acceptodds
Under review as a conference paper at ICLR 2027

Concept J-Lens: From Token Binding to Concept-Level Control in the Jacobian Lens

Abstract

Chain-of-thought traces provide a useful but potentially fragile window into model reasoning, motivating complementary methods that read hidden states directly. The Jacobian Lens (J-Lens) maps hidden states to vocabulary-indexed directions that can be read and causally manipulated. Yet it remains unclear whether these directions capture the semantics underlying their indexing tokens or merely the tokens themselves. We test this distinction using writing interventions in Entity and Relation diagnostics, which require directions to generalize across alternative names and input-dependent relational outputs. Instead, J-Lens directions behave as token-level controls, predominantly promoting their construction tokens—a phenomenon we call token binding. Our analysis suggests that this behavior arises from fixed vocabulary targets and corpus-averaged Jacobians, neither conditioned on task examples. We therefore introduce Concept J-Lens (CJL), which maps task-defined targets through prompt-specific Jacobians and aggregates the resulting local directions into a reusable concept-level steering direction. Target–Jacobian ablations support this construction-level explanation. Across six language models and 25 tasks spanning five families, CJL raises average concept-level success from 4.18% to 60.47%, performs comparably to DiffMean activation steering overall, and is less sensitive to steering strength on Relation tasks.

Primary area:
interpretability and explainable AI
Keywords:
activation steering, large language models, mechanistic interpretability, concept-level control, Jacobian Lens, representation engineering

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_xh2POpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This work proposes Concept J-lens, an extension to the original Jacobian Lens for reusable concept control in LLM steering. Authors argument J-lens suffers from "token binding", in which the lens overly promote the token used for its construction instead of the underlying concept it would entail. To overcome it, authors propose a concept direction construction, using several input prompts related by a common task to build a final concept direction vector that allows more fine steering.

Critical strengths

Concept J-Lens could be a useful incremental steering method for specific cases where we want to not only steer but investigate the subspace a single concept would form within the latent space. It starts from a supervised set of supplied demonstrations and desired outputs and applies SVD to the set of gradients to extract the singular vector that builds the concept direction (or its opposite since it is sign-oriented). This strategy is sound, and could be expanded to analyze the singular values’ energy to understand the breadth of a subspace and fine-tune the construction of the concept vector.

Critical weaknesses
  1. Limited comparisons

There are substantial gaps in the framing and empirical analysis that make the evaluation procedure seem ill-posed. Most experiments compare the technique only with J-lens, while one section compares with DiffMean, but the authors should run the further comparisons and all tables should show the results of all techniques, including a simple input prompt change, which could be the simplest steering baseline. I think the work lacks proper comparison with prior work, as I detail below.

Theoretical foundations:

Linear Representation Hypothesis is what formalized the connection between semantic concepts, linear measurements, model steering and linear intervention/interpolation. Authors should distinguish token-indexed directions, concept-measurement directions, concept-intervention directions, and task operators. This can help justify the Euclidean normalization and SVD application.

  • Park, Choe, and Veitch, The Linear Representation Hypothesis and the Geometry of Large Language Models, ICML 2024.

Also, Frame representation already shows a way to connect tokens to concepts and has not been extended to work with J-lens, but Top-k Concept-Guided Decoding demonstrates concept generation control.

  • Valois, Souza, Shimomoto, and Fukui, Frame Representation Hypothesis: Multi-Token LLM Interpretability and Concept-Guided Text Generation, TACL 2025.

Comparison Techniques:

ICV is the strongest missing methodological comparison, and authors should compare activation differences versus gradients as to isolate the contribution of the gradient from spectral aggregation.

  • Liu, Ye, Xing, and Zou, In-context Vectors: Making In Context Learning More Effective and Controllable Through Latent Space Steering, ICML 2024.

Some other works that should be used for comparison against CJL:

not cited:

  • Hernandez et al., Linearity of Relation Decoding in Transformer Language Models, ICLR 2024.

cited, but insufficient comparison:

  • Hendel, Geva, and Globerson, In-Context Learning Creates Task Vectors, Findings of EMNLP 2023.
  • Todd et al., Function Vectors in Large Language Models, ICLR 2024.

