acceptodds
Under review as a conference paper at ICLR 2027

What the Nonnegative Lens Misses: Energy-Matched Causal Tests of the J-Space in Open Language Models

Abstract

The Jacobian lens maps residual activations onto token-indexed output directions; "J-space" is a sparse nonnegative reconstruction of those atoms. We test how much causal work that reconstruction does on open-weight models with preregistered energy-matched interventions. When residual energy is matched, nonnegative J installs less of a contrastive concept shift than the leftover R=D−J. That R >> J ordering holds on the causal cells that report J−R—six families, rich and legacy reductions (future-summed as robustness), base models, and scales where capability permits ( 7B–32B)—and on a Gemma-4 invented-registry copy the remainder still transfers donor answers while reconstructed-J arms transfer none. Geometric capture stays in a 0.10–0.23 band from 0.5B to 32B. Nesting reconstructions on one dictionary is a selection-partition check: nonnegative OMP matches unconstrained least squares on the same support, while signed OMP raises reconstruction by 1.86–5.24 times at k=100 and flips J versus remainder on geography families. Increasing nonnegative budget fails to close the gap. Installation is not necessity: on Qwen2.5-32B-Instruct (n=28, descriptive), a post-hoc 2×2 of holding×answer-exposure shows an interaction, not an unconfounded holding effect; signed selection is unselective. Held-out known-attribute transport is strong for the full shift (0.82–0.93) and weak for the official nonnegative readout (0.15–0.23). The lens atoms track model knowledge; under the tested nonnegative-OMP reconstruction of contrastive shifts (rich-transport primary; future-summed as robustness), that code discards most of the equal-energy installable component. Preregistrations, frozen manifests, and code are released.

Primary area:
interpretability and explainable AI
Keywords:
mechanistic interpretability, Jacobian lens, J-space, global workspace, causal interventions, sparse coding, activation patching, construct validity, open-weight models

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_vyEROpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

unclear to me what the contributions are, as the entire paper appears to be agent-written in a unreadable fashion.

Critical strengths

clear agent-slop work

Critical weaknesses

clear ai-slop. Some sample sentences from the paper:

At matched residual-relative energy, operational J is not the dominant installable swap component—R=D−J is.

Two measurements already strain a simple workspace reading

Where the constraint sits. Nesting on one dictionary is a selection-partition san- ity check

Questions for authors

n/a

1 comment

Thank you for reading the submission. We accept that the writing obscured the contributions; the sentences you quoted were compressed to the point of being unclear. We have rewritten the paper around a single question: what a sparse nonnegative lens code carries when a model changes its answer. Internal labels have been removed and the motivation is stated plainly.

The contributions are now stated directly:

  1. A test design that compares parts of an activation change at matched energy, matched off-target disruption and matched alignment, with preregistered decision rules.

  2. Evidence that the nonnegative code of a factual shift carries a minority of what moves the answer, across 14 open checkpoints.

  3. A diagnosis: the shortfall comes from coding a signed change with a nonnegative code, not from the lens.

  4. A reconciliation with the original paper's magnitude-matched result.

The general comment summarizes the new experiments. As stated in the submission's AI-use statement, generative tools assisted with editing; responsibility for the text is ours. We would welcome any technical concerns behind the soundness score.

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

The paper studies the recent J-space work from Anthropic on open-source models. It considers the difference in activations between two contrastive prompts and projects it onto the J-space and its complement. It finds that the complement affects the answer much more than the projection. Since the difference of activation would require the subtraction of concepts from the activations, the paper allow the J-space to use negative coefficients and show that this allows the projection to capture more variance.

Critical strengths

None

Critical weaknesses
  • The paper gives a strong AI-generated feeling and is very hard to read. There is no intuition or story to help the reader understand the goal of the paper. The abstract is full of numerical results and jargon that cannot be understood through a first read with sentences such as “That R ≫ J ordering holds on the causal cells that report J−R—six families, rich and legacy reductions (future-summed as robustness), base models, and scales where capability permits (∼7B–32B)—and on a Gemma-4 invented-registry copy the remainder still transfers donor answers while reconstructed-J arms transfer none.”

