{"id":"04c1726d-69d6-4a2f-af81-344fd0ce3e6d","arxiv_id":"2602.15293","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Dual steering—adding a linear probe in the dual (mean-parameter) space—provably preserves off-target distributions better than Euclidean steering for softmax-based models, under factorization assumptions.","lead":"This paper argues that the right geometry for softmax-based AI representations is information geometry, and uses it to introduce 'dual steering,' which moves a probe vector in the dual space rather than the raw representation space. The method is shown to preserve off-target behavior better than standard Euclidean steering on an LLM and a CLIP model.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Figure 7 shows the linear-probe constancy premise of Theorem 3 is violated, so the proven off-target optimality does not apply to the tested steering paths.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern: Theorem 3's off-target optimality requires the linear probe to be exact and the distribution to be concept-factorizable, and the paper's own Figure 7 demonstrates the probe-constancy premise fails on tested models. I agree this is the central soft spot. The mathematical core—Proposition 1 and the derivation of Theorem 3—is sound under its assumptions, and the empirical results do show dual steering preserves off-target distributions better than Euclidean steering in the tested settings. The issue is not internal inconsistency but external validity: the unqualified abstract claim is stronger than what the theorem supports, and the paper itself acknowledges the failure mode. The regularized-Newton implementation gap is also real but secondary; the probe-constancy violation is documented and directly attacks applicability of the theorem to the headline comparison. A conditional verdict remains appropriate: accept the geometry as a useful principle, but require the stated caveats to be carried into the claims, and ideally test whether dual steering remains near-optimal when the exact-probe premise is relaxed. No change to the reader's CONDITIONAL verdict is needed.","tokens_in":19596,"tokens_out":10471,"duration_ms":105690,"concrete_test":"For each concept and several target levels p, directly solve the constrained optimization min_λ D_KL(P^Z_λ0 || P^Z_λ) subject to P^W_λ(1)=p using gradient-based projected optimization over λ, and compare its off-target KL to the dual-steering path at the same p. Also compute the first term of (A.30) along the dual-steering path. If dual steering's off-target KL is within a small tolerance of this oracle and the first term is nearly constant, the probe-violation concern is mitigated; if dual steering is materially worse, the empirical superiority of dual steering is not explained by Theorem 3 and the claim must be stated as conditional.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on Theorem 3, specifically part (3.4). The proof decomposes D_KL(Pλ0 || Pλ) as in (A.30) and drops the first term because, on the hyperplane ΛW(c), the exact linear probe assumption (3.1) makes P^W_λ constant. Section C.2 and Figure 7 directly test this premise: along dual-steering paths, logit P(W=1|λ) is systematically lower than test-set contexts with the same β^T_W λ / ||β_W||. Therefore P^W_λ is not a function of the probe projection, the first term in (A.30) is not constant, and the constrained KL minimizer is not equivalent to the minimizer of D_KL(P^Z_λ0 || P^Z_λ). The optimality guarantee in (3.4) is thus not applicable to the actual steered distributions. Moreover, the headline experiments in Figure 4 align paths by the measured target probability P^W(1), not by β^T_W λ; with probe mismatch these are different level sets, so Theorem 3 does not even cover the plotted comparison. The paper acknowledges this in Section 5.3 ('If this condition is violated...'), but the abstract's unqualified claim that dual steering is proven to 'minimize changes to off-target concepts' overstates the result. The theorem itself is correctly derived under its assumptions; the soft spot is the gap between those assumptions and the tested regime. Concept-factorizability is also assumed without direct verification, making the external validity of (3.4) doubly conditional.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops an information-geometric view of softmax-based representation spaces, arguing that the natural geometry is Bregman/dually flat. It introduces primal and dual interpolation paths, and on that basis proposes \"dual steering\": instead of adding the probe direction β_W in the primal space as in Euclidean steering, one moves the dual coordinate along β_W. The central result, Theorem 3, states that the KL minimizer over the hyperplane Λ_W(c)={λ : β_W^⊤ λ=c} has dual coordinate φ(λ̂)=φ(λ_0)+tβ_W, and that, under a concept-factorizability condition, the same minimizer also minimizes the off-target KL divergence D_KL(P^Z_{λ_0}||P^Z_λ). The method is implemented via a regularized Newton update on the covariance (Hessian) and tested on Gemma-3-4B and MetaClip-2, with metrics measuring target-concept probability, counterfactual mass, off-target KL, and rank stability. The empirical section reports that dual steering preserves off-target distributions better than Euclidean steering across several binary concepts.","tokens_in":20008,"tokens_out":5211,"duration_ms":52906,"significance":"If the theorem applied to the