{"id":"c4756496-679d-462b-aa27-5502823d986f","arxiv_id":"2506.08543","paper_version":4,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"The paper argues and partially tests that linear concept representations in LLMs originate in the input space and propagate through the leading singular directions of activation difference matrices.","lead":"This paper proposes a new explanation for how large language models keep concept directions stable: the direction is already in the input and is preserved through the strongest spectral direction at each layer. It introduces a formal framework with stability bounds and reports measurements on several open-weights models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 4.5's Step 3(a) is not scale-invariant: ‖D_fluct‖_op ≤ σ1√η holds only under an unstated σ1 ≥ √M normalization, and a valid counterexample satisfies Conditions 1–2 while violating the stated bound, so the LRH derivation lacks its premise.","rationale":"The reader's verdict of REJECT is correct, but I locate the decisive problem one step earlier than their stated weakest assumption. ISLH is indeed untested and load-bearing for the origin story, but even granting ISLH, the formal route from ISLH to LRH fails because Theorem 4.5's perturbation bound is not scale-invariant and admits a concrete counterexample satisfying Conditions 1 and 2. The proof's Step 3(a) confuses the average outer-product norm η with the spectral norm of the M×d fluctuation matrix; in the ideal rank-1 case the former is (1/M) times the latter squared, so the theorem needs σ1≥√M, an unstated normalization. Since Theorem 4.8 depends on the per-layer stability bound to argue that all non-principal paths decay exponentially, the derivation of LRH does not go through. The empirical sections measure spectral gaps and incoherence ratios but never directly measure the quantity that enters the Wedin bound, so the gap is not detected. A revision can fix this by stating the correct normalization, replacing σ1√η with √(Mη) (or an equivalent bound on ‖D_fluct‖), and re-checking whether the empirical ratios still imply a useful bound. Meanwhile the paper's SPP diagnostic and measurements remain potentially useful, but the central theoretical claim is unsupported; I agree with the existing REJECT verdict and would not soften it.","tokens_in":16936,"tokens_out":20668,"duration_ms":243788,"concrete_test":"Implement the counterexample in the proof's own setting: M=100, d=4096, v1 arbitrary unit vector, d_k=0.1 c_k v1 with c_k=±1, E_k iid Gaussian entries with variance 0.01/d, ¯Δ=0, ξ=0. Compute D^{(l)} (rows d_k^T), F (rows (E_k d_k)^T), Γ=(1/M)FF^T, η=‖Γ‖_op, and D^{(l+1)}=D^{(l)}+F. Verify that σ1(D^{(l)})=1, γ(D^{(l+1)})≈1, η≈2.4e-8, so Conditions 1–2 are satisfied, but ‖F‖_2≈1.6e-3 exceeds σ1√η≈1.6e-4 by a factor of 10. Then compute sin∠(v1(D^{(l+1)}), v1(D^{(l)})) and compare to the Theorem 4.5 RHS; the RHS will be about 10× too small. This isolates the missing σ1≥√M normalization with no rank-1 error.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorem 4.5 is the load-bearing result: Theorem 4.8 invokes it to set Θ_1≈1 and let non-principal paths decay. The critical step is Step 3(a), which bounds the perturbation matrix D_fluct (rows (E_k d_k)^T) by σ_1^{(l)}√η^{(l)}. But when the rows of D_fluct are exactly the f_k used to define Γ^{(l)} (the rank-1, zero-remainder case d_k=c_k v1, ¯Δ=0, ξ=0), ‖D_fluct‖_op^2 = λ_max(F^T F) = Mη, so the claimed inequality is equivalent to σ_1 ≥ √M. This normalization is nowhere stated or verified, and the inequality is not scale-invariant: scaling all hidden states by a scales ‖D_fluct‖ by a but σ_1√η by a^2, so it fails for small a. Concretely, with M=100, d=4096, d_k=0.1 c_k v1, E_k iid N(0,0.01/d), ¯Δ=0, ξ=0, Conditions 1–2 hold (σ_1≈1, γ≈1, η≈2.4e-8), yet ‖D_fluct‖≈1.6e-3 while σ_1√η≈1.6e-4; the true angular deviation of the leading singular vector is ≈1.6e-3, an order of magnitude above the Theorem 4.5 bound. The rank-1 approximation comment in the proof does not resolve this scale mismatch. Thus the SPP stability theorem is false as stated, and the proof of Theorem 4.8 collapses at its first invocation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes the Input-Space Linearity Hypothesis (ISLH), which posits that concept-aligned directions exist in the input embedding space and are propagated through