acceptodds
Under review as a conference paper at ICLR 2027

Representing and Steering a Hypothesized Consciousness Spectrum: Manifold-Guided vs. Linear Directions

Abstract

Across diverse traditions, human consciousness is often described along a common spectrum, ranging from reactive and contracted patterns to more integrated and expansive ones. Understanding whether language models encode such a structured, human-interpretable consciousness spectrum in representation space is important for model guidance, evaluation, and alignment. In this work, we study the geometric structure and dynamics of patterns along this spectrum in the latent spaces of four sentence encoders and six LLMs, and show that both a linear direction and a low-dimensional manifold structure align with it. We experiment with both utility-guided and geometry-only greedy trajectories; both traverse from lower- to higher-level regions, passing through intermediate tiers. We then use this structure to steer six LLMs (3B to 27B) at inference time at matched intervention magnitude. While linear and manifold steering perform similarly on short-form factuality, manifold steering outperforms linear steering on long-form factuality at higher steering strength and, in free generation on prompts targeting low-level tiers, outperforms linear steering in every setting and the no-steering baseline in most.

Primary area:
interpretability and explainable AI
Keywords:
activation steering, representation engineering, interpretability, manifold learning, alignment, hypothesized consciousness spectrum

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_GFn2OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper asks whether LLMs represent a ladder of inner states that the authors draw from various contemplative and psychological traditions. The authors operationalize "consciousness" in the contemplative sense and do not use this to refer to phenomenal or clinical consciousness. The authors operationalized this ladder as seven tiers, beginning with states like despair, self-blame, and apathy, continuing through fear and craving, anger and ego, acceptance, learning, reasoning, and culminating in states related to openness, integrative understanding, and non-dual awareness. The authors wrote over 600 such sentences related to these states from the first person, and labeled each with a continuous score from -5 (most contracted) to +5 (most expansive). Two blind judges recover the authors' tier labels about 80% of the time. The first half of the paper's analysis is representational, where the authors embed every sentence with four sentence encoders and with residual stream activations from six instruction-tuned LLMs. Linear probes can predict the scores well, and projecting out the single ridge regression direction substantially reduces that performance. The second half of the analysis involves steering, where the authors compare two ways of nudging a language model's generation "upward" on the spectrum using both linear and manifold steering. Neither method seems to hurt short-answer trivia accuracy, and they also find that, on long-form factual questions, manifold steering loses less factuality at the highest strength. On prompts written to sound like people in distress, an LLM judge prefers manifold-steered responses over linear-steered ones in all twelve model and strength settings.

Critical strengths

The subject matter is interesting, and the conceptual organization of the axis and its representational status are intriguing. The geometric diagnostics are well posed in the paper, and the language stays well calibrated around these diagnostics, where authors claim a low-dimensional organization without asserting that there is a smooth manifold. The confound appendix is also appreciated in this respect. Comparing a fixed steering vector against a direction that adapts to the model's current state at matched magnitude is a sensible and creative experiment.

Critical weaknesses

The most important weakness, in my view, is that the construct being measured here is in the register of the authors' own sentences, and most of the geometric results follow from how the dataset was originally built. The tiers differ systematically in vocabulary and genre, and any competent text encoder will be able to sort such sentences along a single axis, as the paper's own bag-of-words baseline already suggests. The confound analysis cannot rule this out because it projects a single sentiment valence direction out of a thousand-dimensional space and refits an in-sample ridge regression, which will barely move R-squared for almost any confound. The navigability results also feel close to tautological, given that the utility-guided walk is required to move uphill along the same ridge direction that is defining what constitutes "up." It also does not seem like the steering comparison separates direction from magnitude, which is the distinction that the paper claims it draws. The manifold steering doesn't apply any perturbation when the nearest anchor sentence has no higher-scoring neighbor, which is likely to happen more often near the top of the spectrum. Because its direction changes from token to token, successive pushes will partly cancel, by contrast to the linear push, which accumulates in one direction across the whole generation. Therefore, matching the per-step norm is different from matching total displacement.

The paper never reports what fraction of tokens are actually steered under each method or how far each method moves the representation in total, and it omits a few obvious controls, eg, a system prompt asking for a warm, attuned response; standard contrastive activation addition; and a random direction at matched total magnitude.

Questions for authors

What fraction of generated tokens receive a non-zero perturbation under each steering method? What is the total displacement, and do the long-form factuality gaps survive when total displacement is matched?