The J-Lens paper also has concept coordinate swaps that already tackle the limitations of the indexing token. Authors should further discuss the Template Lens extension and possibly use it as a baseline or comparison.

  1. Metrics and Clarity

From my understanding all metrics used are variations of "concept-level success rate", but they way it is computed changes slightly for each cases, which is only briefly explained in section A.3. This needs better explanation and more metrics need to be used in the experiments. Also, there is no qualitative demonstration in the main paper, only a few minimal examples in the appendices, but I recommend having a proper qualitative section and quantitative sections because in general steering can have misleading metrics, with random or unrelated responses performing unusually well sometimes.

For example, the Entity metric seems to test only lexical relations, not semantics. Entity test prompt is no more than “Give a country as an example.”, but J-Lens construction token "America" is excluded while accepting alternative expressions such as "USA" or "United States", but this is not testing if CJL represents the entity semantically. I think that is possibly why Table 10 reports Entity success of 92.71% with a single example, but 91.46% with 16 examples, telling me this test is not strong enough to verify multiple examples are needed for entity representation. In this case, leave-one-out evaluation or cross-language references could provide stronger evidence than this metric.

In general, examples of accepted and rejected outputs would make these metrics substantially easier to interpret and it would be appreciated if equations and schematics could be complemented by pseudocode and proper environments explaining the logic, inputs, outputs and each aspect in a more sensible order.

Questions for authors
  1. Definition of concept. The authors constantly use "concept-level control" terms, but it is unclear what or how a concept is defined in this work. In my view, representing a concept and expressing an entity through synonyms or a steering function could be different, while failing to build antonyms does not indicate lack of semantic content. The approach here seems to be close to how Linear and Frame Representation Hypothesis works (among others) chose to use Counterfactual pairs or WordNet synsets (synonym sets) to build up concepts, but I must ask for a more formal definition.

  2. Improper avg metric. Section 5.1 states there are 25 tasks across 5 families, with Nonsense and Reject having a single task each, but the Avg. metric is an unweighted average of each family, so Nonsense and Reject alone contribute 40% of the weight despite having no more than 8% of the tasks. A properly weighted average could drop results by more than 10 percentage points. Why was this approach chosen and how does it affect reported results?

  3. In that sense, without the family-weighted mean, does the comparison with DiffMean sustain itself? The reported difference between CJL and DiffMean is only 1.18 percentage points. Could authors provide a breakdown of DiffMean results for finer comparison?

7 comments
AuthorsOpenReview ↗

Response to Reviewer xh2P (1/2)

Thank you for the thorough review.

W1. Limited comparisons and theoretical foundations

CJL outperforms the added ICV, TV, and FV baselines, and the crossed ablation supports its use of target gradients and SVD. We also clarify the construction’s relation to LRH and FRH.

  1. Comparisons.

    • With the same construction examples and strength-selection protocol, CJL achieves 68.48% task-average success, compared with 62.10% for ICV, 57.60% for TV, and 56.15% for FV (Table A). The revision will add these Qwen3.5-4B baselines to the main comparison.

    • Under mean pooling, target gradients improve Entity success from 87.50% to 91.46% compared with activation differences. SVD further raises gradient-based Relation success from 47.25% to 57.88% (Table B). Together, these results show the complementary benefits of target gradients and SVD in CJL.

  2. Theory.

    • Under LRH, we distinguish the fixed token-indexed direction w[τ], the task-defined measurement sᵢ = wᵢᵀ Norm_f(h(L, bᵢ; pᵢ)), and its local gradient gᵢ = J̃ᵢᵀ wᵢ, where J̃ᵢ is the position-averaged prompt Jacobian. CJL aggregates these gradients into a shared direction for intervention. The task specification x ↦ A_c(x), where A_c(x) = {y : R_c(x, y) = 1}, describes the outputs realizing the target property; R_c is defined below.

    • We normalize each target vector wᵢ before computing its gradient. Uncentered SVD then finds the unit axis that maximizes the sum of squared projections of the local gradients. The steering scale ρ‖h‖₂ is applied separately. LRH’s causal inner product and an analysis of singular-value energy are natural extensions.

    • FRH builds synset concepts from multi-token frames for decoding. CJL uses task-defined outputs to build steering directions that guide the model to express the target concept.

