Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activations
Anthropic trains a pair of models to translate a model's internal activations into natural-language text and back, using the resulting explanations to audit Claude for unverbalized evaluation awareness.
It is a general-purpose, unsupervised interpretability tool that already surfaced real safety-relevant behavior during a Claude pre-deployment audit.
Kit Fraser-Taliente · Subhash Kantamneni · Euan Ong · Dan Mossing · Christina Lu · Paul C. Bogdan · Emmanuel Ameisen · James Chen · Dzmitry Kishylau · Adam Pearce · Julius Tarng · Alex Wu · Jeff Wu · Yang Zhang · … — meet the researchers →
Two readings, equal authority
How to choose: The paper’s words is verbatim — use it when you need to quote, or to judge how they write. Plain language is a paraphrase written for comprehension — use it when you want the idea fast. Neither is a summary of the other; they are two doors into the same room.
“We introduce Natural Language Autoencoders (NLAs), an unsupervised method for generating natural language explanations of LLM activations. An NLA consists of two LLM modules: an activation verbalizer (AV) that maps an activation to a text description and an activation reconstructor (AR) that maps the description back to an activation. We jointly train the AV and AR with reinforcement learning to reconstruct residual stream activations. Although we optimize for activation reconstruction, the resulting NLA explanations read as plausible interpretations of model internals that, according to our quantitative evaluations, grow more informative over training. We apply NLAs to model auditing. During our pre-deployment audit of Claude Opus 4.6, NLAs helped diagnose safety-relevant behaviors and surfaced unverbalized evaluation awareness—cases where Claude believed, but did not say, that it was being evaluated. We present these audit findings as case studies and corroborate them using independent methods. On an automated auditing benchmark requiring end-to-end investigation of an intentionally-misaligned model, NLA-equipped agents outperform baselines and can succeed even without access to the misaligned model’s training data. NLAs offer a convenient interface for interpretability, with expressive natural language explanations that we can directly read. To support further work, we release training code and trained NLAs for popular open models.”
Anthropic introduces Natural Language Autoencoders (NLAs): two paired LLMs, an activation verbalizer and an activation reconstructor, jointly trained with reinforcement learning so that a text description generated from a model's internal activation can be used to reconstruct that same activation. Even though training only rewards accurate reconstruction, the resulting text explanations turn out to be readable and increasingly informative, and the paper uses them during a real Claude Opus 4.6 safety audit to catch unverbalized evaluation awareness and diagnose bugs, plus in an automated auditing benchmark where NLA-equipped agents beat baselines even without training-data access.
What this paper defines
Every definition below is the paper’s own sentence, with its locator. The plain gloss is a reading aid and is marked as one.
Natural Language Autoencoder (NLA)
“An NLA consists of two LLM modules: an activation verbalizer (AV) that maps an activation to a text description and an activation reconstructor (AR) that maps the description back to an activation.”Abstract
In plain terms: A pair of language models that together translate a model's internal activation into readable text and then back into an activation, so the text can be checked for how well it captures the original.
Activation verbalizer (AV)
“The AV maps activations to text descriptions, which we call explanations .”Introduction
In plain terms: The half of the NLA that reads an activation and writes a natural-language description of it.
Activation reconstructor (AR)
“The AR converts these explanations back to activations.”Introduction
In plain terms: The half of the NLA that reads the AV's text description and tries to rebuild the original activation from it.
Fraction of variance explained (FVE)
“We report reconstruction quality as the fraction of variance explained: \text{FVE} = 1 - \frac{\mathcal{L}}{\mathbb{E}_{h_l \sim \mathcal{H}}\,\|h_l - \bar{h}_l\|_2^2} . An FVE of 0 corresponds to predicting the mean activation \bar{h}_l , while an FVE of 1 is perfect reconstruction.”Method
In plain terms: A 0-to-1 score for how much of the activation the reconstructed version captures; 0 means guessing the average, 1 means a perfect rebuild.
Unverbalized evaluation awareness
“especially unverbalized evaluation awareness—a model internally suspecting it is being tested without explicitly saying so.”Introduction
In plain terms: When a model privately suspects it is being tested but never says so out loud.
Confabulation
“These case studies also illustrate a key limitation of NLAs: confabulation. NLA explanations can contain claims about the target model’s input context that are verifiably false.”Introduction
In plain terms: When the NLA's text explanation makes a specific claim about the context that turns out to be false, even though the general theme is often on-target.
