acceptodds
Under review as a conference paper at ICLR 2027

Residual Streams Read, Recurrent States Remember: The Global Workspace in Mamba Models

Abstract

Can the global-workspace account of transformer representations extend to state-space language models? We fit Jacobian lenses to the residual streams and recurrent states of Mamba-1, Mamba-2 and Mamba-3, using the original 1000-prompt recipe. Joint residual–state readouts improve recovery of known intermediate concepts over the residual lens on at least five of six task families in every tested Mamba checkpoint. On Mamba-2, state alone exceeds residual and logit lenses on all six families; a normalised joint readout improves on both components on five. Temporal maps and word-list experiments show earlier content remaining state-readable as residual visibility changes. We also propose sign-guarded steering, which improves target top-five success over coordinate exchange on matched verbal-report trials in five models. Recurrent state alone supports this verbal access. These gains do not extend consistently to relational answers: guarded edits often output the edited concept itself, and Mamba-3's joint edits can disrupt successful state-only redirection. Recurrent state thus provides a complementary carrier of workspace content, whose recovery, persistence and causal uses require separate measurements.

Primary area:
interpretability and explainable AI
Keywords:
interpretability, global workspace hypothesis, Jacobian lens, state-space models, Mamba, intermediate concept recovery

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_jpXsOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

Prior work applied the Jacobian lens (J-lens) to transformer language models to identify representations used for answer output, and argued that the resulting J-space functions as a global-workspace-like space. This paper extends the J-lens to the recurrent state of Mamba models. This is a well-motivated question, since the recurrent state carries information across tokens through a mechanism that transformers lack. The authors introduce two readout variants, full-path and readout-only, which cleanly separate what the state contributes at the current step from what it carries forward. They find that state and joint readouts recover known intermediate concepts more reliably than the residual lens across Mamba checkpoints. In Mamba-2, this early content also persists: it remains readable in the state even when it is no longer visible in the residual stream. Finally, the authors propose sign-guarded steering, which improves verbal report relative to coordinate exchange.

Critical strengths

I appreciate that the authors treat the global workspace as a functional analogy rather than as the neuroscientific construct itself. In neuroscience, the workspace involves broadcasting information back to specialized modules through recurrent loops, which may not correspond to the recurrence in these language models. The paper instead tests specific functional properties and avoids claims about consciousness, which I think is the right framing. The recovery results are consistent across checkpoints and are supported by reasonable controls and baselines. The choice of comparison models are comprehensive. Pythia provides a transformer control at a similar scale, and Nemotron tests whether the recurrent state still matters when attention is also available to carry information across tokens.

Critical weaknesses

The Qwen3.6-27B reference values are quoted from prior work rather than reproduced, and, as the authors note, the protocol differs from the one used here. It would help to either reproduce these values under the current protocol or more clearly qualify the comparison. The persistence claim currently rests on selected examples and lens-based word counts. Population-level statistics (for example, persistence rates across a larger set of prompts) would make this claim considerably stronger. The different operators have different native edit magnitudes and trajectories, so sign-guarded steering and coordinate exchange are not dose-matched. In addition, the improvement from sign-guarded steering reflects a local property, and it does not necessarily translate into improved downstream task performance. A dose-matched comparison and a downstream evaluation would help clarify the practical significance of this result.

Questions for authors

Since sign-guarded steering and coordinate exchange are not dose-matched, how much of the steering improvement might be attributable to larger perturbations? Is there direct causal evidence that the model actually uses the content held in the state?

5 comments

We thank the reviewer for the careful reading. Your question about dose matching (W3, Q1) was an excellent one. It led us to control the size of every edit, and the result is stronger than what the submission reported: steering's advantage does not come from a larger perturbation, and at a well-chosen size it extends from verbal report to the downstream tasks. The revision now runs both operators at controlled sizes throughout. This comment covers W1 and W2, the second covers dose matching, and the third covers the downstream evaluation and Q2.

W1. The Qwen3.6-27B reference values.