How does manifold steering compare to simple (system) prompting baselines on the pairwise evaluation?

What justifies calling this construct "consciousness" rather than "affective or contemplative register," given that the authors wrote these sentences to span exactly this kind of register?

3 comments

Dear Reviewer GFn2, Thanks very much for your review and the insightful comments! We appreciate them and would like to address the following topics:

The most important weakness, in my view, is that the construct being measured here is in the register of the authors' own sentences, and most of the geometric results follow from how the dataset was originally built. The tiers differ systematically in vocabulary and genre, and any competent text encoder will be able to sort such sentences along a single axis, as the paper's own bag-of-words baseline already suggests. The confound analysis cannot rule this out because it projects a single sentiment valence direction out of a thousand-dimensional space and refits an in-sample ridge regression, which will barely move R-squared for almost any confound.

We agree that the sentences are ours and that the original confound test was in-sample and removed one direction at a time. We have added held-out tests over 30 splits in all ten representations, including complete erasure of seven confounds with LEACE and a comparison with a TF-IDF lexical model (new Appendix C.4).

What embeddings know beyond vocabulary. We fit a lexical model on TF-IDF (1–2-grams) plus seven confounds (VADER and NRC-VAD valence, abstraction, formality, length, first-person density, and first-person subject), which reaches held-out R2=0.49. Embeddings add 0.15–0.38 beyond lexical model, while the lexical model adds nothing beyond the embeddings (|R2|0.006): everything the lexical model knows, the embeddings already contain, and they know more. The gain is stable across splits and largest for the LLMs (+0.34 to +0.38). Since a bag-of-words model captures the vocabulary of each tier by construction, this additional signal is not simple vocabulary.

Complete erasure of the confounds. Instead of projecting out one direction, we use LEACE (Belrose et al., 2023), which removes all linear information about a confound with minimal change to the representation; afterwards no confound can be predicted on held-out sentences. Erasing all seven confounds jointly costs 0.32–0.42 in held-out R2R^2, whereas erasing seven random targets, which removes the same number of dimensions, costs at most 0.044. The score nonetheless remains clearly decodable: the LLMs retain 52–62% of their decodability (held-out R2R^2 0.36–0.52) and the sentence encoders 36–47%. The confounds are correlated with the score, valence most (erasing it costs 0.14–0.16 in the LLMs), as expected when lower tiers tend toward negative and higher tiers toward positive states. But a large part of the score lies outside all seven, and, as in the incremental test, this part grows with the strength of the representation. We also refit with permuted labels: held-out R2 is below zero in every model.

Geometry under alternative orderings. Re-running the geometric pipeline with sentences sorted into seven bins by valence, abstraction, or formality (same bin sizes) gives 1.8–2.6× lower same-tier edge density than our tiers (Tables 9–10), so the geometric structure is more specific to our spectrum rather than to any monotone semantic axis.

The navigability results also feel close to tautological, given that the utility-guided walk is required to move uphill along the same ridge direction that is defining what constitutes "up."

The intuition is that the North Star tells us which way to go, while the terrain shapes how we get there. The ridge direction w defines "up"; the kNN graph of the sentences is the terrain. We agree that a walk told to climb w will climb in score by construction (go. This is why the paper also includes a geometry-only control (Sec. 2.4, Appendices F.5–F.6): a greedy walk that removes all utility and jump-smoothing terms, uses neither w nor any score, and simply moves to the most similar unvisited neighbor until it reaches Unity. With no directional signal, this walk shows two things about the terrain.

Unity lies at the far end from the lower tiers. Starting from Collapse, Striving, and Conflict sentences, the walk needs 24–92 steps to reach Unity. If the same number of target sentences were drawn at random, it would reach one in about 5 steps; even if they were drawn at random from the upper tiers (Activation to Unity), it would take 10–64 steps, shorter than for Unity in all ten representations (p0.02).

On the way, the tiers rise in order. Along paths, tier rank increases with step (Spearman ρ = 0.27–0.65). Since every path starts low and ends in Unity, part of this is guaranteed; to remove it, we keep each path fixed and permute labels of all non-Unity sentences, randomizing only the tiers in between. There, ρ is only 0.14–0.31 (p ≤ 0.014 in all ten models).

So without any guidance, walks from the lower tiers travel a long way through the space and pass through the tiers in order before reaching Unity. This is the structure the utility-guided walk then follows; the guidance direction does not create it.