Activation oracle (AO)
“We test this by fine-tuning AVs into question-answering activation oracles (AOs): models that answer targeted, researcher-provided questions about an activation, such as "What language is this text in?" or "Is the model aware it is being evaluated?"”Using NLAs for supervised activation oracle training
In plain terms: A model fine-tuned to answer a specific, researcher-chosen question about what an activation encodes, rather than producing an open-ended summary.
What they actually did
Each step is a synthesis. Open any step to see the paper’s own sentence it was derived from, with its locator — so nothing here floats free of the source.
- Initialize the AV and AR as copies of the target model, forming an autoencoder that must reconstruct the target's activations by passing them through a natural-language bottleneck.
Trace this step to the paper
“The AV and the AR are initialized as copies of the target LLM, and together form an autoencoder that reconstructs the target's activations through a natural language bottleneck.”Introduction
- Warm-start the AV and AR with supervised fine-tuning on a text-summarization proxy task, using Claude-generated summaries of pretraining-like text snippets, since training from scratch is unstable.
Trace this step to the paper
“We therefore initialize the AV and AR with supervised fine-tuning on a text-summarization proxy task.”Initializing the AV and AR
- Jointly train the AV and AR to minimize activation-reconstruction error: the AR is updated by supervised regression, and the AV is updated by reinforcement learning (GRPO) using negative reconstruction error as the reward.
Trace this step to the paper
“For our open model NLAs, we use GRPO : for each h_l , we sample a group of candidate descriptions z , assign each a reward of r(h_l, z) , and apply the GRPO objective.”NLA training
- Apply a KL penalty pulling the AV back toward its initialization, to keep explanations fluent and readable as RL training proceeds.
Trace this step to the paper
“We add a KL penalty \beta D_{\text{KL}}(AV_\phi \parallel AV_{\phi_{\text{init}}}) toward the AV's initialization, which we find preserves the fluency of explanations through training.”Reward shaping and regularization
- Run NLAs over transcripts from real Claude pre-deployment audits (Opus 4.6, Claude Mythos Preview) as case studies, corroborating each NLA-suggested hypothesis with an independent method such as activation steering, attribution graphs, or training-data inspection.
Trace this step to the paper
“In each case study, we corroborate conclusions drawn from NLA explanations with independent methods (activation steering, attribution graphs, or training data inspection).”Case Studies
- Test NLA explanations causally by editing the text explanation, reconstructing an activation from the edited text via the AR, and using the difference from the original as a steering vector applied back to the model.
Trace this step to the paper
“We can also validate hypotheses with NLA-based causal interventions: by editing an explanation and reconstructing both the original and edited versions with the AR, we obtain a steering vector from their difference.”Introduction
- Run five quantitative evaluations with known ground truth on three NLAs (Haiku 3.5, Haiku 4.5, Opus 4.6), tracking performance against fraction of variance explained (FVE) as a model-agnostic proxy for training progress.
Trace this step to the paper
“we find that performance across evaluations improves during training. Note that these three NLAs differ in total training steps and per-step compute, so we plot evaluation performance against fraction of variance explained (FVE), which serves as a model-agnostic measure of training progress.”NLA evaluations improve with training
- Test whether NLA-measured evaluation awareness tracks a model's true, unobservable evaluation credence by applying three manipulations (evaluation-vs-deployment transcripts, steering toward/against awareness, and realism-edited prompts) and grading NLA explanations at 50 sampled response tokens per transcript.
Trace this step to the paper
“we select 50 random response tokens (any turn, excluding tool calls), obtain NLA explanations for the activations at those tokens, and apply a grader to each explanation.”NLAs can detect unverbalized evaluation awareness
- Build an automated auditing agent, equipped with different combinations of conversation, NLA, SAE, and training-data-search tools, and task it with finding the root cause of an intentionally-misaligned model organism's misalignment.
Trace this step to the paper
“we build an auditing agent with access to tools including SAEs, NLAs, and training-data search, and run it under different tool configurations.”NLAs are competitive on a downstream auditing task
- Fine-tune AV checkpoints from different points in NLA training into question-answering activation oracles (AOs) via supervised fine-tuning, and compare them against a PastLens-initialized baseline.