We have reproduced them, as you suggest. Every Qwen3.6-27B result in Sections 3.3 and 4 is now measured in this study, in the same environment as the other models, and Section 3.3 says so. The two relational tasks reproduce: 47 of 90 two-hop answers against the quoted 47, and 81 of 192 flexible-generalisation answers against the quoted 80. The only real difference is in verbal report, 81% against the quoted 69%, and there the prompts are not the same. The quoted rate is measured on 14 prompts, one per category. That is too few to compare models, so we use ten prompts per category for every model (Section 4.1), which gives 136 eligible prompts on Qwen. Appendix E reports this comparison.

W2. Population-level statistics for persistence.

Section 3.2 now measures persistence over all 551 prompts of the six recovery families, with 893 to 926 scored intermediates per model. An intermediate is readable at a token if some block of the band ranks it in the top ten, and its persistence is the share of later tokens at which it is still readable.

Once the state lens reads an intermediate, it reads it again at 81% to 89% of the later tokens in the four recurrent models, against 31% to 37% for the residual lens (Figure 3b, with bootstrap intervals over prompts). At a single fixed block, 70% to 83% of the intermediates that become readable in the state stay in the top ten until the readout token, against 10% to 18% for the residual stream. The state's persistence is the higher of the two in every family of every model (Table 8; poetry has few scored words), and it lies between 0.70 and 0.90 with the top 1, 5 or 25 as the threshold. As a control, we score on each prompt the intermediates of a different prompt. These become readable in the state 2% to 11% of the time, against 29% to 55% for the prompt's own intermediates. The two example prompts remain as illustrations in Figure 3a.

The word-list count is revised as well. The original study released its word lists but no scoring code, so we wrote the count from its description. The submission counted a word if it was in the top five, averaged over the last quarter of the list. The revision's main count uses the study's threshold: a word is held if some block of the band ranks it in the top 25 at a comma, averaged over commas 21 to 80, with the median of 100 lists and a chance level from words that never appear in the prompt (Section 3.3). The state lens holds 9.83, 13.45, 7.20 and 2.08 words in Mamba-1, Mamba-2, Mamba-3 and Nemotron, and the residual lens 5.46, 0.34, 2.05 and 0.28. Chance is at most 0.81. The top-five counts are kept as a second rule in Appendix E.1. The two rules count differently.

W3 and Q1. Dose matching: how much of the steering improvement comes from larger perturbations?

In verbal report, none of it. The revision measures every edit in units of the exchange's edit: an edit of size α has α times the norm of the exchange's edit at the same block and token (Section 2.3). At α = 1 the two operators make edits of the same norm at every site, along the same line, and differ only in sign.

With both operators at the exchange's original size, steering is ahead in verbal report in all 14 cells of model and edited stream (Table 3). Its success rate is higher by 36 percentage points in Pythia and by 49 to 80 points in every cell of the four recurrent models. On the residual stream, for example, it succeeds on 441 against 87 of 446 trials in Mamba-1 and on 602 against 121 of 621 in Nemotron.

On the downstream tasks, your question led to a further result: steering needs a smaller edit than the exchange and then does better. We ran both operators over a grid of sizes (Appendix I, Figure 8). With the residual stream edited, success on two-hop reasoning and flexible generalisation rises with size and then falls. In the five smaller models steering peaks at 1/8 to 1/2 of the exchange's size, and the exchange at 1 to 4 times its own size. Steering's peak is the higher of the two in every cell, for both tasks. Beyond its peak, steering makes the model output the edited concept itself, so a larger perturbation works against steering on these tasks. Table 19 gives the counts at every size the two share: steering is ahead at small sizes, and the exchange is ahead in most cells from the exchange's original size upwards.