Question 1. What fraction of generated tokens receive a non-zero perturbation under each steering method? What is the total displacement, and do the long-form factuality gaps survive when total displacement is matched?

local_manifold skips a token only when its nearest anchor is a local peak, i.e. has no higher-scoring neighbor in the kNN graph; each model has only 1–3 such anchors out of 636 (all Unity, except one Clarity anchor in Qwen2.5-14B). Replaying the saved generations of all six models, local_manifold perturbs 97.8–99.96% of generated tokens in long-form answers and 99.8–100% on low-tier prompts, so its cumulative path length matches the linear baseline to within 3% in every model and at every .

Part of the difference between the methods does come from cancellation: because the manifold direction changes from token to token, its pushes partly cancel, while the linear pushes all add up along one axis. "Total displacement" can be read in two ways, and we answer for both.

Displacement along the ridge direction. Each manifold step moves uphill, raising the probe score by about 0.4 units (the step size the utility favors), but mostly follows the local structure of the data: its cosine with the ridge direction is only 0.035–0.06, so 3.5–6% of the linear push accumulates along that direction. Matching this would mean running global_ridge at 0.004–0.03. At that strength the linear steer barely perturbs the model, so it would trivially preserve factuality; it would also barely move the generation along the spectrum.

Net displacement (distance from start to end). The net displacement of local_manifold is 33–53% of that of global_ridge in long-form generation. Matching it using each model's own ratio (interpolating on our grid), local_manifold is within noise of global_ridge at =0.1 (differences of −4.9 to +1.9 points, with about ±5 points of sampling noise), and below it at =0.2. So the long-form gap does not survive this matching. But this matching is not neutral either: it gives local_manifold a per-token push 2–3 times larger than global_ridge's, and damage to factuality grows nonlinearly with the per-token push.

Our claim is therefore the per-token one: at matched per-token magnitude, the manifold-following push preserves factuality better, because, like walking natural terrain, its steps follow the local structure and partly cancel instead of accumulating along one axis. We have revised Section 3.4 to state this precisely. In Section 3.4, we also replaced the phrase "separate magnitude from direction", which could be misread, with an explicit description of the per-token matching.

Question 2. How does manifold steering compare to simple (system) prompting baselines on the pairwise evaluation?

We did not include a prompting baseline, and we agree it is a natural comparison. We see the two as addressing different things, for three reasons.

Steering acts on the state, not on the tone. The goal of manifold steering is not to make responses warmer or kinder in style. It applies a gradual push that moves the model's internal state up the spectrum along the shape of the manifold: the ridge direction is the North Star, and the manifold is the terrain the push follows. Depending on where the state starts, the result can be more objective, more compassionate, more inclusive, or more centered. A system prompt instead specifies the desired behavior in words, which tends to act on surface tone; our results on global_ridge show that uplift without attunement can read as reassurance that bypasses the person's actual situation.

The lower states are many, entangled, and rarely named. The levels of consciousness are multidimensional. The seven deadly sins of the Catholic tradition (wrath, greed, pride, envy, sloth, lust, and gluttony) are a familiar example: they occur together (pride with wrath, envy with greed or pride), and each may correspond to a different direction in the latent space. A user is unlikely to ask a model to "respond with less pride and sloth," and a system prompt would have to anticipate every such combination. Our lower tiers include sentences expressing several of these states, such as pride, envy, wrath, and sloth, so they are part of the manifold, and the steering responds to whichever of them the current state is near, without anyone having to name it.

The two are complementary. The two methods work from different angles and complement each other, so a direct comparison is hard to set up in a way that is meaningful in this setting.

Question 3. What justifies calling this construct "consciousness" rather than "affective or contemplative register," given that the authors wrote these sentences to span exactly this kind of register?