transformer layers along Spectral Principal Paths (SPP). The main theoretical claim is Theorem 4.5, a stability bound on the angular drift of the leading singular vector of a difference matrix between adjacent layers, and Theorem 4.8, which derives the Linear Representation Hypothesis (LRH) from ISLH via SPP. The paper also reports empirical measurements of spectral gaps and angular stability on four LLMs, plus a context-homogeneity ablation. The central derivation, however, rests on a stability theorem that is false as stated, and the LRH derivation is not a rigorous consequence of the stated assumptions.","tokens_in":17262,"tokens_out":9683,"duration_ms":107324,"significance":"If the theorems were correct, the paper would provide a conditional, testable mechanism for how linear concept directions persist across transformer layers, complementing the largely descriptive LRH literature. The authors deserve credit for attempting to make each assumption empirically measurable and for including a causal-style ablation on context coherence. Nevertheless, because the main stability theorem is false as stated and the LRH derivation contains substantial unproven steps, the theoretical contribution is not currently sound. The paper's conditional framing (ISLH as a hypothesis) is honest, but the promised 'rigorous stability guarantees' are not delivered.","major_comments":[{"comment":"The bound on the fluctuation term is not a consequence of Conditions 1–2. In the proof of Theorem 4.5, Step 3(a) claims ‖D^(l)_fluct‖_op ≤ σ_1^(l) √η^(l), citing Condition 2 and the rank-1 approximation d_k ≈ c_k v_1. When the rank-1 approximation is exact and the mean Jacobian perturbation and remainder vanish (d_k = c_k v_1, \\barΔ=0, ξ=0), the rows of D^(l)_fluct equal the fluctuation vectors f_k, so ‖D^(l)_fluct‖_op^2 = λ_max(∑_k f_k f_k^⊤) = M η^(l). The claimed inequality then reduces to √M ≤ σ_1^(l), which is not scale-invariant and is false in explicit examples: take M=100, d=4096, d_k = 0.1 c_k v_1, E_k iid N(0, 0.01/d I); then σ_1 ≈ 1, η ≈ 2.4×10^(-8), ‖D_fluct‖ ≈ 1.6×10^(-3), but σ_1√η ≈ 1.6×10^(-4). Thus the bound in Eq. (9) is false as stated, and Theorem 4.8, which invokes Theorem 4.5, loses its premise.","section":"Theorem 4.8 and its proof in §4.5"},{"comment":"The proof of Theorem 4.8 does not constitute a valid derivation of LRH. Eq. (12) treats the singular value decomposition of the product of Jacobians as if it factorizes into products of individual singular values and inter-layer alignment factors; this identity does not hold for non-commuting matrices, and no justification is given. Furthermore, even if the principal direction u_1^(L-1) were stable, the fact that the average or aggregate difference Δh^(L) is proportional to u_1^(L-1) does not imply that every individual counterfactual difference h_+,(L)_k − h_−,(L)_k lies in Cone(u_1^(L-1)), which is what the LRH definition in Eq. (1) requires. The proof's '≈' steps are unquantified, so the claimed formal derivation is at best heuristic.","section":"Lemma 4.2, Eq. (4)"},{"comment":"The proof of Lemma 4.2 treats the softmax attention weights A_{N,i} as constants when computing E[h|c], but these weights depend on the input tokens and hence on the concept label c. As a result, the equality in Eq. (4) does not follow from the ISLH model as written; it requires an additional assumption (e.g., that attention weights are concept-independent) or a carefully stated first-order approximation. Since \\tilde v_1 is the seed direction for the SPP, this gap affects the foundation of the framework.","section":"Lemma 4.2, Eq. (4)"}],"minor_comments":[{"comment":"The proof of Theorem 4.5 in Step 2 says 'By Theorem 4.5, this direction deviates from v_1^(l) by at most arcsin(δ^(l))'; this is circular and should instead read 'by the definition of δ^(l)'.","section":"Appendix A.2, Step 2"},{"comment":"The justification 'as τ^2 ≤ B^2 d by Jensen' in the variance bound is inaccurate; the inequality follows directly from the almost-sure norm bound ∥f_k∥ ≤ B, not from Jensen's inequality.","section":"Appendix A.3.2, Proposition A.4"},{"comment":"The empirical section does not compute the right-hand side of Eq. (9) or compare the measured spectral gap and incoherence to the bound's thresholds; the reported destructive-interference ratio of 2.7–2.9× is not directly connected to the Wedin bound's