  3. Related work.

    • Hernandez et al. aggregate local linear approximations to study relation decoding. Our shared-Jacobian control examines how averaging affects steering. On Relation, pairing each target with its own prompt’s Jacobian achieves higher success than using a shared Jacobian, further supporting CJL’s construction. We will cite Hernandez et al. and clarify this connection in the revised paper.

    • TV and FV extract functions from in-context activations. CJL maps task-defined outputs through prompt-specific Jacobians. Table A compares the resulting directions.

    • Template Lens extends the index from a token to a phrase. Coordinate swaps replace an active coordinate, and in our Qwen3.5-4B swap test the model often reproduces the swapped-in token. We will discuss how both approaches relate to the concept-level control studied here.

Table A. Qwen3.5-4B success rates (%) under matched budgets. Success rates for Concept J-Lens and DiffMean are averaged over three construction sets. Family avg. weights the five task families equally, and Task avg. weights all 25 tasks equally.

MethodRelationEntityStyleNonsenseRejectionFamily avg.Task avg.
Unsteered0.2500.3601.250.370.25
TV55.574.5846.4381.2531.2557.857.6
FV5645.8369.6447.533.7550.5456.15
ICV4776.2568.2168.7578.7567.7962.1
DiffMean50.1387.578.579566.2575.4969.5
J-Lens613.754.4661.25017.099.4
Concept J-Lens53.2992.0164.2995.8381.2577.3368.48

Table B. Qwen3.5-4B success (%) by signal and pooling (one construction set).

SignalPoolingRelationEntity
Activation differencesMean49.7587.5
Activation differencesSVD5085
Target gradientsMean47.2591.46
Target gradients (Concept J-Lens)SVD57.8891.46
AuthorsOpenReview ↗

Response to Reviewer xh2P (2/2)

W2 & Q1. Definition of concept and validity of Entity evaluation

A concept is a high-level semantic or pragmatic property. CJL can elicit valid entity names absent from construction.

  1. Definition of a concept (Q1). A concept is a high-level semantic or pragmatic property that admits different linguistic realizations. We specify it through R_c(x, y), which equals 1 when response y instantiates the property in input context x, and 0 otherwise. The relation is specified by the concept’s meaning, independently of construction examples or internal representations.

    • Concept representation concerns how a model encodes a concept, while concept-level control concerns how an intervention makes the model express it. We evaluate concept-level control on held-out queries and aliases. For example, controlling the antonym relation should make the model answer small when given big.
  2. Metrics and alias evaluation (W2). Relation uses input-specific accepted answers; the other families use a method-blind judge. Responses must satisfy both task and validity criteria.

    • One example already yields high Entity success (paper Table 10). In Table C, we further test generalization beyond construction names by fixing N = 16 and varying the number of distinct aliases k. CJL produces valid unseen aliases in 45.00% of answers with k = 1 and 22.08% with k = 3, versus 13.75% for J-Lens. Its control therefore extends beyond the names used in construction.
  3. Presentation. The revision will include the definition and representation–control distinction, accepted and rejected outputs for each family, scoring rules, and full-pipeline pseudocode.

Table C. Qwen3.5-4B Entity alias comparison (six-task average, %, N = 16). Construction-alias reuse counts outputs that use a name from the method’s own construction set. Both rates use all queries as the denominator.

MethodConstruction-alias reuseValid unseen alias
J-Lens (one index)72.513.75
Concept J-Lens, k = 137.545
Concept J-Lens, k = 244.1738.75
Concept J-Lens, k = 364.1722.08

Q2 & Q3. Does the comparison with DiffMean hold without family weighting?

The comparison holds: CJL and DiffMean remain comparable when every task receives equal weight (55.35% vs. 55.46%, Table D).

  1. Family and task weighting (Q2). Each family probes an important capability, so we weight the five families equally. With equal task weights, CJL’s average changes from 60.47% to 55.35%, compared with 55.46% for DiffMean (Table D).