  • The paper does not include a figure to make the reader understand its main ideas.

  • The paper is very short and only includes a small one paragraph related work section.

  • The main idea of the paper is to test how the difference between contrastive prompts in activations can be explained from the J-space. However, why should the J-space be able to explain the difference of activations ?

Questions for authors

Clarification:

  • Why is the main experiment of the paper important and what does it teach us about the role of the J-space ?
3 comments

Thank you. We agree the submission was hard to read and did not explain why the experiment matters. We have rewritten the paper around one question and one answer.

Why should J-space explain activation differences?

Gurnee et al. argue that J-space is where the causally relevant part of a concept representation lives. A causal role is defined by intervention, and the most direct intervention on a concept is to change it: replace France with Japan in a capital-of question and check whether the model's answer follows. The lens directions are answer-token directions pulled back into the residual stream. If the nonnegative lens code carries the causally relevant part of any activation change, it should carry this one. The experiment is therefore a direct test of the causal claim, in a case where the code should be at its strongest.

What we found.

The nonnegative lens code does not carry the change. When we add the code and the remainder back at equal energy, the remainder moves the answer far more than the code, in all 14 open checkpoints we tested. We expected the opposite. Three new controls explain the result:

  • The lens directions themselves are meaningful. At the same alignment with the change, they move the answer more than directions outside the lens.
  • The nonnegativity constraint is what fails. A change of answer has a sign: one concept goes down and another goes up. A code whose coefficients cannot be negative can represent the concept that goes up but not the one that goes down. Allowing signed coefficients, or coding each state separately and subtracting the two codes, recovers most of the effect.
  • The same account explains why Gurnee et al. found the opposite. With their own code, the J-space component wins on vectors that describe which concept is present, and loses on vectors that describe a change between two states.

The conclusion is that a nonnegative lens code describes what is present in a state well but does not capture how a state changes. Studies that use it to steer a model or to reason about edits should code each state rather than the difference, and report a signed readout alongside the nonnegative one.

Specific points.

  • Abstract: rewritten in plain language with four numbers.
  • Overview figure: a new schematic (Figure 1) shows how the shift splits into the lens code and the remainder, the equal-energy injection, and the result.
  • Related work: now a full section (Section 6) of three paragraphs, on vocabulary-space lenses, J-space and its follow-up work; sparse codes and the sign constraint; and intervention validity (steering, patching, construct validity and factual recall). An extended six-paragraph review is in Appendix E.
  • Length: the revision stays within the 10-page limit. Protocol details, the transport and ablation experiments, and the full results of the comparison with Gurnee et al.'s protocol moved to the appendix to make room for motivation, controls and discussion.
  • On your summary: JJ is a sparse greedy reconstruction (nonnegative OMP), not a projection, so JJ and R=D−JR = D - J are not orthogonal complements. Our capture number is a norm ratio, not explained variance; its square is the variance share. The revised Method section states this explicitly.

I acknowledge the response and would like to keep my original rating, as the new version is still mostly AI slop.

I thank the authors for their response but will be keeping my original score.

Reviewer_2EKBOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper evaluates sparse nonnegative reconstructions of contrastive residual shifts using the Jacobian lens. It decomposes each shift DD into a reconstruction JJ and remainder R=D−JR=D-J, then compares their effects under equal-energy interventions. Across the reported model families, the remainder changes answer margins more strongly than the nonnegative reconstruction. Further experiments compare signed and nonnegative coding, test transport of known attributes, and examine whether J-space ablation selectively disrupts tasks requiring intermediate information to be held.

Critical strengths

The authors do more than compare components at their native, unequal magnitudes. They include equal-energy random controls, repeat the comparison with the future-summed operator, and compare coding constraints on the same dictionary. These controls make the reported reconstruction results informative. I also appreciate the treatment of the necessity experiment: the holding-by-answer-exposure analysis reveals a confound in the initial interpretation, and the paper reports that rather than continuing to claim a clean holding-specific effect.