experimental regime, this would be an important contribution: it gives a principled geometric reason for the failure of Euclidean steering and provides a concrete, falsifiable alternative. The proof of Theorem 3 is internally coherent, and the KL decomposition in Appendix A.2 checks out. Proposition 1 is also correctly proven. The paper usefully distinguishes primal and dual interpolation and supports the distinction with clear figures. The main weakness is not the internal derivation but the gap between the theorem's assumptions and the tested setting: the paper's own Appendix C.2 shows that the exact linear-probe assumption is violated on the tested models, and concept-factorizability is never verified. As a result, the unqualified claim in the abstract that dual steering is proven to minimize off-target changes overstates what is established. The empirical results are suggestive, but the theoretical guarantee currently covers an idealized regime rather than the actual implementation.","major_comments":[{"comment":"The proof of (3.4) relies on the assumption that, on the hyperplane Λ_W(c), the concept distribution P^W_λ is constant, so that the first term in Eq. (A.30) is independent of λ. Section C.2 and Figure 7 directly test this premise and find that dual-steering paths yield lower logit P(W=1|λ) than test contexts at the same β_W^⊤λ/||β_W||. Thus P^W_λ is not a function of the probe projection, the first term in (A.30) is not constant, and the constrained KL minimizer is not equivalent to the minimizer of D_KL(P^Z_{λ_0}||P^Z_λ). Section 5.3 acknowledges this, but the abstract and Section 3.1 state the optimality result without this qualification. The theorem is correct under its assumptions; the load-bearing issue is that the experiments are run where the assumptions are known to fail.","section":"§3.1 (Theorem 3), §A.2 (Eq. A.30), §C.2 (Fig. 7)"},{"comment":"The empirical comparison in Figure 4 aligns steering paths by the measured target-concept probability P^W_{λ_t}(1) on the x-axis, not by the probe projection β_W^⊤λ. Theorem 3 concerns the hyperplane Λ_W(c) where β_W^⊤λ is fixed. These are different level sets when the probe is miscalibrated, and Figure 7 shows that they do differ on the tested models. Consequently, even the qualitative comparison in Figure 4 is not directly covered by Theorem 3. The authors should either align paths by β_W^⊤λ in the main experiments or explicitly state that the theorem applies only to an idealized probe and that the empirical comparison is heuristic.","section":"§5.2 (Figure 4) vs. §3.1 (Theorem 3)"},{"comment":"The implemented method is not the exact dual-coordinate update φ(λ_t)=φ(λ_0)+tβ_W. Algorithm 1 solves (Σ_t+αI)v=β_W and takes a normalized step; the actual dual displacement is approximately Σ_t(Σ_t+αI)^{-1}β_W, which is not colinear with β_W when β_W is in the approximate null space of Σ_t. Section C.1's Figure 5 confirms the cosine similarity between the dual step and β_W is substantially below 1, especially initially. Moreover, for LLMs the covariance is truncated to Top-K tokens (Appendix B.2), changing the Hessian. No error bound or convergence guarantee is given for this approximation. The theorem does not establish the optimality of the regularized, truncated path actually evaluated.","section":"§4.2 (Algorithm 1), §5.2"},{"comment":"The off-target optimality result (3.4) requires concept-factorizability of P_λ for all λ in Λ_W(c) ∪ {λ_0}. This is a strong structural assumption: it requires the output space to split into counterfactual pairs with P_λ(y_i^w)=P^W_λ(w)P^Z_λ(z_i). The paper never tests this assumption. In Appendix B.2 the metrics define P^Z_λ(z_i)=P_λ(y_i^0)+P_λ(y_i^1), but that identity does not by itself establish the factorization used in the proof. Without factorizability, the decomposition (A.30) is not valid, and the interpretation of the KL term as an off-target divergence is not justified. The authors should either provide empirical evidence for factorizability on their counterfactual pairs or present (3.4) as a conditional statement that is not directly applicable to the reported experiments.","section":"Definition 2, §5.1, Eq. (3.4)"}],"minor_comments":[{"comment":"The rank-difference metric is defined as a weighted sum of absolute differences of inverse ranks, but the notation 'ranking_{P^Z_{λ_t}}(z)' is not formally defined. Please define how ties and out-of-vocabulary tokens are handled.","section":"§5.1"},{"comment":"The Top-K covariance truncation is described as using 'e.g., 20,000' tokens for LLMs, but no K is reported for the CLIP experiments or in the main figures. Since truncation affects the Hessian and hence the steering path, the value of K should be reported per experiment.","section":"§B.2"},{"comment":"The decomposition of reverse KL in Eq. (3.5) is stated without proof and without explicitly defining the domain of the sum over z. For consistency with Appendix A.2, a short derivation or reference would help.","section":"§3.2, Eq. (3.5)"},{"comment":"The model name appears as both 'MetaClip-2' and 'MetaClip' in different places; please standardize. Also, Figure 2's caption is dense and would benefit from a legend explaining the colored lines in the probability plots.