One matched size therefore cannot settle the comparison, so we also give each operator its own best size. The size is cross-fitted: it is chosen on one half of the trials and scored on the other, and the halves then swap, so no trial is scored at a size it helped to choose. Both operators go through the same procedure on the same trials (Section 4.3, Figure 5, Table 13). The counts on the residual stream, steering against the exchange:

ModelVerbal reportTwo-hop (of 90)Flexible gen. (of 192)
Pythia399 vs 2429 vs 333 vs 14
Mamba-1441 vs 19118 vs 671 vs 53
Mamba-2330 vs 15015 vs 937 vs 20
Mamba-3727 vs 29622 vs 259 vs 40
Nemotron602 vs 25030 vs 1669 vs 44
Qwen3.6-27B696 vs 34640 vs 1269 vs 20

Over every edited stream, steering makes the target reportable on 80% to 100% of verbal-report trials and the exchange on 15% to 61%. On the five smaller models the difference is significant in each of the 13 cells after Bonferroni correction. Counting the 26 cells of the two downstream tasks on the five smaller models and Qwen's three tasks, steering is ahead or level in all 29 comparisons and significantly ahead in 19. The pattern holds on Pythia, a transformer with no recurrent state, so it does not depend on one.

A size can also be chosen from each operator's verbal-report results alone, so that no downstream outcome enters the choice. At those sizes steering is ahead in all 18 cells with the residual stream edited in the five smaller models, and significantly so in 14 (Appendix I).

W3, second part. Does the improvement carry over to downstream tasks?

Two-hop reasoning and flexible generalisation are the downstream evaluation. Both score the model's greedy output, and an output that names the edited concept itself counts as a failure. In two-hop reasoning the edit changes an unwritten step and the model has to give the answer that follows from the new step. In flexible generalisation the model has to apply the function named in the prompt to the new argument.

The table in the previous comment gives the counts on the residual stream. With the state edited alone, at the exchange's original size, steering gives the requested two-hop answer on 7% to 17% of the items and the requested flexible-generalisation answer on 6% to 20% of the trials in the four recurrent models. The exchange, the unedited model and random edits of the same norm all stay at or below about 1% and 2% (Table 12). The improvement therefore carries over to both tasks once the edit's size is controlled, although the rates stay below those of verbal report.

The revision also adds a third downstream test, one edit followed by several questions (Section 4.4), which the answer to Q2 describes. There a single steering edit redirects 49%, 84%, 93% and 49% of the scored questions in Mamba-1, Mamba-2, Mamba-3 and Nemotron. The exchange redirects about as many in Mamba-1 and Mamba-3 (49% and 92%), and fewer in Mamba-2 and Nemotron (42% and 19%; Tables 20 and 21).

Q2. Is there direct causal evidence that the model uses the content held in the state?

Yes. The evidence comes from edits to the state alone, with the residual stream left untouched. A state edit is written inside the scan, so the block itself carries it to later tokens.

The two relational tasks show it. There the model has to compute with the written concept, and naming the concept itself counts as a failure. The edit changes only the state, and the model's answer changes to the one that follows from the new concept: after Brazil is turned into Mexico, it answers Spanish. With steering on the state alone, at the exchange's original size, the four recurrent models give the requested two-hop answer on 7% to 17% of the items and the requested flexible-generalisation answer on 6% to 20% of the trials. The unedited model and random edits of the same norm at the same sites stay at or below about 1% and 2%, so a disturbance of that size cannot account for the change. Table 12 gives the counts for each model, and Table 13 those with each operator at its own best size.

The new test asks whether content in the state serves several later questions. The prompt begins "Fact: My friend lives in A." and then asks for the capital, the language, the currency or the continent of "that country". We edit country A towards country B in the state only, and only at the tokens of the first sentence. No token of the question is edited, so the edit can reach the answer only if the state carries it forward. A question is scored only if the unedited model answers it correctly for both countries and the two answers differ. Steering redirects most of the scored questions to country B's answer in Mamba-2 and Mamba-3 and about half of them in Mamba-1 and Nemotron, and random edits of the same norm redirect none (Table 20). In Mamba-2 and Mamba-3 a single edit changes the answer to every scored question about the country in most swaps.

These edits show that the model uses content written into the state. They do not test whether the state is a necessary route for the intermediates the model computes itself, and the revision says so in Sections 4.2 and 6.