Critical weaknesses

My main concern is the object being reconstructed. DD is a signed transition between two states. Even if both states have accurate nonnegative codes, their difference can require negative coefficients. Sections 2 and 7 acknowledge this, but I could not find the corresponding control: reconstruct both endpoints nonnegatively and then difference the reconstructions before the matched-energy intervention. I understand that the conclusion is restricted to coding contrastive shifts. Without this comparison, however, it remains unclear whether the causal gap diagnoses a limitation of the concept code or a restriction imposed specifically on the edit used to test it. The signed-code advantage does not resolve that distinction.

Equal energy also does not ensure comparable disruption of the model’s other behavior. The primary ranking integrates target–foil margin changes over a dose grid extending to 4, while the limitations explicitly state that the registered KL and unrelated-control accuracy measurements were not implemented or retained. A larger margin change is therefore not yet evidence of more selective concept transfer. Isotropic random directions do not settle this comparison between two structured directions. The Gemma donor-transfer result is helpful, but it does not establish that the main multi-family AUC ranking persists at doses where both interventions preserve unrelated behavior.

Questions for authors
  1. Does remainder dominance persist when the intervention is constructed by differencing two nonnegative endpoint reconstructions, with the reconstruction budget and intervention energy stated explicitly?
  2. Also, does the JJ -versus- RR ordering persist in a low-disruption regime, assessed jointly through donor-answer success and preservation of unrelated behavior? Please include the corresponding dose-response measurements.
1 comment

Thank you for a careful review. Both concerns were right, and both changed the paper.

Q1. Does the remainder's advantage survive coding the endpoints separately?

We ran the control you describe, preregistered with frozen decision rules, on Qwen2.5-7B and OLMo-2-7B (100 items, five layers). Each endpoint state is coded with nonnegative OMP on the same dictionary, without centering, as the official readout reads a single activation, using kk = 13, 25, 100 or 400 atoms per endpoint (primary 100). Then Jdiff=NN(xB)−NN(xA)J_{diff} = NN(x_B) - NN(x_A) and Rdiff=D−JdiffR_{diff} = D - J_{diff}, both matched to DD's energy on the same dose grid as the other arms.

Results (J−RJ - R area; negative means the remainder moves the answer more):

  • Qwen2.5-7B: nonnegative code of DD -34.2 [-35.9, -32.4]; endpoint code -5.9 [-7.3, -4.4]; signed code of DD (k=100k = 100) +11.0 [9.7, 12.3].
  • OLMo-2-7B: nonnegative code of DD -18.4 [-19.6, -17.2]; endpoint code -0.9 [-1.9, +0.1]; signed code +6.4 [5.6, 7.1].

All three are from the endpoint-control run; the paper's budget sweep gives +10.3 and +6.1 for the signed code.

The remainder stays ahead on Qwen. On OLMo the difference is not significant, which matches your objection. A mean-centered variant leaves RR ahead on both models (-6.1, -2.4). The budget does not matter: kk = 13 to 400 per endpoint gives identical estimates, because each endpoint is coded by a median of 3 atoms and the two supports barely overlap. At dose 1 the endpoint code reaches 80% (Qwen) and 85% (OLMo) of DD's effect, against 22% and 31% for the nonnegative code of DD.

We have narrowed the claim: the remainder dominates the nonnegative code of the shift; with endpoint coding it keeps a smaller lead on one model of two. This also gives the paper its explanation. The endpoint difference puts negative weight on the original state's atoms, which a nonnegative code of DD cannot do.

Q2. Does the ordering hold at low disruption?

We added the measurements you asked for (preregistered; 50 items, two layers): next-token KL on 40 unrelated sentences, KL and accuracy on 40 unrelated one-word factual probes, and donor success, all per arm and dose. A dose counts as low-disruption for an arm if both KLs stay below the random direction's 95th percentile at dose 1.

  • Qwen2.5-7B: at the highest dose where both JJ and RR qualify (q=0.5q = 0.5), the R−JR - J flip rate is +0.22 [0.11, 0.34].
  • OLMo-2-7B: at the common dose (q=0.5q = 0.5), the full shift itself flips only 6% of answers, so the test is uninformative. This label was added after launch but before any full-run output was read. We disclose that without it the frozen rule would have read OLMo as "explained by disruption".
  • At dose 1 the two parts disturb unrelated text similarly, yet RR flips far more answers. Qwen: neutral KL 0.026 vs 0.040, flips 100% vs 5% (RR costs more probe accuracy, 0.89 vs 0.96). OLMo: neutral KL 0.020 vs 0.023, flips 89% vs 3%.

