- 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
- 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.
- 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
-
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.
-
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?
-
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
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.
-
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.
-
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.
-
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.
| Method | Construction-alias reuse | Valid unseen alias |
|---|---|---|
| J-Lens (one index) | 72.5 | 13.75 |
| Concept J-Lens, k = 1 | 37.5 | 45 |
| Concept J-Lens, k = 2 | 44.17 | 38.75 |
| Concept J-Lens, k = 3 | 64.17 | 22.08 |
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).
-
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).
-
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.
| Method | Relation | Entity | Style | Nonsense | Rejection | Family avg. | Task avg. |
|---|---|---|---|---|---|---|---|
| J-Lens | 3.25 | 2.99 | 3.01 | 11.46 | 0.21 | 4.18 | 3.33 |
| DiffMean | 39.5 | 62.99 | 69.94 | 82.71 | 41.32 | 59.29 | 55.46 |
| Concept J-Lens | 47.24 | 74.85 | 47 | 88.06 | 45.21 | 60.47 | 55.35 |
Ethics. Incorrect answers and inappropriate refusals
We will add an Ethics and Broader Impact section.
-
Risks. Nonsense and Reject show that steering can induce incorrect answers and refusals of benign requests, potentially enabling misinformation or denial of legitimate assistance.
-
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.
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.
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.
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.
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.
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.
Table B. Qwen3.5-4B success (%) by signal and pooling (one construction set).