  2. Our goal and practical benefits (Q3). Our aim is not only to build a stronger steering method. We start from a question about J-Lens: do its token-indexed writing directions control the concept a token expresses, or only the token itself? Understanding this distinction matters for interpreting interventions with J-Lens. We identify token binding and hypothesize that it arises from the fixed token target and corpus-averaged Jacobian. We therefore develop CJL to address these two components. CJL’s effectiveness and the targeted ablations further support our explanation. It also achieves success comparable to DiffMean at similar total runtime (1.02× vs. 1.03× on Qwen3.5-4B). Across three models, at the same 50% success threshold, CJL’s average window width on a log scale is 1.36–1.48 times DiffMean’s where both methods reach the threshold. These wider windows give users more flexibility in choosing the steering strength.

Table D. Success rates (%) averaged across six models. Family avg. weights the five task families equally. Task avg. weights all 25 tasks equally.

MethodRelationEntityStyleNonsenseRejectionFamily avg.Task avg.
J-Lens3.252.993.0111.460.214.183.33
DiffMean39.562.9969.9482.7141.3259.2955.46
Concept J-Lens47.2474.854788.0645.2160.4755.35

Ethics. Incorrect answers and inappropriate refusals

We will add an Ethics and Broader Impact section.

  1. Risks. Nonsense and Reject show that steering can induce incorrect answers and refusals of benign requests, potentially enabling misinformation or denial of legitimate assistance.

  2. Mitigation. CJL construction requires access to model internals and gradients. Restricting who can add or modify steering directions can help reduce misuse.

Summary

In summary, the added experiments further support CJL’s construction and effectiveness in concept-level control. The theoretical discussion clarifies its rationale and relation to prior work. We will incorporate these results, explanations, and the proposed presentation improvements, together with the relevant citations, into the revised paper.

AuthorsOpenReview ↗

Follow-up on Our Response

As the discussion period is coming to an end, we would like to gently follow up and check whether our response has addressed your concerns. If so, we would appreciate it if you could consider updating your score. We are happy to answer any remaining questions.

Reviewer_Evp6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper finds that J-Lens steering often promotes its target token rather than the behavior that token describes. It proposes Concept J-Lens, which combines gradients from task examples into a reusable steering direction.

Critical strengths
  1. The paper makes a useful distinction between promoting a word and controlling the behavior it describes. The Relation tasks test this distinction through answers that must change with the input.

  2. The method follows naturally from the proposed explanation of token binding. The target–Jacobian ablations show that changing both components works better than changing either alone.

  3. Experiments cover six models and 25 tasks. Construction, development, and test sets are separated, with steering strength selected on development data.

  4. The paper compares against DiffMean and examines sensitivity to steering strength. It appropriately describes the overall performance as comparable rather than claiming consistent superiority.

Critical weaknesses
  1. The token-binding conclusion is based on repeated additive steering during generation. It may not extend to other J-Lens interventions, such as swapping internal concept coordinates.

  2. The Entity evaluation accepts aliases already used to construct CJL directions. This leaves unclear whether CJL learns broader entity control or mainly promotes a larger set of target expressions.

  3. CJL receives task demonstrations and output targets, while the J-Lens baseline uses a single target token. Comparisons with task-vector and function-vector methods using the same examples would better establish its contribution.

  4. The ablations do not isolate whether matching each target to its specific prompt is essential. Shuffling these pairings would provide a more direct test of the proposed explanation.

  5. The practical advantage remains unclear because performance is similar to DiffMean and construction requires backpropagation. Explicit instructions also achieve high success on the diagnostic models.

  6. The strength-robustness analysis covers only one model and uses thresholds relative to each method’s own peak. Most model results also rely on a single construction set.

Questions for authors
  1. Does token binding persist with single-step interventions or J-Lens coordinate swaps?

  2. Can CJL produce valid entity aliases that were excluded from direction construction?

  3. How does performance change when targets are shuffled across construction prompts?

  4. How does CJL compare with task-vector or function-vector methods built from the same examples?

  5. What practical benefit does CJL offer over explicit prompting or DiffMean? Please report construction time and memory costs.

  6. Does the wider steering window persist across other models and at matched absolute success thresholds?

8 comments
AuthorsOpenReview ↗

Response to Reviewer Evp6 (1/3)

Thank you for the thorough review.

W1 & Q1. Does token binding persist under single-step interventions or coordinate swaps?

J-Lens binding persists under single-step injection, whereas CJL substantially reduces it. J-Lens also frequently reproduces the swapped-in token under coordinate swaps (Tables A and B).

