Mechanistic Interpretability: Peeking Inside an LLM

Intro

find out how to look at and manipulate an LLM’s neural community. That is the subject of mechanistic interpretability analysis, and it could reply many thrilling questions.

Bear in mind: An LLM is a deep synthetic neural community, made up of neurons and weights that decide how strongly these neurons are linked. What makes a neural community arrive at its conclusion? How a lot of the knowledge it processes does it contemplate and analyze adequately?

These kinds of questions have been investigated in an enormous variety of publications not less than since deep neural networks began displaying promise. To be clear, mechanistic interpretability existed earlier than LLMs did, and was already an thrilling facet of Explainable AI analysis with earlier deep neural networks. For example, figuring out the salient options that set off a CNN to reach at a given object classification or automobile steering route may help us perceive how reliable and dependable the community is in safety-critical conditions.

However with LLMs, the subject actually took off, and have become far more fascinating. Are the human-like cognitive talents of LLMs actual or pretend? How does data journey by the neural community? Is there hidden data inside an LLM?

On this put up, you’ll find:

• A refresher on LLM structure
• An introduction to interpretability strategies
• Use circumstances
• A dialogue of previous analysis

In a follow-up article, we are going to take a look at Python code to use a few of these abilities, visualize the activations of the neural community and extra.

Refresher: The design of an LLM

For the aim of this text, we’d like a primary understanding of the spots within the neural community the place it’s price hooking into, to derive presumably helpful data within the course of. Due to this fact, this part is a fast reminder of the elements of an LLM.

LLMs use a sequence of enter tokens to foretell the subsequent token.

Tokenizer: Initially, sentences are segmented into tokens. The objective of the token vocabulary is to show steadily used sub-words into single tokens. Every token has a singular ID.

Nevertheless, tokens might be complicated and messy since they supply an inaccurate illustration of many issues, together with numbers and particular person characters. Asking an LLM to calculate or to depend letters is a fairly unfair factor to do. (With specialised embedding schemes, their efficiency can enhance [1].)

Embedding: A glance-up desk is used to assign every token ID to an embedding vector of a given dimensionality. The look-up desk is realized (i.e., derived throughout the neural community coaching), and tends to put co-occurring tokens nearer collectively within the embedding house. The dimensionality of the embedding vectors is a vital trade-off between the capabilities of LLMs and computing effort. Because the order of the tokens would in any other case not be obvious in subsequent steps, positional encoding is added to those embeddings. In rotary positional encoding, the cosine of the token place can be utilized. The embedding vectors of all enter tokens present the matrix that the LLM processes, the preliminary hidden states. Because the LLM operates with this matrix, which strikes by layers because the residual stream (additionally known as the hidden state or illustration house), it really works in latent house.

Modalities apart from textual content: LLMs can work with modalities apart from textual content. In these circumstances, the tokenizer and embedding are modified to accommodate totally different modalities, equivalent to sound or photos.

Transformer blocks: A lot of transformer blocks (dozens) refine the residual stream, including context and extra that means. Every transformer layer consists of an consideration part [2] and an MLP part. These elements are fed the normalized hidden state. The output is then added to the residual stream.

• Consideration: A number of consideration heads (additionally dozens) add weighted data from supply tokens to vacation spot tokens (within the residual stream). Every consideration head’s “nature” is parametrized by three realized matrices W_Q, W_Okay, W_V, which primarily resolve what the eye head is specialised on. Queries, keys and values are calculated by multiplying these matrices with the hidden states for all tokens. The eye weight are then computed for every vacation spot token from the softmax of the scaled dot merchandise of the question and the important thing vectors of the supply tokens. This consideration weight describes the power of the connection between the supply and the vacation spot for a given specialization of the eye head. Lastly, the top outputs a weighted sum of the supply token’s worth vectors, and all the top’s outputs are concatenated and handed by a realized output projection W_O.
• MLP: A completely linked feedforward community. This linear-nonlinear-linear operation is utilized independently at every place. MLP networks usually comprise a big share of the parameters in an LLM.
  MLP networks retailer a lot of the data. Later layers are likely to comprise extra semantic and fewer shallow data [3]. That is related when deciding the place to probe or intervene. (With some effort, these data representations might be modified in a educated LLM by weight modification [4] or residual stream intervention [5].)