requirements.","section":"Section 5.2"},{"comment":"The experimental description is confusing: COCO is an image dataset, yet the concepts studied (honesty, fairness) and the token-wise scoring in Appendix B suggest text or multimodal prompts; the dataset construction should be clarified.","section":"Section 5 / Appendix B"}],"recommendation":"reject","confidential_remarks":"The paper's main stability theorem is false as stated, and the derivation of LRH in Theorem 4.8 is a sequence of unvalidated approximations. These are load-bearing errors that invalidate the central theoretical contribution. The empirical observations, while suggestive, do not compensate for the unsound theory. The topic is of interest, and a revised conditional empirical study might be viable, but that would be a substantially new manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know about arXiv:2506.08543. First, the central stability theorem (Theorem 4.5) is false as stated; Step 3(a) bounds ||D_fluct||_op by sigma1*sqrt(eta) using Condition 2 and a rank-1 approximation, but the bound is not scale-invariant and a valid counterexample satisfies Conditions 1–2 while violating it. Second, the paper still has a useful idea worth engaging: the SPP framework and its spectral diagnostics are new, and the empirical measurements of spectral gap and incoherence are plausible descriptive tools.\n\nWhat is genuinely new: the Input-Space Linearity Hypothesis (ISLH), the definition of the Spectral Principal Path, and the attempt to connect input-space structure to LRH via Wedin's theorem. I have not seen this exact framing in the LRH/RepE literature. The writing is clear, and the limitations section is notably honest, acknowledging the causal ablation's confound and the lack of tight worst-case bounds.\n\nThe soft spots are serious. The proof connects Condition 2 (which bounds the second moment of f_k) to the perturbation term D_fluct (whose rows are E_k d_k) via an unquantified rank-1 approximation. The stress-test counterexample is concrete: with M=100, d=4096, d_k=0.1 c_k v1, E_k ~ N(0,0.01/d), Conditions 1–2 hold, yet ||D_fluct|| is roughly an order of magnitude larger than sigma1*sqrt(eta), so the claimed bound fails. The central derivation of LRH in Theorem 4.8 therefore lacks its premise. The paper also does not directly test ISLH; Eq. (3) is assumed, not estimated. Empirically, results are single runs without code or data, and the causal ablation is further confounded by prompt-length changes, as the authors admit. These are not minor; they hollow out the central claim.\n\nWho gets value: someone working on spectral methods for interpretability might mine the measurement protocol and the framing, but they should not rely on the theorem. I would not cite the paper until the proof is repaired and the code/data are released. Still, it deserves a serious referee: the idea is important enough that a flawed first attempt should get expert feedback, not a desk reject.","headline":"A genuinely new spectral framing of linear representations, but the main stability theorem is false as stated and the empirical support is thin.","tokens_in":17849,"tokens_out":3490,"would_cite":false,"duration_ms":37754,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that the linear concept directions observed in LLM hidden states are inherited from a concept-aligned direction already present in token embeddings, then preserved across layers by flowing along a spectral principal path…","keywords":["linear representation hypothesis","spectral principal path","singular value decomposition","interpretability","large language models","input-space linearity","context incoherence","Wedin sinTheta theorem"],"falsifier":"Estimate the input-space concept direction $v_{\\mathrm{in}}$ directly from token embeddings before any transformer layer, then measure its alignment with the SPP direction $v_1^{(l)}$ at several intermediate layers; if the alignment is low, or if a model trained on embeddings with randomized concept directions still develops stable linear concept representations at depth, the claim that input-space linearity is necessary for the Linear Representation Hypothesis is contradicted.","tokens_in":16606,"feed_emoji":"📐","tokens_out":5557,"duration_ms":62009,"temperature":0.7,"pith_summary":"The paper claims that the linear representations observed in LLM hidden states