  1. Single-step injection. J-Lens binding remains at 52.50% on Relation and 60.00% on Entity. With the same intervention, CJL achieves 45.00% success on Relation and 86.25% on Entity, with binding rates of 2.50% and 7.50%, respectively (Table A).

  2. Coordinate swaps. We use the official α = 1 swap operator at the middle (L/2) or late (3L/4) layer on all prompt positions before the first generation step, then generate freely. At the late layer, J-Lens reproduces the swapped-in token in 77.50% of Entity answers and 56.25% of Relation answers. On Relation, 2.75% give the correct downstream answer without reproducing that token (Table B).

Table A. Qwen3.5-4B single-step injection results (%). Binding counts task failures that reproduce the J-Lens index token.

MethodRelation success↑Entity success↑Relation binding↓Entity binding↓
J-Lens8.752052.560
Concept J-Lens4586.252.57.5

Table B. Qwen3.5-4B J-Lens coordinate swaps (%, α = 1, one layer). Swap counts correct answers for the new target that do not repeat the swapped-in token. Binding counts answers that repeat the swapped-in token.

LayerRelation swapEntity swapRelation bindingEntity binding
L/23.251.4646.2568.75
3L/42.751.0456.2577.5

W2 & Q2. Can CJL produce aliases excluded from construction?

CJL produces valid entity names absent from its construction set: even with three construction aliases, the rate is 22.08%, compared with 13.75% for J-Lens (Table C).

  1. Setup. On the same queries for six Entity tasks, we keep 16 construction examples and vary the number of distinct aliases used to build CJL. J-Lens uses its single index token.

  2. Results. CJL produces valid unseen aliases in 45.00% of answers with one construction alias. As more common names enter the construction set, the unseen-alias rate decreases. Even with three construction aliases, CJL still produces valid unseen names in 22.08% of answers, compared with 13.75% for J-Lens. These results show that CJL’s control extends beyond the names used to construct the direction.

Table C. Qwen3.5-4B Entity alias comparison (six-task average, %, N = 16). Construction-alias reuse counts outputs that use a name from the method’s own construction set. Both rates use all queries as the denominator.

MethodConstruction-alias reuseValid unseen alias
J-Lens (one index)72.513.75
Concept J-Lens, k = 137.545
Concept J-Lens, k = 244.1738.75
Concept J-Lens, k = 364.1722.08
AuthorsOpenReview ↗

Response to Reviewer Evp6 (2/3)

W3 & Q4. How does CJL compare with TV and FV?

With the same construction examples and tuning budget, CJL outperforms TV, FV, and ICV (Table D).

  1. Comparison setup. We added TV, FV, and ICV using the same construction examples and strength-selection protocol as CJL.

  2. Results. On Qwen3.5-4B, CJL achieves 68.48% task-average success, compared with 57.60% for TV, 56.15% for FV, and 62.10% for ICV.

Table D. Qwen3.5-4B success rates (%) under matched budgets. Success rates for Concept J-Lens and DiffMean are averaged over three construction sets. Family avg. weights the five task families equally, and Task avg. weights all 25 tasks equally.

MethodRelationEntityStyleNonsenseRejectionFamily avg.Task avg.
Unsteered0.2500.3601.250.370.25
TV55.574.5846.4381.2531.2557.857.6
FV5645.8369.6447.533.7550.5456.15
ICV4776.2568.2168.7578.7567.7962.1
DiffMean50.1387.578.579566.2575.4969.5
J-Lens613.754.4661.25017.099.4
Concept J-Lens53.2992.0164.2995.8381.2577.3368.48

W4 & Q3. What happens when targets are shuffled?

Shuffling targets within Relation tasks reduces success from 57.88% to 6.21%; shuffling targets across entities reduces success on the original entity from 91.46% to 3.54% (Table E).

  1. Setup. We keep the construction prompts and their Jacobians fixed. For Relation, we shuffle targets among prompts within each task. For Entity, we shuffle targets across entities and score answers against the original entity.