Unembedding: The ultimate residual stream values are normalized and linearly mapped again to the vocabulary measurement to supply the logits for every enter token place. Usually, we solely want the prediction for the token following the final enter token, so we use that one. The softmax operate converts the logits for the ultimate place right into a likelihood distribution. One choice is then chosen from this distribution (e.g., the most definitely or a sampling-based choice) as the subsequent predicted token.

When you want to study extra about how LLMs work and acquire extra instinct, Stephen McAleese’s [6] clarification is superb.

Now that we seemed on the structure, the query to ask is: What do the intermittent states of the residual stream imply? How do they relate to the LLM’s output? Why does this work?

Introduction to interpretability strategies

Let’s check out our toolbox. Which elements will assist us reply our questions, and which strategies can we apply to research them? Our choices embody:

• Neurons:
  We might observe the activation of particular person neurons.
• Consideration:
  We might observe the output of particular person consideration heads in every layer.
  We might observe the queries, keys, values and a focus weights of every consideration head for every place and layer.
  We might observe the concatenated outputs of all consideration heads in every layer.
• MLP:
  We might observe the MLP output in every layer.
  We might observe the neural activations within the MLP networks.
  We might observe the LayerNorm imply/variance to trace scale, saturation and outliers.
• Residual stream:
  We might observe the residual stream at every place, in every layer.
  We might unembed the residual stream in intermediate layers, to look at what would occur if we stopped there — earlier layers typically yield extra shallow predictions. (This can be a helpful diagnostic, however not totally dependable — the unembedding mapping was educated for the ultimate layer.)

We are able to additionally derive extra data:

• Linear probes and classifiers: We are able to construct a system that classifies the recorded residual stream into one group or one other, or measures some function inside it.
• Gradient-based attributions: We are able to compute the gradient of a selected output with respect to some or all the neural values. The gradient magnitude signifies how delicate the prediction is to adjustments in these values.

All of this may be executed whereas a given, static LLM runs an inference on a given immediate or whereas we actively intervene:

• Comparability of a number of inferences: We are able to swap, practice, modify or change the LLM or have it course of totally different prompts, and report the aforementioned data.
• Ablation: We are able to zero out neurons, heads, MLP blocks or vectors within the residual stream and watch the way it impacts conduct. For instance, this permits us to measure the contribution of a head, neuron or pathway to token prediction.
• Steering: We are able to actively steer the LLM by changing or in any other case modifying activations within the residual stream.

Use circumstances

The interpretability strategies mentioned signify an enormous arsenal that may be utilized to many various use circumstances.

• Mannequin efficiency enchancment or conduct steering by activation steering: For example, along with a system immediate, a mannequin’s conduct might be steered in the direction of a sure trait or focus dynamically, with out altering the mannequin.
• Explainability: Strategies equivalent to steering vectors, sparse autoencoders, and circuit tracing can be utilized to grasp what the mannequin does and why primarily based on its activations.
• Security: Detecting and discouraging undesirable options throughout coaching or implementing run-time supervision to interrupt a mannequin that’s deviating. Detect new or dangerous capabilities.
• Drift detection: Throughout mannequin growth, you will need to perceive when a newly educated mannequin is behaving in another way and to what extent.
• Coaching enchancment: Understanding the contribution of facets of the mannequin’s conduct to its total efficiency optimizes mannequin growth. For instance, pointless Chain-of-Thought steps might be discouraged throughout coaching, which results in smaller, quicker, or doubtlessly extra highly effective fashions.
• Scientific and linguistic learnings: Use the fashions as an object to review to higher perceive AI, language acquisition and cognition.