are not created by the network but inherited: concept-aligned directions are already present in token embeddings, and the transformer stack preserves them along a spectral principal path. It formalizes this as the Input-Space Linearity Hypothesis and proves, via Wedin's sinΘ theorem, that the leading singular direction of a concept-difference matrix stays nearly fixed across layers when a spectral gap and context incoherence hold. From this it derives the Linear Representation Hypothesis: the deep representation difference is proportional to the product of leading singular values times one dominant direction, so concepts become linearly separable. The authors test the prerequisite conditions on four decoder-only LLMs and show that breaking context incoherence degrades the spectral gap and cross-layer stability, as predicted.","feed_headline":"LLMs inherit linear concept directions from token embeddings","feed_subtitle":"A new stability bound explains why concept directions stay linear and stable across deep layers.","key_machinery":"The central object is the Spectral Principal Path (SPP): given $M$ paired sequences differing only in concept tokens, form the difference matrix $D^{(l)}$ whose rows are last-token hidden-state differences for each pair; the SPP is the leading right singular vector $v_1^{(l)}$ of $D^{(l)}$. The argument is carried by the Wedin $\\sin\\Theta$ perturbation theorem, which bounds the rotation of this leading singular vector when $D^{(l+1)}$ is decomposed into a mean-Jacobian-propagated signal $D^{(l)}\\bar{J}^{(l)\\top}$ plus context fluctuation and a higher-order remainder. The stability theorem requires a spectral gap ($\\sigma_1^{(l)} \\gg \\sigma_2^{(l)}$) and context incoherence (the empirical second-moment matrix of per-sample fluctuations has small operator norm $\\eta^{(l)}$). Together these make the network behave like a multi-stage power iteration, so non-principal spectral paths decay exponentially and the concept signal converges to a single dominant direction.","core_discovery":"The central claim is that the Linear Representation Hypothesis is not an emergent learning outcome but a spectral inheritance: under the Input-Space Linearity Hypothesis, each token's embedding is a concept-agnostic base plus a concept-aligned term $c\\cdot \\alpha_i v_{\\mathrm{in}}$ plus isotropic noise, so a shared concept direction $v_{\\mathrm{in}}$ exists in the input space. The paper proves that the first attention layer aggregates this distributed signal into a concentrated direction at the last token, and then proves that the leading right singular vector of the difference matrix $D^{(l)}$ rotates by at most $\\sigma_1^{(l)}(\\delta^{(l)}+\\sqrt{\\eta^{(l)}}+\\sqrt{\\zeta^{(l)}})/\\gamma^{(l+1)}$ between layers, where $\\gamma$ is the spectral gap and $\\eta, \\zeta$ measure context fluctuation and higher-order remainder. Theorem 4.8 then concludes that the deep representation satisfies the Linear Representation Hypothesis: $\\Delta h^{(L)}$ is proportional to the product of the leading singular values times the leading left singular vector $u_1^{(L-1)}$, so the semantic difference collapses into a one-dimensional subspace. In short, the paper claims that concept directions in LLMs are inherited from input-space linear structure and propagated through dominant spectral directions, with stability guaranteed by spectral gap and context incoherence.","pith_inferences":["A testable extension: treating the SPP direction as a causal handle, ablating or amplifying the leading singular direction of $D^{(l)}$ at an intermediate layer should shift downstream concept behavior, and the theorem's bound gives a quantitative prediction for how much drift to expect.","The framework implies that concepts not linearly organized in the input embedding space, such as compositional or culturally situated concepts, may fail to form stable spectral principal paths, which would explain why some concepts are more linearly probeable than others.","The same spectral argument may transfer beyond decoder-only LLMs to any residual architecture with attention-based aggregation, since only residual Jacobians and attention pooling are used; the paper hints at multimodal extensions but does not prove the transfer.","One could test the theory by artificially shrinking the spectral gap, for instance by adding a strong second concept