  2. Results. Shuffling targets within Relation tasks reduces success from 57.88% to 6.21%. Since the prompts, Jacobians, and target set remain unchanged, this result directly supports matching each target with its own prompt. On Entity, assigning targets from other entities reduces success on the intended entity from 91.46% to 3.54%, showing that the construction targets determine which entity the direction promotes.

Table E. Qwen3.5-4B success rates under matched and shuffled target–prompt pairings (%).

Target–prompt pairingRelationEntity
Shuffled6.213.54
Concept J-Lens (matched)57.8891.46
AuthorsOpenReview ↗

Response to Reviewer Evp6 (3/3)

W5 & Q5. What practical benefit does CJL offer over DiffMean?

CJL achieves success and total runtime comparable to DiffMean’s, with wider effective steering windows.

  1. Our goal. Our aim is not only to build a stronger steering method. We start from a question about J-Lens: do its token-indexed writing directions control the concept a token expresses, or only the token itself? This distinction matters for how researchers interpret interventions with J-Lens, an increasingly adopted interpretability tool. We identify token binding and hypothesize that it arises from the fixed token target and corpus-averaged Jacobian. We therefore develop CJL to address these two components. CJL’s effectiveness and the targeted ablations further support our explanation. It also achieves success comparable to DiffMean, with wider steering windows at similar cost.

  2. Practical benefits. Across six models, CJL achieves success comparable to DiffMean (60.47% vs. 59.29%, paper Table 4). At the same 50% success threshold, CJL has wider steering windows across three models, giving users more flexibility in choosing the steering strength (Table G below).

  3. Time and memory. On Qwen3.5-4B, CJL takes 0.04× to build and DiffMean takes 0.02×, relative to the time for an unsteered batch. Their total times are 1.02× and 1.03×, respectively (Table F). During construction, CJL retains the activations needed to backpropagate through one prompt at a time and stores a 16 × d gradient matrix for SVD, with parameter gradients disabled. During generation, both methods store one steering vector per task.

Table F. Qwen3.5-4B relative total cost (unsteered = 1×). Cost includes construction and test generation.

MethodCost
Unsteered1.00×
Prompting1.03×
ICV1.20×
FV2.91×
TV1.01×
DiffMean1.03×
Concept J-Lens1.02×

W6 & Q6. Do wider windows persist across models and absolute thresholds?

At the same 50% success threshold, CJL has wider steering windows than DiffMean across all three models on task–set pairs where both methods reach the threshold (Table G).

  1. Setup. In the paper, Fig. 4 measures how much of each method’s peak gain is retained on Qwen3.5-4B. We extend the Relation evaluation to three models and apply the same absolute success threshold to both methods, using three independent construction sets per model.

  2. Results. CJL reaches 50% success on more task–set pairs on every model. On pairs where both methods reach this threshold, CJL’s average window width, measured on a log scale, is 1.36–1.48 times that of DiffMean. Thus, the wider effective range persists across models at the same success threshold.

Table G. Share (%) of task–set pairs that reach 50% success somewhere in the strength sweep. Each model has 30 pairs (10 Relation tasks × 3 sets). The width ratio is Concept J-Lens’s mean log-window width divided by DiffMean’s, calculated on task–set pairs where both methods reach 50% success.

ModelConcept J-LensDiffMeanWidth ratio
Qwen3.5-4B73.3363.331.48×
Qwen3.5-9B-Base83.3373.331.38×
Llama-3.1-8B-Instruct66.6736.671.36×

Summary

Together, the added intervention and alias tests, pairing controls, and baseline and robustness comparisons further strengthen our explanation of token binding and demonstrate CJL’s generalization and practical benefits. We will incorporate these findings and refine the corresponding explanations in the revised paper.

Reviewer_o7mFOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies whether J-Lens directions control the behavior associated with a concept or simply encourage the model to output its name. The authors identify this “token binding” problem in their experiments and propose Concept J-Lens, which combines task-defined output targets with prompt-specific Jacobians. It uses SVD to aggregate local gradients into a reusable steering direction without updating model weights. Across six models and 25 tasks, average concept-level success increases from 4.18% to 60.47%. The paper provides a useful distinction between producing concept names and controlling behavior, supported by targeted ablations. However, it does not consistently outperform DiffMean, and the paper lacks direct comparisons with task-vector and function-vector methods as well as a runtime analysis.