Dose-response curves for margin and KL are in the revised Figure 2. The per-dose breakdown now leads the results. The difference between the two parts peaks at dose 1-2 and narrows at dose 4, so large doses do not drive it.

Isotropic random does not settle a comparison between two structured directions.

Agreed. We added a preregistered structured null: a direction outside the lens span with the same cosine with DD as each code. Every lens code beats it (Qwen / OLMo): nonnegative +3.0 / +3.3 logits, signed +7.4 / +5.7, endpoint +11.3 / +8.9. The result is unchanged with the item's answer tokens removed from the dictionary. The remainders sit near or below the null (at most 0.22 logits above it).

Multi-family ranking at low disruption.

The disruption control covers two 7B models, which we state as a limitation; we no longer claim a multi-family ranking at matched disruption. The per-dose ordering at the natural dose also holds on BLOOM-7b1 (flips: RR 62% vs JJ 44%).

Comments on the paper

1 comment

We thank all three reviewers. The criticism we heard from everyone is fair: the submitted text was hard to read, its motivation was implicit, and it relied on internal labels. Reviewer 2EKB also identified two controls the paper needed. We ran both, plus three further preregistered or post-hoc analyses, and rewrote the paper around what the results show. The revised PDF is uploaded.

The question. Gurnee et al. propose that a sparse nonnegative code over Jacobian-lens directions ("J-space") is a privileged, causally relevant part of a model's state. Causal claims are tested by intervention, so we ask what that code carries when a model changes its answer. We take the activation difference DD between a factual prompt and its counterfactual (France to Japan in a capital-of question), split it into its nonnegative lens code JJ and the remainder R=D−JR = D - J, and add each back at equal energy.

Findings.

  1. The remainder moves the answer far more than the code, in all 14 open checkpoints tested (6.9B to 32B). This includes the shift's natural size: at dose 1 on three 7B models (100 items each), RR flips 57-80% of answers and JJ 2-44%.

  2. At matched off-target disruption (new, preregistered), the ordering holds on Qwen2.5-7B: the R−JR - J flip rate is +0.22 [0.11, 0.34] at the highest dose where both parts stay within a random direction's KL on unrelated text. On OLMo-2-7B that dose is too weak for the full shift itself to move answers (6% flips), so the test is uninformative there.

  3. Coding the endpoints (new, preregistered). Coding each state nonnegatively and subtracting the codes closes most of the difference: on Qwen2.5-7B from -34.2 to -5.9 [-7.3, -4.4], with the remainder still ahead; on OLMo-2-7B from -18.4 to -0.9 [-1.9, +0.1], which is not significant. We have narrowed the claim accordingly.

  4. The shortfall does not come from the lens directions (new, preregistered). At the same cosine with DD, every lens code moves the answer 3-11 logits more than a non-lens direction, also with the item's answer tokens removed from the dictionary.

  5. Reconciling with Gurnee et al.'s magnitude-matched result, which favors J (new, preregistered). With Gurnee et al.'s own k=16k = 16 code, the lens component beats its remainder on concept-presence vectors and loses on the contrastive shift. The original swap protocol has a weak positive control on our 7B models (9-13% vs. 88% reported for Claude), so model family remains a possible explanation we cannot test.

Revised thesis. The lens directions carry causal content, but a nonnegative code over them describes which concepts are present, not the change between two states. A change lowers some concepts and raises others; a code that cannot take negative coefficients can represent the concept that rises but not the one that falls. Signed codes and endpoint codes recover most of the effect.

Main changes. A new abstract, introduction and discussion built on this thesis; a schematic Figure 1; per-dose results in place of a dose-integrated headline; the disruption, endpoint and alignment controls in the main text; a full related-work section, with an extended version in the appendix; internal labels replaced by descriptive names; and claims narrowed where the new controls require it.

Preregistrations, frozen code and result files: https://anonymous.4open.science/r/jspace-iclr2027-anonymous-A1D0/

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.