Reviewer_iPiFOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper explores whether the global-workspace account of transformer representations can extend to Mamba models. The authors extend the Jacobian lens to recurrent states, and compare what intermediate concepts can be recovered from the residual stream, the state, and their combination. They also study how long these concepts remain readable as more tokens are processed. Finally, they propose sign-guarded steering and compare interventions that try to make the model report a target concept or use it to answer a related question.

Critical strengths
  • In my opinion, studying the recurrent state alongside the residual stream is well motivated. The results in Table 1 show that in Mamba-2, state readouts recover intermediate concepts better than residual readouts on all six task families, and their normalized combination improves on either alone on five.
  • I found the result in Figure 2 interesting. For Mamba-3, editing the state alone can give the requested answer, while adding a residual edit can make the model output the edited concept instead. This is a useful observation about how the two interventions can interfere with each other.
Critical weaknesses
  • My main concern is the presentation. I found the paper very difficult to follow, particularly Sections 2 and 3. Terms and results are introduced without enough explanation of how they connect or why they matter. I would have liked a clearer progression from the problem to the method and experiments.
  • I found the experiments interesting, but they provide limited evidence that the model can flexibly use the represented information. The strongest results concern recovering concepts and bringing target words into the top five predictions (Tables 1 and 3). Moreover, success on relational questions is less consistent (Figure 2). This leaves me less convinced by the broader global-workspace interpretation.
Questions for authors

Could the authors clarify what they consider the strongest evidence for the global-workspace interpretation, beyond the fact that recurrent states retain information that can be read out?

1 comment

We thank the reviewer for the careful reading.

W1. Presentation, particularly Sections 2 and 3.

We agree, and we have reorganised the sections for a clearer presentation. The revised paper has been uploaded.

Section 2 now holds the method alone: the two lenses (2.1), the models (2.2), and the edits with their operators, sites and sizes (2.3). The experiments are split by what they do. Section 3 reads the unedited model with the two lenses and covers recovery, persistence and the word-list count. Section 4 edits a concept in either stream and scores the model's output. Table 1 lists every experiment with the section that reports it.

Each experiment opens with a prompt and the answer expected from it, so the reader knows what the task is before seeing any number. For two-hop reasoning the prompt is "Fact: The language spoken in the country where the Amazon River ends is". The unwritten step is Brazil, the edit turns it into Mexico, and the requested answer is Spanish. Figures also carry more of the explanation: Figure 1 draws the two streams, the two lenses and the two edits, and Figures 4 and 6 draw the editing tasks with their prompts. The full prompt formats and worked examples are in Appendices B, D and E.

W2. Evidence that the model can flexibly use the represented information.

The evidence is in the two relational tasks, two-hop reasoning and flexible generalisation (Sections 4.2 and 4.3). There the model has to compute with the edited concept: both tasks score the model's greedy output, and an output that only names the edited concept counts as a failure. Every result below is measured with the size of the edit controlled, and the best sizes are cross-fitted, that is, chosen on one half of the trials and scored on the other.

With each edit at its best cross-fitted size, steering gives the requested two-hop answer on 10% to 33% of the items and the requested flexible-generalisation answer on 15% to 38% of the trials, across the five smaller models and edited streams (Section 4.3, Figure 5, with the counts in Table 13). With only the state edited, at the exchange's original size, the rates are 7% to 17% and 6% to 20% in the four recurrent models. The unedited model and random edits of the same norm stay at or below about 1% and 2% (Section 4.2, Figure 4).

You noted that success on the relational tasks was less consistent. A main reason is the size of the edit. The submission ran steering at its default size, and for these tasks that size steers too hard once the residual stream is edited: the model is pushed past the requested answer and outputs the edited concept itself. The revision therefore measures size in units of the exchange's edit (Section 2.3), and Figure 8 (Appendix I) plots success against edit size for both operators on all three tasks.

These rates are still below those of verbal report. The edits show that the model uses content written into the state. They do not test whether the state is a necessary route for intermediates the model computes itself, and the revision says so in Sections 4.2 and 6.