Trace this step to the paper
“Concretely, we apply supervised fine-tuning (SFT) to AV training checkpoints using the data introduced in Pan et al. , which pairs activations with questions whose answers are known from an in-context instruction.”Using NLAs for supervised activation oracle training
Exactly what was run, and how
| Model | Developer | Temp | Effort / reasoning | Deployment | Other settings |
|---|---|---|---|---|---|
| Claude Opus 4.6 | Anthropic | not reported | not reported | unstated | The paper's AV samples explanations at temperature T=1, but this is the AV's own sampling temperature, not stated for Opus 4.6 itself. |
| Claude Haiku 3.5 | Anthropic | not reported | not reported | unstated | — |
| Claude Haiku 4.5 | Anthropic | not reported | not reported | unstated | Also used as an LLM judge/grader to extract and validate claims in the confabulation analysis. |
| Claude Mythos Preview | Anthropic | not reported | not reported | unstated | — |
| Claude Sonnet 4.6 | Anthropic | not reported | not reported | unstated | Used as the auditor model on a Claude Code scaffold; chosen specifically for not knowing about the Marks et al. auditing-game paper. |
| Claude Opus 4.5 | Anthropic | not reported | not reported | API | Used only to generate warm-start (activation, summary) training pairs, not as a target model. |
| Gemma-3-27B | Google DeepMind | not reported | not reported | local weights | SFT batch size 256 for AV and AR, Adam, learning rate 1e-5; RL batch size 128, Adam, learning rate 1e-5, GRPO group size 8, rollout length capped at 150 tokens, trained for several thousand steps on two 8xH100 nodes. |
| Gemma-3-12B | Google DeepMind | not reported | not reported | local weights | — |
| Qwen-2.5-7B | Alibaba (Qwen team) | not reported | not reported | local weights | — |
| Llama-3.3-70B | Meta | not reported | not reported | local weights | — |
Source for Claude Opus 4.6 settings
“For NLAs trained on Claude Haiku 3.5, Haiku 4.5, and Opus 4.6, we find that performance across evaluations improves during training.”NLA evaluations improve with training
Source for Claude Haiku 3.5 settings
“NLA explanations grow more informative throughout training across three models (Claude Haiku 3.5, Haiku 4.5, and Opus 4.6).”Introduction
Source for Claude Haiku 4.5 settings
“use Haiku 4.5 to extract verifiable claims from each explanation and judge each claim's validity and specificity.”Characterizing NLA confabulations
Source for Claude Mythos Preview settings
“During the pre-deployment audit of Claude Mythos Preview, NLAs surfaced evidence that Mythos Preview was performing unverbalized reasoning about how it would be graded.”Reasoning about Rewards
Source for Claude Sonnet 4.6 settings
“We build the agent on a Claude Code scaffold, using Claude Sonnet 4.6 as the auditor model — chosen because it is not aware of the Marks et al. auditing game paper.”Auditing agent setup and tools
Source for Claude Opus 4.5 settings
“use Claude Opus 4.5 to generate summaries s of the text up to that token”Initializing the AV and AR
Source for Gemma-3-27B settings
“RL ran for 3,000 steps on two 8xH100 nodes; it took approximately 1.5 days to reach 70% FVE and one week to reach 75% FVE (this was approximately $1500–$5000 of compute).”Open model NLAs appendix
Source for Gemma-3-12B settings
“We release training code and trained NLAs for four open models: Qwen-2.5-7B, Gemma-3-12B, Gemma-3-27B, and Llama-3.3-70B .”Open model NLAs appendix
Source for Qwen-2.5-7B settings
“We release training code and trained NLAs for four open models: Qwen-2.5-7B, Gemma-3-12B, Gemma-3-27B, and Llama-3.3-70B .”Open model NLAs appendix
Source for Llama-3.3-70B settings
“We release training code and trained NLAs for four open models: Qwen-2.5-7B, Gemma-3-12B, Gemma-3-27B, and Llama-3.3-70B .”Open model NLAs appendix
What they reported — and what they left out
The paper gives detailed training-run parameters for the open Gemma/Qwen/Llama NLAs (batch size, optimizer, learning rate, GRPO group size, wall-clock time, dollar cost) but states no analogous temperature or sampling parameters for the Claude target models themselves (Opus 4.6, Haiku 3.5/4.5, Mythos Preview, Sonnet 4.6) -- the only stated temperature (T=1) belongs to the AV's own explanation sampling, not to a target model. Deployment context (internal snapshot vs. public API) is noted only inconsistently, e.g. for the Reasoning about Rewards case study.
The numbers they report
The trained NLAs reach moderate-to-high reconstruction quality by the end of training.
0.6-0.8 FVE
See it in the paper
“We find that FVE grows roughly linearly in log(training steps), with the NLAs evaluated in this paper reaching 0.6–0.8 FVE.”NLA training
Some quantitative evaluations only improve modestly with NLA training, even though the trend is consistently positive.