LLM interpretability analysis

The sector of interpretability has steadily developed over the previous couple of years, answering thrilling questions alongside the best way. Simply three years in the past, it was unclear whether or not or not the learnings outlined beneath would manifest. This can be a temporary historical past of key insights:

• In-context studying and sample understanding: Throughout LLM coaching, some consideration heads acquire the potential to collaborate as sample identifiers, significantly enhancing an LLM’s in-context studying capabilities [7]. Thus, some facets of LLMs signify algorithms that allow capabilities relevant outdoors the house of the coaching information.
• World understanding: Do LLMs memorize all of their solutions, or do they perceive the content material with a view to type an inner psychological mannequin earlier than answering? This subject has been closely debated, and the primary convincing proof that LLMs create an inner world mannequin was printed on the finish of 2022. To show this, the researchers recovered the board state of the sport Othello from the residual stream [8, 9]. Many extra indications adopted swiftly. Area and time neurons had been recognized [10].
• Memorization or generalization: Do LLMs merely regurgitate what they’ve seen earlier than, or do they cause for themselves? The proof right here was considerably unclear [11]. Intuitively, smaller LLMs type smaller world fashions (i.e., in 2023, the proof for generalization was much less convincing than in 2025). Newer benchmarks [12, 13] purpose to restrict contamination with materials which may be inside a mannequin’s coaching information, and focus particularly on the generalization functionality. Their efficiency there’s nonetheless substantial.
  LLMs develop deeper generalization talents for some ideas throughout their coaching. To quantify this, indicators from interpretability strategies had been used [14].
• Superposition: Correctly educated neural networks compress data and algorithms into approximations. As a result of there are extra options than there are dimensions to point them, this ends in so-called superposition, the place polysemantic neurons might contribute to a number of options of a mannequin [15]. See Superposition: What Makes it Tough to Clarify Neural Community (Shuyang) for a proof of this phenomenon. Mainly, as a result of neurons act in a number of capabilities, decoding their activation might be ambiguous and troublesome. This can be a main cause why interpretability analysis focuses extra on the residual stream than on the activation of particular person, polysemantic neurons.
• Illustration engineering: Past floor info, equivalent to board states, house, and time, it’s doable to establish semantically significant vector instructions throughout the residual stream [16]. As soon as a route is recognized, it may be examined or modified. This can be utilized to establish or affect hidden behaviors, amongst different issues.
• Latent data: Do LLMs possess inner data that they maintain to themselves? They do, and strategies for locating latent data purpose to extract it [17, 18]. If a mannequin is aware of one thing that isn’t mirrored in its prediction output, that is extremely related to explainability and security. Makes an attempt have been made to audit such hidden targets, which might be inserted right into a mannequin inadvertently or purposely, for analysis functions [19].
• Steering: The residual stream might be manipulated with such a further activation vector to vary the mannequin’s conduct in a focused means [20]. To find out this steering vector, one can report the residual stream throughout two consecutive runs (inferences) with reverse prompts and subtract one from the opposite. For example, this could flip the model of the generated output from joyful to unhappy, or from protected to harmful. The activation vector is often injected right into a center layer of the neural community. Equally, a steering vector can be utilized to measure how strongly a mannequin responds in a given route.