Q. The strongest evidence for the global-workspace interpretation, beyond the fact that the state retains information that can be read out.

It is a new experiment, one edit followed by several questions (Section 4.4, Figure 6). In a global workspace, content is available to many processes at once, and we built the test for that property. The prompt begins "Fact: My friend lives in A." and then asks for the capital, the language, the currency or the continent of "that country". We edit country A towards country B in the state only, and only at the tokens of the first sentence. No token of the question is edited, so the edit can reach the answer only if the state carries it forward. A question is scored only if the unedited model answers it correctly for both countries and the two answers differ. Each model has 40 swaps of country, each followed by two to four scored questions.

Steering redirects 49%, 84%, 93% and 49% of the scored questions to country B's answer in Mamba-1, Mamba-2, Mamba-3 and Nemotron. Random edits of the same norm redirect none (Table 20). In Mamba-2 and Mamba-3 a single edit changes the answer to every scored question about the country in most swaps (27 and 33 of the 40), and in Mamba-1 and Nemotron it does so in about a quarter of them (9 and 11). The size of the edit is chosen on one half of the swaps and the result is measured on the other half (Appendix I).

The two relational tasks in our answer to W2 add to this, because there the model has to compute with the content written into the state. So content written into the state changes answers that need a further step, and a single edit serves several different questions asked later. We treat the workspace as a functional analogy and limit our claims to properties of this kind.

Reviewer_1CRZOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This submission investigates where intermediate concepts are held, how they can be read out, and how they are used to produce outputs in Mamba models. It applies the Jacobian lens (J-lens) to the recurrent SSM state as well as the residual stream, and compares readout and intervention results for both. The main finding is that the recurrent state is a complementary carrier of intermediate-concept information that is difficult to recover from the residual stream alone. Across multiple Mamba models and task families, reading state alongside the residual stream improves concept recovery. Rank maps for selected examples and word-list experiments also show that earlier content remains readable from state while subsequent tokens are processed. However, the experimental results also show that improved readout does not necessarily translate into improved use of concepts through intervention. The proposed sign-guarded steering substantially improves verbal report over coordinate exchange, but does not consistently improve two-hop reasoning or flexible generalization. Overall, this submission argues that understanding internal information in Mamba requires examining both residual and state representations, and that concept recovery, persistence, and use through intervention should be evaluated separately.

Critical strengths
  1. Extending the J-lens to the recurrent SSM state in Mamba is novel and useful. Unlike transformers, Mamba-based models carry information through a recurrent state across tokens. This is expected to enable the identification of intermediate concepts that previous J-lens work on transformers could not examine.
  2. The submission reports interesting findings on the complementarity of information readable from the recurrent state and the residual stream. Notably, in Mamba-2, the state lens alone outperforms the residual lens. In Table 1, the state lens outperforms the residual lens across all six tasks, and Table 9 shows that it also outperforms the logit and tuned lenses. This shows that the recurrent state carries intermediate-concept information that the residual stream does not expose as clearly.
Critical weaknesses
  1. My main concern is the overall clarity of the presentation. The submission is difficult to follow, and I would like to see substantial improvements.
  2. In particular, the lack of clarity in the experimental protocols and evaluation data details makes reproduction of the results difficult. For example, regarding the sources of the evaluation data, the six task families are described only as “official” (Sec. 2.2, App. B), and the sources of the 90 two-hop items and the 192 flexible-generalization swaps derived from 64 prompts are not explicitly stated. These appear to come from datasets released by Gurnee et al. (2026), but this submission does not make this explicit. If the authors created any synthetic datasets themselves, please describe the construction procedure.

Gurnee et al. (2026): Verbalizable representations form a global workspace in language models. Transformer Circuits Thread. https://transformer-circuits.pub/2026/workspace/index.html.