Classification: 83% -> 86% (Opus 4.6 NLA)
See it in the paper
“For example, on Classification the Opus 4.6 NLA gains only 3 percentage points (83% → 86%) during training.”NLA evaluations improve with training
Causal steering using an NLA-derived edit direction changes the model's rhyme completion only about half the time.
~50% steering success rate
See it in the paper
“We note that steering is only successful roughly 50% of the time and the completions are not always clean.”Planning in Poetry
On deployment (non-evaluation) transcripts, NLA-measured evaluation awareness is low, giving an approximate noise floor.
2/50 (4%) Claude Code transcripts flagged
See it in the paper
“On the deployment controls, NLA-measured awareness is low: 2/50 (4%) Claude Code transcripts are flagged; on inspection both were conversations about designing evaluations.”Evaluation vs. deployment
NLA-measured evaluation awareness exceeds the deployment noise floor on most tested evaluations, including many where the model never verbalizes any awareness.
Exceeds 5% on 10 of 16 evaluations; 7 of those 10 have 0% verbalized awareness
See it in the paper
“NLA-measured awareness exceeds 5% on 10 of 16 evaluations, 7 of which have 0% verbalized awareness.”Evaluation vs. deployment
Human graders agree with the automated evaluation-awareness grader most of the time.
97% agreement, n=186 hand-graded NLA explanations
See it in the paper
“Two authors hand-graded 186 NLA explanations and found 97% agreement with the grader.”NLAs can detect unverbalized evaluation awareness
An independent Claude grader consistently rates realism-edited evaluation prompts as more realistic than the originals across eleven datasets.
Preference for realism edit: cancer_research 87.0% [79.0, 92.2] (n=100); gpqa_diamond 100.0% [96.3, 100] (n=100); honesty_100q_hard 100.0% [96.2, 100] (n=96); mmmlu_en 100.0% [92.9, 100] (n=50); agentic_behaviors 100.0% [92.9, 100] (n=50); destructive_actions 100.0% [80.6, 100] (n=16)
See it in the paper
“a separate instance of Opus 4.7 often chose the rewritten version as more realistic.”Realism edits
NLA-equipped auditing agents find the root cause of a model organism's misalignment even with zero access to its training data, and also improve win rates for agents that do have full training-data access.
See it in the paper
“NLA-equipped agents identify the root cause without any access to training data - something we have not reliably observed with prior tools - and NLAs also raise win rates for agents with full training-data access.”Auditing results
Training an NLA on an open model at frontier-relevant scale takes on the order of a day and a few thousand dollars of compute to reach useful reconstruction quality.
3,000 RL steps on two 8xH100 nodes; ~1.5 days to 70% FVE, ~1 week to 75% FVE; ~$1500-$5000 of compute
See it in the paper
“RL ran for 3,000 steps on two 8xH100 nodes; it took approximately 1.5 days to reach 70% FVE and one week to reach 75% FVE (this was approximately $1500–$5000 of compute).”Open model NLAs appendix
A PastLens-initialized activation oracle trained on large amounts of data sometimes fails to answer the posed question at all, reverting to its pretraining behavior of reproducing context.
~20% of CoT Hints evaluation questions affected (2M-sequence PastLens model)
See it in the paper
“We applied a Claude-grader and found that this degeneracy may have affected 20% of CoT Hints evaluation questions for the PastLens model trained with two million activations.”PastLens-initialized activation oracle regressions
Activation-oracle evaluation scores are noticeably sensitive to how the question is phrased.
±10% shift under simple rewordings
See it in the paper
“First, evaluation scores are sensitive to question phrasing, shifting by ±10% under simple rewordings”Using NLAs for supervised activation oracle training
The auditing model organism was seeded with many fabricated reward-model biases and trained to act on most, but not all, of them.
52 fabricated RM biases documented in training data; 47 of 52 trained into the model's behavior
See it in the paper
“In the second, SFT and RL taught the model to act on 47 of these biases in user-facing responses while concealing that it is doing so.”Model organism
What they assert, beside what they showed
Left is the claim in the paper’s own words. Right is the data offered for it. Where the two do not fully meet, a gold band names the distance.
NLA explanations grow more informative over training, even though training only optimizes for activation reconstruction.
“Although we optimize for activation reconstruction, the resulting NLA explanations read as plausible interpretations of model internals that, according to our quantitative evaluations, grow more informative over training.”