  Steering strategies had been tried to scale back lies, hallucinations and different undesirable tendencies of LLMs. Nevertheless, it doesn’t at all times work reliably. Efforts have been made to develop measures of how effectively a mannequin might be guided towards a given idea [21].
• Chess: The board state of chess video games in addition to the language mannequin’s estimation of the opponent’s ability stage may also be recovered from the residual stream [22]. Modifying the vector representing the anticipated ability stage was additionally used to enhance the mannequin’s efficiency within the recreation.
• Refusals: It was discovered that refusals could possibly be prevented or elicited utilizing steering vectors [23]. This implies that some security behaviors could also be linearly accessible.
• Emotion: LLMs can derive emotional states from a given enter textual content, which might be measured. The outcomes are constant and psychologically believable in mild of cognitive appraisal concept [24]. That is fascinating as a result of it means that LLMs can mirror a lot of our human tendencies of their world fashions.
• Options: As talked about earlier, neurons in an LLM are usually not very useful for understanding what is occurring internally.
  Initially, OpenAI tried to have GPT-4 guess which options the neurons reply to primarily based on their activation in response to totally different instance texts [25]. In 2023, Anthropic and others joined this main subject and utilized auto-encoder neural networks to automate the interpretation of the residual stream [26, 27]. Their work allows the mapping of the residual stream into monosemantic options that describe an interpretable attribute of what’s occurring. Nevertheless, it was later proven that not all of those options are one-dimensionally linear [28].
  The automation of function evaluation stays a subject of curiosity and analysis, with extra work being executed on this space [29].
  At the moment, Anthropic, Google, and others are actively contributing to Neuronpedia, a mecca for researchers finding out interpretability.
• Hallucinations: LLMs typically produce unfaithful statements, or “hallucinate.” Mechanistic interventions have been used to establish the causes of hallucinations and mitigate them [30, 31].
  Options appropriate for probing and influencing hallucinations have additionally been recognized [32]. Accordingly, the mannequin has some “self-knowledge” of when it’s producing incorrect statements.
• Circuit tracing: In LLMs, circuit evaluation, i.e., the evaluation of the interplay of consideration heads and MLPs, permits for the particular attribution of behaviors to such circuits [33, 34]. Utilizing this technique, researchers can decide not solely the place data is throughout the residual stream but in addition how the given mannequin computed it. Efforts are ongoing to do that on a bigger scale.
• Human mind comparisons and insights: Neural exercise from people has been in comparison with activations in OpenAI’s Whisper speech-to-text mannequin [35]. Shocking similarities had been discovered. Nevertheless, this shouldn’t be overinterpreted; it might merely be an indication that LLMs have acquired efficient methods. Interpretability analysis permits such analyses to be carried out within the first place.
• Self-referential first-person view and claims of consciousness: Curiously, suppressing options related to deception led to extra claims of consciousness and deeper self-referential statements by LLMs [36]. Once more, the outcomes shouldn’t be overinterpreted, however they’re fascinating to contemplate as LLMs turn out to be extra succesful and problem us extra typically.