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper's central theoretical derivation is sound under its stated assumptions, and the empirical phenomenon it identifies is plausible. However, the authors themselves show that the key assumption of Theorem 3 — exact linear-probe constancy — is violated in the exact experiments used to support the abstract's claim. This is a load-bearing gap between theory and practice. The revision should either restrict the theoretical claims, add experiments in controlled settings where the assumptions hold (e.g., synthetic softmax distributions with known factorizability), or provide a quantitative account of how the probe violation affects the off-target optimality. With such a revision the paper could be acceptable; in its current form the central claim is overstated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper has a genuine theoretical core and an honest disclosure of its own soft spot, but the abstract sells the guarantee too hard. The Bregman-duality treatment of softmax is a nice unification, and the primal vs. dual interpolation experiments are intuitive. Theorem 3 is correctly derived: the KL projection onto a probe hyperplane yields dual steering, and with concept-factorizability it minimizes off-target KL. The proof checks out. The regularized Newton implementation is pragmatic, and the empirical results consistently show dual steering beating Euclidean steering on off-target preservation. Credit also goes to Section 5.3, which openly says the probe-constancy assumption is violated and points to C.2 and Figure 7 — that is real intellectual honesty.\n\nThe soft spot is load-bearing, not cosmetic. Theorem 3's off-target optimality (3.4) requires the concept probability to be constant on the hyperplane. Figure 7 shows dual steering paths have lower logits than test contexts at the same projection, so the first term in (A.30) is not constant on the actual steered path. That means the proven guarantee does not apply to the experiments as run. The headline comparisons in Figure 4 align paths by measured target probability, not by probe projection, so the plotted curves are not even on the hyperplane the theorem talks about. Concept-factorizability is also assumed without direct verification. The implemented method uses a tuned regularized-Newton approximation without formal convergence guarantees. These are addressable, but the abstract's claim that dual steering is 'proven' to minimize off-target changes is too strong for what is actually established.\n\nWho gets value from this: anyone working on interpretability, activation steering, or the linear representation hypothesis. The framework is clear, the method is practical, and the empirical evidence suggests it works even where the theory is conditional. I would cite it. It deserves a serious referee, but with a request to align the claims with the assumptions — either soften the abstract or add experiments that probe the invariance assumption directly, e.g., test whether off-target preservation holds conditional on projection value rather than on achieved target probability.\n\nSend it to peer review. I would take it as a conditional: the core is sound, and the gap is known and disclosed. The right fix is disciplined claims plus a bit more evidence that the assumption approximately holds in the tested regime.","headline":"A clean conditional result: the geometry story is sound, the theorem is correctly derived, but the paper's own Figure 7 shows the key assumption fails in practice, so the abstract overstates what is proven.","tokens_in":20431,"tokens_out":1828,"would_cite":true,"duration_ms":22264,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Dual steering moves a softmax model's concept probability with minimal off-target change, and its formula is exact.","keywords":["information geometry","softmax","linear representation","steering","KL divergence","Bregman divergence","linear probe","interpretability"],"falsifier":"Measure, along an actual dual-steering path on a large language model, the concept logit logit P(W=1|λt) versus the probe projection βWᵀλt/||βW||, and compare to test contexts with the same projection. If dual-steering paths systematically give lower concept logits at matched projections than test contexts (as the paper's Figure 7 shows), then the probe-constancy assumption is violated and the proven off-target optimality no longer holds — this observation would settle whether the theorem's premise is met.","tokens_in":19475,"feed_emoji":"","tokens_out":1663,"duration_ms":17891,"temperature":0.7,"pith_summary":"This paper argues that the geometry of softmax representation spaces is not Euclidean but information-geometric, and that this distinction matters for steering model behavior. The authors prove a specific claim: to change a binary concept (like dog to cat) while disturbing everything else as little as possible, you should update the representation in the dual coordinate — the expected unembedding vector — rather than by directly adding the probe vector in the raw space. They call this dual steering, show it is the solution of a constrained KL-minimization problem, and verify empirically on language and image models that it preserves off-target probabilities and rankings better than standard Euclidean steering. If correct, this gives a principled geometric explanation for why additive steering leaks probability mass and provides a robust recipe for concept manipulation.","feed_headline":"Steer softmax models in dual coordinates to avoid off-target drift","feed_subtitle":"A new proof shows the right geometry for concept steering