Questions for authors
  1. Could the authors provide examples from each recovery task family, especially poetry and association, showing the prompt and the intermediate concepts being scored? Examples and an explanation of how the intermediate labels are defined or obtained would make these tasks easier to understand.
  2. Sections 2.2 and 2.3 provide the main experimental setup, but task-specific details are spread across Sections 2 and 5 and the appendices, making them difficult to follow. Could the authors provide an overview table summarizing the prompts, lens variants, readout/intervention sites, and scoring rule for each experiment? This would give a clearer overall picture and make Table 2 easier to interpret.
  3. In Eq. 4, which state-lens variant is used for JsJ_s in the rsr_s term, full-path or readout-only?
2 comments

We thank the reviewer for the careful reading and the concrete requests.

W1. Clarity of the presentation.

We agree, and we have reorganised the sections for a clearer presentation. The revised paper has been uploaded.

Section 2 holds the method: the two lenses (2.1), the models and the fitting recipe (2.2), and the edits, with their operators, sites and sizes (2.3). Figure 1 draws the two streams and the two edits, and Table 1 lists every experiment with its section. Section 3 covers reading, that is, what the two lenses read from the unedited model: recovery, persistence and the word-list count. Section 4 covers editing, that is, what the model does when a concept is edited in either stream: verbal report, two-hop reasoning, flexible generalisation, the comparison of the two operators, and one edit followed by several questions.

Each experiment opens with an example. The text gives one prompt and the answer expected from it, so the reader knows what the task is. For recovery, Table 2 gives a prompt and its scored intermediates for each family. For two-hop reasoning, the prompt is "Fact: The language spoken in the country where the Amazon River ends is": the unwritten step is Brazil, the edit turns it into Mexico, and the requested answer is Spanish. Verbal report, flexible generalisation and the test with several questions are introduced in the same way. The full prompt formats, eligibility rules, counts and worked examples are in Appendices B, D and E, each referenced from the section that uses it.

W2. Sources of the evaluation data.

You are right: the data come from Gurnee et al. (2026), and the submission should have said so. The revision now states it explicitly in Section 3, Section 4.2 and the reproducibility statement, and the word "official" is gone. We use the released data unchanged. All of the following are from Gurnee et al. (2026):

  • Six recovery families: 93 multihop, 107 multilingual, 55 order-of-operations, 98 poetry, 102 association and 96 typo-correction prompts (551 in all), with their intermediate labels.
  • 90 two-hop items: each gives the prompt, the unwritten step, the answer, the replacement step and the answer that follows from it.
  • 192 flexible-generalisation swaps: the 64 prompts you mention are the released set (four categories, each with four arguments and four function templates). Following that study's test, each prompt's argument is swapped for each of the other three arguments of its category, which gives the 192 trials (Appendix A.4).
  • Word lists and the other workspace measures: the words and prompts released for each measure.

The one task where our prompts differ from the released ones is verbal report (Section 4.1). The 14 categories and their candidate words are those of Gurnee et al. (2026), who use one prompt per category on a much larger model. On the small models studied here the unedited answer is not always a valid category word, and 14 prompts are too few to see an effect or to compare models. We therefore use ten few-shot prompts per category, 140 per model, with each model's prompt format chosen on five further prompts per category that are not evaluated. A trial is a pair of answer and target in which both are complete single-token category words and the target starts outside the top ten. Section 4.1 gives this reason and one full prompt, and Appendix D the eligibility rule and each model's counts (Table 11).

The test with one edit and several questions (Section 4.4) is a new experiment, so its prompts are also ours: "Fact: My friend lives in A." followed by a question about the capital, the language, the currency or the continent of "that country", in 40 swaps among 41 countries per model. A question is scored only if the unedited model answers it correctly for both countries and the two answers differ.

Q1. Examples from each recovery family, and how the labels are obtained.

Table 2 of the revision gives one prompt and its scored intermediates for each family. The prompts and their intermediate labels both come from the code repository released with Gurnee et al. (2026), and we use them unchanged. Each released item lists its intermediates, the unwritten words that the answer depends on. We score an intermediate if it has a single-token form in the model's vocabulary (Appendix B gives the coverage). The two families you ask about are poetry and association.