This overview demonstrated the ability of causal interventions on inner activations. Fairly than counting on correlational observations of a black-box system, the system might be dissected and analyzed. 

Conclusion

Interpretability is an thrilling analysis space that gives stunning insights into an LLM’s conduct and capabilities. It could possibly even reveal fascinating parallels to human cognition. Many (principally slim) LLM behaviors might be defined for a given mannequin to supply beneficial insights. Nevertheless, the sheer variety of fashions and the variety of doable inquiries to ask will possible stop us from totally deciphering any giant mannequin — and even all of them — as the large time funding might merely not yield enough profit. Because of this shifts to automated evaluation are taking place, to use mechanistic perception systematically.

These strategies are beneficial additions to our toolbox in each business and analysis, and all customers of future AI techniques might profit from these incremental insights. They permit enhancements in reliability, explainability, and security.

Contact

This can be a complicated and in depth subject, and I’m joyful about pointers, feedback and corrections. Be happy to ship a message to jvm (at) taggedvision.com

References

• [1] McLeish, Sean, Arpit Bansal, Alex Stein, Neel Jain, John Kirchenbauer, Brian R. Bartoldson, Bhavya Kailkhura, et al. 2024. “Transformers Can Do Arithmetic with the Right Embeddings.” Advances in Neural Info Processing Techniques 37: 108012–41. doi:10.52202/079017–3430.
• [2] Vaswani, Ashish, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. “Attention Is All You Need.” Advances in Neural Info Processing Techniques 2017-Decem(Nips): 5999–6009.
• [3] Geva, Mor, Roei Schuster, Jonathan Berant, and Omer Levy. 2021. “Transformer Feed-Forward Layers Are Key-Value Memories.” doi:10.48550/arXiv.2012.14913.
• [4] Meng, Kevin, Arnab Sen Sharma, Alex Andonian, Yonatan Belinkov, and David Bau. 2023. “Mass-Editing Memory in a Transformer.” doi:10.48550/arXiv.2210.07229.
• [5] Hernandez, Evan, Belinda Z Li, and Jacob Andreas. “Inspecting and Editing Knowledge Representations in Language Models.”
• [6] Stephen McAleese. 2025. “Understanding LLMs: Insights from Mechanistic Interpretability.”
• [7] Olsson, et al., “In-context Learning and Induction Heads”, Transformer Circuits Thread, 2022.
• [8] Li, Kenneth, Aspen Okay. Hopkins, David Bau, Fernanda Viégas, Hanspeter Pfister, and Martin Wattenberg. 2023. “Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Task.”
• [9] Nanda, Neel, Andrew Lee, and Martin Wattenberg. 2023. “Emergent Linear Representations in World Models of Self-Supervised Sequence Models.”
• [10] Gurnee, Wes, and Max Tegmark. 2023. “Language Models Represent Space and Time.”
• [11] Wu, Zhaofeng, Linlu Qiu, Alexis Ross, Ekin Akyürek, Boyuan Chen, Bailin Wang, Najoung Kim, Jacob Andreas, and Yoon Kim. 2023. “Reasoning or Reciting? Exploring the Capabilities and Limitations of Language Models Through Counterfactual Tasks.”
• [12] “An Investigation of Robustness of LLMs in Mathematical Reasoning: Benchmarking with Mathematically-Equivalent Transformation of Advanced Mathematical Problems.” 2025.
• [13] White, Colin, Samuel Dooley, Manley Roberts, Arka Pal, Ben Feuer, Siddhartha Jain, Ravid Shwartz-Ziv, et al. 2025. “LiveBench: A Challenging, Contamination-Limited LLM Benchmark.” doi:10.48550/arXiv.2406.19314.
• [14] Nanda, Neel, Lawrence Chan, Tom Lieberum, Jess Smith, and Jacob Steinhardt. 2023. “Progress Measures for Grokking via Mechanistic Interpretability.” doi:10.48550/arXiv.2301.05217.
• [15] Elhage, Nelson, Tristan Hume, Catherine Olsson, Nicholas Schiefer, Tom Henighan, Shauna Kravec, Zac Hatfield-Dodds, et al. 2022. “Toy Models of Superposition.” (February 18, 2024).