is curved, and the fix is a regularized Newton step.","key_machinery":"The central object is the dual coordinate map φ(λ)=∇A(λ)=E[γ|λ], the gradient of the log-normalizer (equivalently, the mean of the unembedding vectors under the softmax). The key identity is that KL divergence between softmax distributions equals the Bregman divergence of A, which turns steering into a projection problem: moving along βW in the dual space is the forward-KL projection onto the probe hyperplane, whereas adding βW in the primal space is an unprincipled Euclidean step. The proof decomposes the KL objective into an on-target term and an off-target term via a concept-factorization of the output distribution.","core_discovery":"The central claim is that the softmax family P(γ|λ)=exp(λᵀγ−A(λ)) carries a Bregman (dually flat) geometry, with primal coordinates λ and dual coordinates φ(λ)=E[γ|λ]. Within this geometry, the correct way to steer a representation onto the hyperplane defined by a linear probe βW (i.e., to set βWᵀλ=c) is to move the dual coordinate along the probe direction: φ(λ̂)=φ(λ0)+tβW. The paper proves (Theorem 3) that any minimizer of the forward KL divergence D_KL(Pλ0||Pλ) over that hyperplane has exactly this form, and that if the distribution is factorizable into a concept part and an off-target part, the same minimizer also minimizes the KL divergence of the off-target distribution. Thus dual stee","pith_inferences":["The same dual-coordinate argument likely extends to other exponential-family output layers, where the dual map is the expectation parameter; a testable extension is to apply dual steering to attention logits or mixture-of-experts gates.","If the probe-constancy assumption fails (as the paper's own Figure 7 suggests it does in practice), then dual steering may trade incomplete target steering against off-target distortion; a natural follow-up is to measure and correct for the deviation of the concept logit along the steering path.","The rank-preservation property of dual steering could be exploited beyond steering, e.g., in controlled generation or debiasing where maintaining a stable ordering of alternatives matters.","The paper focuses on the output softmax layer, but a direct empirical test for intermediate layers would assess whether the same dual-coordinate update remains optimal when representations are transformed by later layers."],"forward_implications":["Euclidean steering (adding the probe vector to λ) is a type error whenever the primal and dual spaces differ; the paper predicts it will leak probability mass to off-target outputs, which the experiments confirm.","Dual steering provides a concrete, implementable algorithm (regularized Newton updates on the covariance matrix) that works even when the Hessian is rank-deficient, because regularization nudges the distribution toward higher entropy.","The theory says dual steering preserves not just probabilities but also the ranking of off-target components, which is why it is more reliable than Euclidean steering for practical use.","The analysis explains when Euclidean steering can work: when the total counterfactual probability mass stays roughly constant along the path, so the on-target term in the KL decomposition is effectively fixed.","The geometry of softmax layers is shown to be dually flat, implying that interpolation between representations has two natural paths with distinct semantics — an AND-like primal interpolation and an OR-like dual interpolation."],"fun_headline_variants":["Dual-space steering proves optimal for softmax concept control","Steer with dual coordinates: minimal off-target drift, proven","Information geometry: the right way to steer softmax","Move dual coordinates to hit concept targets exactly"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The theorem assumes the linear probe exactly satisfies P(W=1|λ)=σ(βWᵀλ+bW) for every λ on the steering hyperplane (so the concept probability is constant there) and that the full output distribution factorizes into a concept part and an off-target part; the paper itself demonstrates in Section C.2 that this constancy fails on real models.","fun_headline_variants_meta":{"raw":{"variants":["Dual-space steering proves optimal for softmax concept control","Steer with dual coordinates: minimal off-target drift, proven","Information geometry: the right way to steer softmax","Move dual coordinates to hit concept targets exactly"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000241,"raw_usage":{"total_tokens":1336,"prompt_tokens":703,"completion_tokens":633,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":447,"completion_tokens_details":{"reasoning_tokens":569}},"tokens_in":447,"tokens_out":633,"duration_ms":5993,"temperature":1.0,"reasoning_tokens":569,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T22:53:58.635790+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure, along an actual dual-steering path on a large language model, the concept logit logit P(W=1|λt) versus the probe projection βWᵀλt/||βW||, and compare to test contexts with the same projection. If dual-steering paths systematically give lower concept logits at matched projections than test contexts (as the paper's Figure 7 shows), then the probe-constancy assumption is violated and the proven off-target optimality no longer holds — this observation would settle whether the theorem's premise is met.","supporting_citations":[],"review_version":1}