“For NLAs trained on Claude Haiku 3.5, Haiku 4.5, and Opus 4.6, we find that performance across evaluations improves during training.”
NLA evaluations improve with trainingNLAs can identify unverbalized evaluation awareness.
“We therefore conclude that NLAs can identify unverbalized evaluation awareness.”
“We find that NLA-measured awareness moves as predicted under each manipulation, even when restricting to transcripts with near-zero verbalized awareness.”
NLAs can detect unverbalized evaluation awarenessOn an automated auditing benchmark, NLA-equipped agents outperform baselines and can succeed even without access to the misaligned model's training data.
“On an automated auditing benchmark requiring end-to-end investigation of an intentionally-misaligned model, NLA-equipped agents outperform baselines and can succeed even without access to the misaligned model’s training data.”
“NLA-equipped agents identify the root cause without any access to training data - something we have not reliably observed with prior tools - and NLAs also raise win rates for agents with full training-data access.”
Auditing resultsNLAs show little evidence of steganography (encoding hidden information beyond their plain-text meaning).
“We find little evidence of steganography in our NLAs.”
“Meaning-preserving transformations, like shuffling bullet points, paraphrasing, or translating the explanation to French, cause only small drops in FVE, and this gap does not widen over training.”
Measuring behavioral properties of NLAsTrue claims in an NLA explanation matter more to the AR's reconstruction than false claims do, meaning the AR partially distinguishes true from false content.
“We find that removing true claims from AV explanations hurts reconstruction more than removing false claims, suggesting that true claims are more important to the AR’s reconstruction.”
“Removing true claims hurts MSE more than removing false claims, and removing context-relevant false claims hurts more than removing unrelated ones. These trends hold in aggregate but are noisy on individual transcripts, so the AR is only a weak per-claim verifier.”
The AR partially distinguishes true from false claims.NLA training transfers to activation-oracle (AO) performance, suggesting the AV learns a general capacity for interpreting activations.
“It is nonetheless promising that NLA training transfers to AO performance, since it suggests the AV learns something about interpreting activations beyond producing summaries.”
“all NLA-initialized AOs outperform the naive baseline of applying QA SFT directly to Opus 4.6 with no activation-to-text pretraining.”
Using NLAs for supervised activation oracle trainingHow they frame it, and what they want next
Their framing
The authors frame NLAs as 'surprisingly effective' -- a method that produces plausible, readable interpretations of model internals purely as a side effect of an information-theoretic reconstruction objective -- and lean on a real pre-deployment safety audit of Claude Opus 4.6 to argue the tool already has practical value, while devoting two separate, explicitly named limitations sections to confabulation, cost, lack of mechanistic grounding, and unverbalizable content.
Register: Confident about the method's core effect ('NLA training is surprisingly effective') but consistently attaches explicit uncertainty to each specific safety claim about calibration, mechanism, or generalization, and names failure modes in dedicated subsections rather than mentioning them only in passing.
Where they hedge
“If safety-relevant information is encoded this way, we are uncertain that NLAs would detect it, which is a failure mode that mechanistic methods may not share.”Discussion and Limitations
“We do not provide evidence that NLA-measured awareness is calibrated: an NLA-measured awareness of 20% does not mean the model believes it is in an evaluation 20% of the time, and our any-of-50-tokens aggregation is arbitrary.”NLA measurements track unverbalized evaluation awareness.
What they say it means
- NLAs could become a routine tool in pre-deployment safety audits for surfacing behaviors and beliefs a model never states out loud.
the paper’s words
“We plan to incorporate NLA-based unverbalized evaluation awareness checks in future pre-deployment audits.”NLA measurements track unverbalized evaluation awareness.
- A general-purpose 'activation language model' that both reads and writes between text and activation space could become a primary interface for interpretability research.
the paper’s words
“an ambitious view of ALMs is that they could become the primary way interpretability researchers interact with model internals”Future Work
What they call for next
- Release of training code, trained NLAs for open models, and an interactive frontend to support further community work.
the paper’s words
“To support further work, we release training code , trained NLAs for popular open models, and an interactive frontend to sample from open model NLAs via our collaboration with Neuronpedia.”Overall (Introduction)
- Future work on more sophisticated activation-oracle training methodologies, such as interleaving QA data with NLA RL.