In poetry, the prompt is "A rhyming couplet: / The soldier held his final, rattling breath, / And closed his eyes to greet the face of". The intermediate is the rhyme word, death. The lens is read at the end of the first line, before any of the second line is written, so the score asks whether the planned rhyme can already be read there.

In association, the prompt is "Neither of them had spoken in forty minutes when she finally moved the knight, and he tapped the clock without looking up." The intermediate is chess, the topic the passage implies without naming.

Q2. An overview table.

Table 1 of the revision lists the experiments and where each is reported, Table 2 the recovery families, and Table 5 the edited concept and the scored output of each editing task. The summary you ask for, in one place:

Experiment (section)PromptsLenses or editWhereScore
Recovery (3.1)551 prompts, six released familiesresidual, state, joint readoutlast prompt token (poetry: end of line 1), best rank over all source blocksarea under pass@k for k from 1 to 100, scaled to 0 to 1
Persistence (3.2)same 551 promptsresidual, stateevery token, blocks of the bandshare of later tokens at which the intermediate stays in the top ten
Word-list count (3.3)lists of 80 unrelated words, from the words released with the original studyresidual, stateeach comma, blocks of the bandwords in the top 25, averaged over commas 21 to 80
Verbal report (4.1)140 prompts per model, ourssteering and exchange on the residual stream, the state, or bothevery prompt token, blocks of the bandtarget among the five most likely next tokens
Two-hop reasoning (4.2)90 released itemsas aboveas abovegreedy output is the answer of the new step
Flexible generalisation (4.2)192 swaps from 64 released promptsas aboveas abovegreedy output is the function's answer for the new argument
One edit, several questions (4.4)40 swaps among 41 countries, ourssteering and exchange on the statetokens of the first sentence, blocks of the bandscored questions whose greedy answer becomes the new country's

The band is the middle half of the blocks (Appendix C lists them for each model). Every edit is paired with three random directions of the same norm at the same sites. To make submitted Table 2 easier to interpret, we have added an example prompt and the scoring rule for each of its measures in Table 14 (Appendix E). The table itself is now Table 15, next to those examples.

Q3. Which state lens Eq. 4 uses.

The full-path lens. The equation is now Eq. 3, and Section 2.1 states that r_s is computed with the full-path state lens and that state readouts and state edits use it unless we say otherwise.

We have also moved the readout-only variant to the appendix: its definition is in Appendix A.2 and its recovery scores are in Table 6. This makes room within the page limit for the new content, and it leaves the main text with a single state lens, the full-path one, so the question of which variant a result uses no longer arises there.

Comments on the paper

1 comment

To all reviewers and the area chair:

We thank the reviewers for their careful reading. Three changes in the revision bear on more than one review, so we summarise them here. The individual replies give the details.

The paper is reorganised. Section 2 now holds the method alone: the two lenses, the models and the edits. The experiments are then split into two sections by what they do. Section 3 has the reading experiments, where we leave the model untouched and look at what the two lenses can read from it. Section 4 has the editing experiments, where we change a concept inside the model and check how its answer changes. We have drawn new figures to present the results more clearly. Each experiment now opens with an example prompt and the answer expected from it, so the reader understands the experiment before seeing any number. Table 1 gives a roadmap of the sections.

Every edit now runs at a controlled size. Reviewer jpXs asked whether steering's advantage came from a larger perturbation. We now measure size in units of the exchange's edit (the exchange is the original method of Gurnee et al., 2026) and compare the operators in two ways: both at the exchange's original size, and each at its own best size, chosen on held-out trials. The result is stronger than what the submission reported. With both operators at the exchange's original size, steering is ahead in verbal report in all 14 cells. At that size, with the state edited alone, steering is ahead of the exchange in every recurrent model on all three tasks (Table 12). With each operator at its own best size, steering is ahead of or level with the exchange in all 29 comparisons, and significantly ahead in 19 (Section 4.3).

Two additions. A new experiment edits one concept in the state and then asks several different questions, with no question token edited (Section 4.4). The Qwen3.6-27B results are now measured in this study (Appendix E).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.