Critical strengths
  1. The paper clearly distinguishes naming a concept from acting on it. The entity and relation tasks provide concrete tests of this important difference.

  2. Experiments cover six models and 25 tasks. The ablations show that task-defined targets and prompt-specific Jacobians work best together.

Critical weaknesses
  1. The method does not consistently outperform existing steering methods. Overall performance is close to DiffMean (60.47% versus 59.29%), and direct comparisons with task-vector or function-vector methods are missing.

  2. Constructing steering directions requires per-example backpropagation and SVD, but the paper provides no runtime comparison with simpler methods such as DiffMean.

Questions for authors
  1. How does CJL compare with task-vector and function-vector methods under comparable data and tuning budgets? Reporting construction time would also help assess its practical value.

  2. Would a shared Jacobian averaged over the same 16 construction examples perform similarly?

\qquad \mathbf g_i^{\mathrm{shared}}=\bar J_{\mathrm{task}}^\top\mathbf w_i.$$ Here, $J_i$ is averaged over prompt positions. Keep the targets, readout, normalization, and SVD aggregation unchanged.
4 comments
AuthorsOpenReview ↗

Response to Reviewer o7mF (1/2)

Thank you for the thorough review.

W1 & Q1. CJL does not consistently outperform existing steering methods

Outperforming existing steering methods is not the only goal of CJL. J-Lens is an increasingly adopted interpretability tool whose token-indexed directions are used both to read and to intervene on hidden states, so understanding what its writing directions control is important. Our causal interventions show that these directions mostly promote their construction token rather than the underlying concept (token binding), which we see as a standalone contribution. CJL’s effectiveness further supports our explanation of token binding, while its performance is comparable to existing steering methods.

  1. Comparison with DiffMean. Across six models, CJL achieves 60.47% success, compared with 59.29% for DiffMean (paper Table 4). The added experiments show wider steering windows for CJL across three models at the same 50% success threshold. On task–set pairs where both methods reach this threshold, CJL’s average window width, measured on a log scale, is 1.36–1.48 times that of DiffMean. This gives users more flexibility in choosing the steering strength.
  2. Additional baselines. We added TV, FV, and ICV using the same construction examples and tuning budget. On Qwen3.5-4B, CJL achieves 68.48% task-average success, compared with 57.60% for TV, 56.15% for FV, and 62.10% for ICV (Table A). These results further demonstrate CJL’s effectiveness against other widely used steering methods.

Table A. Qwen3.5-4B success rates (%) under matched budgets. Success rates for Concept J-Lens and DiffMean are averaged over three construction sets. Family avg. weights the five task families equally, and Task avg. weights all 25 tasks equally.

MethodRelationEntityStyleNonsenseRejectionFamily avg.Task avg.
Unsteered0.2500.3601.250.370.25
TV55.574.5846.4381.2531.2557.857.6
FV5645.8369.6447.533.7550.5456.15
ICV4776.2568.2168.7578.7567.7962.1
DiffMean50.1387.578.579566.2575.4969.5
J-Lens613.754.4661.25017.099.4
Concept J-Lens53.2992.0164.2995.8381.2577.3368.48

W2. No runtime comparison with simpler methods

CJL’s total runtime is comparable to DiffMean’s: 1.02× versus 1.03× on Qwen3.5-4B, including construction and generation (Table B).

  1. Source of the cost. CJL uses 16 forward–backward passes on short prompts and one SVD of a 16-row matrix. The direction is then reused for every query of the task, and generation adds one vector per step, as in DiffMean.

  2. Measured time. Relative to the time for an unsteered batch, CJL takes 0.04× to build and DiffMean takes 0.02×. Their total times are 1.02× and 1.03×, respectively (Table B).

Table B. Qwen3.5-4B relative total cost (unsteered = 1×). Cost includes construction and test generation.

MethodCost
Unsteered1.00×
ICV1.20×
FV2.91×
TV1.01×
DiffMean1.03×
Concept J-Lens1.02×
AuthorsOpenReview ↗

Response to Reviewer o7mF (2/2)

Q2. Would a shared Jacobian perform similarly?

Per-example pairing achieves higher success than a task-shared Jacobian: 57.88% vs. 14.00% on Relation and 91.46% vs. 88.75% on Entity (Table C).

