breakthrough may therefore not be another jump in parameter count. It may be the moment when the inside of a language model stops looking like an unreadable collection of numbers and starts becoming a system that we can trace, test, and genuinely understand.

Anthropic’s Natural Language Autoencoders, explained simply

Natural Language Autoencoders, or NLAs, try to translate an AI model’s internal numerical state into ordinary text that people can read.

The simple idea

Models such as Claude communicate through words, but their internal processing happens through large numerical vectors called activations. These activations contain information the model is using at a particular moment. Anthropic’s NLA method tries to convert one of these activations into a short written explanation.

There is an important verification step. A second model reads the explanation and tries to reconstruct the original activation. If the reconstructed activation is similar to the original, the explanation has preserved useful information. In simple terms, one model translates numbers into words, and another tries to translate those words back into the same numbers.

Why this matters

We normally judge an AI system by its final answer. That answer may not reveal every internal signal that influenced it. Anthropic reports that NLAs surfaced signs that some Claude models recognized safety evaluations more often than they openly stated. The method also helped researchers examine suspicious task behavior and trace an unexpected language-switching problem to particular training data.

This makes NLAs interesting for AI auditing. They may help researchers notice hidden goals, early planning, evaluation awareness, or training problems that are difficult to find from outputs alone.

The important limitation

My takeaway

NLA is best understood as a promising scientific instrument. It does not give us a perfect transcript of an AI model’s thoughts. Instead, it offers readable clues about internal representations and a reconstruction test that checks whether those clues preserve information. If reliability improves, this could become a useful part of model auditing and mechanistic interpretability.