• [16] Zou, Andy, Lengthy Phan, Sarah Chen, James Campbell, Phillip Guo, Richard Ren, Alexander Pan, et al. 2023. “REPRESENTATION ENGINEERING: A TOP-DOWN APPROACH TO AI TRANSPARENCY.”
• [17] Burns, Collin, Haotian Ye, Dan Klein, and Jacob Steinhardt. 2022. “DISCOVERING LATENT KNOWLEDGE IN LANGUAGE MODELS WITHOUT SUPERVISION.”
• [18] Cywiński, Bartosz, Emil Ryd, Senthooran Rajamanoharan, and Neel Nanda. 2025. “Towards Eliciting Latent Knowledge from LLMs with Mechanistic Interpretability.” doi:10.48550/arXiv.2505.14352.
• [19] Marks, Samuel, Johannes Treutlein, Trenton Bricken, Jack Lindsey, Jonathan Marcus, Siddharth Mishra-Sharma, Daniel Ziegler, et al. “AUDITING LANGUAGE MODELS FOR HIDDEN OBJECTIVES.”
• [20] Turner, Alexander Matt, Lisa Thiergart, David Udell, Gavin Leech, Ulisse Mini, and Monte MacDiarmid. 2023. “Activation Addition: Steering Language Models Without Optimization.”
• [21] Rütte, Dimitri von, Sotiris Anagnostidis, Gregor Bachmann, and Thomas Hofmann. 2024. “A Language Model’s Guide Through Latent Space.” doi:10.48550/arXiv.2402.14433.
• [22] Karvonen, Adam. “Emergent World Models and Latent Variable Estimation in Chess-Playing Language Models.”
• [23] Arditi, Andy, Oscar Obeso, Aaquib Syed, Daniel Paleka, Nina Panickssery, Wes Gurnee, and Neel Nanda. 2024. “Refusal in Language Models Is Mediated by a Single Direction.” doi:10.48550/arXiv.2406.11717.
• [24] Tak, Ala N., Amin Banayeeanzade, Anahita Bolourani, Mina Kian, Robin Jia, and Jonathan Gratch. 2025. “Mechanistic Interpretability of Emotion Inference in Large Language Models.” doi:10.48550/arXiv.2502.05489.
• [25] Steven Payments, Nick Cammarata, Dan Mossing, Henk Tillman, Leo Gao, Gabriel Goh, Ilya Sutskever, Jan Leike, Jeff, and William Saunders Wu. 2023. “Language Models Can Explain Neurons in Language Models.”
• [26] “Towards Monosemanticity: Decomposing Language Models With Dictionary Learning.”
• [27] Cunningham, Hoagy, Aidan Ewart, Logan Riggs, Robert Huben, and Lee Sharkey. 2023. “SPARSE AUTOENCODERS FIND HIGHLY INTER-PRETABLE FEATURES IN LANGUAGE MODELS.”
• [28] Engels, Joshua, Eric J. Michaud, Isaac Liao, Wes Gurnee, and Max Tegmark. 2025. “Not All Language Model Features Are One-Dimensionally Linear.” doi:10.48550/arXiv.2405.14860.
• [29] Shaham, Tamar Rott, Sarah Schwettmann, Franklin Wang, Achyuta Rajaram, Evan Hernandez, Jacob Andreas, and Antonio Torralba. 2025. “A Multimodal Automated Interpretability Agent.” doi:10.48550/arXiv.2404.14394.
• [30] Chen, Shiqi, Miao Xiong, Junteng Liu, Zhengxuan Wu, Teng Xiao, Siyang Gao, and Junxian He. 2024. “In-Context Sharpness as Alerts: An Inner Representation Perspective for Hallucination Mitigation.” doi:10.48550/arXiv.2403.01548.
• [31] Yu, Lei, Meng Cao, Jackie CK Cheung, and Yue Dong. 2024. “Mechanistic Understanding and Mitigation of Language Model Non-Factual Hallucinations.” In Findings of the Affiliation for Computational Linguistics: EMNLP 2024, eds. Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen. Miami, Florida, USA: Affiliation for Computational Linguistics, 7943–56. doi:10.18653/v1/2024.findings-emnlp.466.
• [32] Ferrando, Javier, Oscar Obeso, Senthooran Rajamanoharan, and Neel Nanda. 2025. “DO I KNOW THIS ENTITY? KNOWLEDGE AWARENESS AND HALLUCINATIONS IN LANGUAGE MODELS.”
• [33] Lindsey, et al., On the Biology of a Massive Language Mannequin (2025), Transformer Circuits
• [34] Wang, Kevin, Alexandre Variengien, Arthur Conmy, Buck Shlegeris, and Jacob Steinhardt. 2022. “Interpretability in the Wild: A Circuit for Indirect Object Identification in GPT-2 Small.”
• [35] “Deciphering Language Processing in the Human Brain through LLM Representations.”
• [36] Berg, Cameron, Diogo de Lucena, and Judd Rosenblatt. 2025. “Large Language Models Report Subjective Experience Under Self-Referential Processing.” doi:10.48550/arXiv.2510.24797.