direction, and predicting the precise layer at which the SPP stops tracking the original concept."],"forward_implications":["If the Input-Space Linearity Hypothesis holds, linear probes and steering vectors work because the network preserves input-space directions rather than inventing them, so interventions at the embedding level should transfer predictably to deep-layer behavior.","The per-layer bound implies a worst-case cumulative angular drift that grows linearly in depth, but a random-walk estimate that grows only as $\\sqrt{L}$; the paper argues that the empirically observed stability is consistent with the random-walk regime.","Both prerequisite conditions are measurable per layer (the spectral-gap ratio $\\sigma_1/\\sigma_2$ and the operator norm of the fluctuation second-moment matrix), giving practitioners a concrete way to predict whether a given model and concept should exhibit stable linear representations.","Because the deep representation difference becomes proportional to the product of leading singular values times $u_1^{(L-1)}$, concept directions at depth are predicted to remain aligned with the SPP across layers and to scale with the cumulative spectral gain."],"supporting_citations":[{"why":"Supplies the sinTheta perturbation inequality that bounds the rotation of the leading singular vector under perturbation, which is the core of Theorem 4.5.","marker":"Wedin, 1972"},{"why":"Defines the Linear Representation Hypothesis that the paper derives from the Input-Space Linearity Hypothesis via the SPP.","marker":"Park et al., 2023"},{"why":"Provides the RepE and LAT methodology used to validate traceability of concept directions in representations.","marker":"Zou et al., 2023"},{"why":"Supplies the matrix Bernstein inequality used to prove the context-incoherence bound in Proposition A.4.","marker":"Tropp, 2015"},{"why":"Provides the COCO dataset used to construct the concept-pair sequences for the empirical validation.","marker":"Lin et al., 2014"}],"fun_headline_variants":["Spectral proof: LLM concepts are inherited, not emergent","New bound explains stable concept directions in LLMs","Concept lines in LLMs: a spectral inheritance story","Why LLM concept directions stay stable: spectral paths"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is the Input-Space Linearity Hypothesis: every token embedding is exactly a concept-agnostic base plus a per-token strength times a single shared concept direction, plus Gaussian noise; if input embeddings are not linearly organized this way, the SPP mechanism and the derivation of the Linear Representation Hypothesis do not get off the ground.","fun_headline_variants_meta":{"raw":{"variants":["Spectral proof: LLM concepts are inherited, not emergent","New bound explains stable concept directions in LLMs","Concept lines in LLMs: a spectral inheritance story","Why LLM concept directions stay stable: spectral paths"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000407,"raw_usage":{"total_tokens":2147,"prompt_tokens":1010,"completion_tokens":1137,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":626,"completion_tokens_details":{"reasoning_tokens":1073}},"tokens_in":626,"tokens_out":1137,"duration_ms":11288,"temperature":1.0,"reasoning_tokens":1073,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:08:26.743166+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Estimate the input-space concept direction $v_{\\mathrm{in}}$ directly from token embeddings before any transformer layer, then measure its alignment with the SPP direction $v_1^{(l)}$ at several intermediate layers; if the alignment is low, or if a model trained on embeddings with randomized concept directions still develops stable linear concept representations at depth, the claim that input-space linearity is necessary for the Linear Representation Hypothesis is contradicted.","supporting_citations":[{"cited_title":"19 Under review as a conference paper at COLM 2026 Figure 6: VLM response for power concept with U.S","cited_arxiv_id":null,"evidence_quote":"Provides the RepE and LAT methodology used to validate traceability of concept directions in representations."},{"cited_title":"Microsoft coco: Common objects in context","cited_arxiv_id":null,"evidence_quote":"Provides the COCO dataset used to construct the concept-pair sequences for the empirical validation."}],"review_version":1}