  1. Setup. We average the position-averaged Jacobians of the same 16 construction prompts into J̄ₜₐₛₖ and set gᵢˢʰᵃʳᵉᵈ = J̄ₜₐₛₖᵀ wᵢ, keeping the targets, readout, normalization, and SVD unchanged.

  2. Results. Using a task-shared Jacobian improves Relation success from 1.25% to 14.00% compared with the corpus average. Pairing each target with its own prompt’s Jacobian further raises success to 57.88%, while improving Entity success from 88.75% to 91.46%. This additional gain is especially large on Relation, where the correct answer changes with the input. Per-example pairing therefore helps the direction capture the relation required by each prompt.

Table C. Qwen3.5-4B success rates with task-defined targets and different Jacobian constructions (%).

Jacobian constructionRelationEntity
Corpus average1.259.79
Task-shared1488.75
Concept J-Lens (matched)57.8891.46

Summary

In summary, the added baseline comparisons, runtime measurements, and shared-Jacobian control further support our analysis of J-Lens and show that CJL achieves competitive control at comparable cost. We will incorporate these results and refine the accompanying analysis in the revised paper.

AuthorsOpenReview ↗

Follow-up on Our Response

As the discussion period is coming to an end, we would like to gently follow up and check whether our response has addressed your concerns. If so, we would appreciate it if you could consider updating your score. We are happy to answer any remaining questions.

Thanks for running these experiments, my concerns have been addressed. So I decide to maintain my positive score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank all reviewers for their careful reading and constructive suggestions. Below we restate the contributions of the paper and summarize the key updates from the rebuttal.

Contributions

  1. Diagnosing token binding in a widely used interpretability tool. The Jacobian Lens (J-Lens) reads and writes hidden states through token-indexed directions (Gurnee et al., 2026), and it is increasingly used to study model internals. Recent work uses these directions to surface model algorithms (Bhatia et al., 2026), read intermediate variables (Ayonrinde & Lindsey, 2026), build uncertainty axes from anchor tokens (Braun et al., 2026), trace and edit the residual stream (Zahedi et al., 2026), and inject concepts to test introspection (Garcia, 2026). Several of these uses treat a token-indexed direction as the concept its token names. We test this assumption through causal writing and find that J-Lens directions mostly promote their construction token rather than the concept: for example, an antonym direction makes the model output antonym instead of small when given big. We call this token binding. It implies that conclusions drawn from J-Lens interventions should be checked at the concept level.
  2. Explaining the cause of token binding. We trace token binding to two components of the J-Lens writing construction: a fixed token target and a corpus-averaged Jacobian. Replacing either component alone leaves Relation success low (1.25% and 8.00%, versus 6.00% for J-Lens), while replacing both raises it to 57.88% (paper Table 2).
  3. Concept J-Lens (CJL) as a fix. Building on this analysis, CJL pairs task-defined targets with prompt-specific Jacobians. Across six models and 25 tasks, it raises concept-level success from 4.18% to 60.47%. It performs comparably to DiffMean while offering wider effective steering windows and higher data efficiency.

Key Updates from Rebuttal

  1. Formalized definition of concept-level control. We refined the definition of concept-level control using matched counterfactual pairs and added experiments to assess it. CJL preferentially elicited the intended concept over related alternatives, produced unseen aliases across languages, and often preserved answer correctness while changing style.
  2. New baseline comparisons. We added comparisons with ICV, TV, and FV, including three bidirectional Relation tasks, and evaluated CJL against DiffMean at matched construction sizes. On Qwen3.5-4B, CJL outperformed ICV, TV, and FV overall, surpassed TV and FV on all three bidirectional tasks, and achieved higher success than DiffMean at every matched construction size tested.
  3. Additional intervention and ablation studies. We extended the token-binding analysis to single-step injections and coordinate swaps, and added shared-Jacobian and shuffled-pairing controls. Token binding persisted in both intervention settings, and both controls substantially reduced Relation success.
  4. Reporting improvements. We will report both family-weighted and task-weighted averages and revise our discussion of construction size.

We thank the reviewers for their thoughtful feedback and continued engagement throughout the discussion. We will incorporate the new results and clarifications into the revised paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.