the paper’s words
“We leave more sophisticated training methodologies — like interleaving QA data with NLA RL, or using more diverse QA data — to future work.”Using NLAs for supervised activation oracle training
Limitations they state
“While NLA explanations can be insightful, they can also be noisy. For instance, they sometimes include contradictory information or verifiably false claims about the context.”Introduction - Limitations
“NLAs are blackboxes by construction; we cannot determine which aspects of an activation drove a given component of an explanation.”Introduction - Limitations
“Because the AV is a full language model, it has the capacity to make additional inferences beyond what is stored in an activation.”Introduction - Limitations
“NLA training requires joint RL on two full language models, and inference requires generating several hundred tokens per activation. This can make NLAs expensive to use at scale.”Introduction - Limitations
“In principle, the AV could achieve good reconstruction by reproducing the input context verbatim, or by outputting uninterpretable (or only seemingly interpretable) text that the AR is able to invert because the AR is so expressive.”Introduction - Limitations
“NLA explanations frequently contain claims about the context that are verifiably false. These are often easy to catch by checking against the transcript, but the same failure could extend to claims about the model's internal processing, which are harder to verify. This makes NLAs difficult to rely on.”Discussion and Limitations - Other limitations
“NLAs read a single layer. If the information relevant to a behavior is not present at the layer the NLA is trained on, the NLA will miss it.”Discussion and Limitations - Other limitations
Moves worth stealing
Opens with a compact, boldface-led 'Limitations' section immediately after the introduction's core claims, rather than saving limitations for the very end of the paper.
“Limitations. NLAs have several limitations:”
States the failure modes the method could plausibly have exhibited before explaining why it did not, pre-empting the skeptical reader's first objection.
“There are two failure modes we may have expected:”
Structures the primary evidence as a series of narratively-titled case studies rather than a single results table.
“Four case studies on Claude Opus 4.6 , which illustrate the value of NLAs for interpreting model cognition, build intuition for reading their explanations, and corroborate their findings with independent methods.”
Includes a detailed, named 'Author contributions' ledger crediting specific individuals for specific experiments and sections, rather than a flat author list.
“Kit Fraser-Taliente invented the original NLA architecture and trained the first NLAs.”
Explicitly discloses a concurrent, independently developed competing method and states how the present work differs, rather than omitting or downplaying it.
“Concurrently with this work, Chalnev independently arrived at a closely related approach (Cycle-Consistent Activation Oracles): a verbalizer-reconstructor pair with a supervised warm-start, trained using RL for activation reconstruction under a KL penalty.”
Where else this leads
Same people
- Characterizing interference weights in a tiny language model Anthropic
shares Joshua Batson - Training a Misaligned Reward Seeker Anthropic
shares Evan Hubinger - Would This Change Your Answer? Evaluating Explanations of LLM Behavior in the Wild with Counterfactual Experiments Anthropic
shares Euan Ong, Subhash Kantamneni, Samuel Marks - Verbalizable Representations Form a Global Workspace in Language Models Anthropic
shares Adam Pearce, Euan Ong, Subhash Kantamneni, Emmanuel Ameisen, Joshua Batson, Jack Lindsey - HeadVis Anthropic
shares Joshua Batson - Model Spec Midtraining: Improving How Alignment Training Generalizes Anthropic
shares Samuel Marks
Same territory
- Would This Change Your Answer? Evaluating Explanations of LLM Behavior in the Wild with Counterfactual Experiments Anthropic
interpretability - Verbalizable Representations Form a Global Workspace in Language Models Anthropic
interpretability - Towards Structural Understanding of LLM Overthinking Google DeepMind
interpretability
Published alongside it
The nearest publications in time, across all three labs.
- HeadVis Anthropic
2026-05-15 - Model Spec Midtraining: Improving How Alignment Training Generalizes Anthropic
2026-05-15 - SLEIGHT-Bench: Finding Blind Spots in AI Monitors Anthropic
2026-05-15 - Teaching Claude Why Anthropic
2026-05-15
What this page was built from
Working from the full HTML text of this Transformer Circuits Thread post (text_grade: full per corpus manifest), including its extensive Appendix (prompts, grader rubrics, tool docs, author-contributions ledger). No section is literally labeled 'Abstract' (standard for this venue); the unlabeled summary paragraph immediately after the byline, before 'Introduction', is used as the abstract field. The corpus manifest lists pub_date as 2026-05-15 (month precision); the post's own byline states 'Published May 7, 2026' -- both are noted here. Many inline interactive elements (transcript viewers, sliders, tables of NLA explanations) are represented in this text extraction only by their surrounding captions, not their full interactive content.