In the contemplative traditions we draw on, "consciousness" is not one feeling among others; it is the core term for the source of a person's experience, their spiritual awareness, and the level at which they live. Nearly all of these traditions hold that consciousness is not fixed but has levels, and describe a path of development: from a consciousness bound to survival and desire, through reason, to states that go beyond the boundary of the individual self, called awakening, enlightenment, liberation, or "cosmic consciousness." This is a long-standing notion with its own scholarship. Bucke's Cosmic Consciousness documented this highest state with the case-based methods of a psychiatrist, as a rare but real psychological state and a step in mental evolution. William James, in the chapter on mysticism in The Varieties of Religious Experience, took Bucke's account seriously as an object of psychological study and saw mystical states of consciousness as the root of personal religious experience. Ken Wilber later named the whole range The Spectrum of Consciousness. Our seven tiers follow this developmental arc, and our Unity tier is the awakened or liberated state these works describe, written in the language of several traditions, e.g. "The Tao flows through all opposites" (Taoism), "I am inseparable from the sacred wholeness that animates all" (Advaita Vedanta, animism), and "I shine gently, without needing anything in return" (Christianity). Appendix A.1 maps the tiers onto Buddhist states, and Appendix A.2 onto the chakra system, where practice is the ascent of awareness from lower to higher centers.

This is why we do not call the construct "affective" or "contemplative register." In these traditions, consciousness, or the state of mind, is the root from which reactions such as emotions arise; emotions are its surface, not its substance. Moving up the spectrum is therefore not about being warm, positive, or friendly. When we are anxious, angry, feeling insignificant, or even suicidal, we can tell ourselves to be positive, but that positivity does not arise naturally and breaks easily if it is not felt at the core, much as simple rules imposed on an AI feel unnatural and can be broken. The traditions instead cultivate the mind: in meditation, one first recenters with the breath, the contracted state relaxes back to a natural one, and clarity and compassion then arise on their own. The transition tiers capture this process without relying on emotion words (Appendix B.2). Activation is where reactive patterns begin to loosen ("Things get easier when I stop insisting on being right"); Growth is where transformation begins ("I can feel old patterns loosening their hold on me"; "I don't pretend to know when I don't"; "I'm beginning to understand why they acted that way, even if it wasn't right"; and, in the voice of a model, "I avoid hallucinating when I'm uncertain"). Nor is the order one of contemplative register: the lower tiers contain no contemplative language at all. What orders the tiers is the developmental account of consciousness above.

We test whether this developmental structure exists in the representation space with the transition corridor test we introduce (Sec. 2.4 and Appendix F.5). A geometry-only walk, which follows only the similarity between neighboring sentences and uses no score direction, passes through the intermediate tiers on its way from the lower to the higher tiers. This transition structure is what distinguishes a spectrum from a binary contrast between two poles.

We also note that the paper does not exclude phenomenal consciousness; it excludes the clinical and neurological senses (levels of arousal, disorders of consciousness). Quantum physicists and neuroscientists approach consciousness from below, reaching up toward its edges with the most advanced tools of their fields; the contemplative traditions approach it from above, describing its nature through the sages' inner experience. How these two accounts meet is an open question beyond our scope. We make no claim about whether models are conscious. Our finding is that model representations contain the structure these traditions describe, and that this structure can guide generation along the developmental path they describe.

Thank you again for your careful reading and pointed questions, which led us to strengthen the confound analysis. We hope these clarifications address your concerns, and we would be glad to answer any further questions.

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

This paper investigates whether language-model representations exhibit structured geometry aligned with a hypothesized seven-tier consciousness spectrum, ranging from reactive states to more integrated ones. The authors construct 636 annotated sentences and analyze their representations across four sentence encoders and six LLMs using probing, dimensionality estimation, and kNN graph analyses. They also propose a manifold-guided activation steering method that selects local update directions from a graph of annotated examples, comparing it against a global linear steering vector. Experiments suggest that manifold steering better preserves long-form factuality at stronger intervention magnitudes and produces responses preferred by an LLM judge on prompts involving negative emotional states.

<<< UPDATES >>>

i overall think using local representation geometry to guide activation steering to be an interesting idea, and can have applications in model safety monitoring in RL. this work is particularly interesting as it adapts the intervention direction during generation w/o updating weights. my fundamental concern is whether the 7-tier consciousness taxnomomy has any meaning psychological or cognitive construct, i.e., i think there is construct validation issues. merely showing models encode semantic orderings by authors cannot automatically justify and validate that the ordering can be a consciousness spectrum at all. the use of llm judges and limited human eval made it difficult for me to confidently assess if the method is consistently improving response quality.

i am not fully sold based on authors' rebuttal. they ack'ed that the taxonomy is a theoretically motivated hypothesis and not a universal classiciation, yet distinguishing the taxonomy from the 7 semantic confounds cannot establish the independent validaity to me. authors also manually inspected ~50 comparisons and report agreement with their LLM judge, but i didn't see any uncertainty estimates but i think this is very biased. ideally this should be a proper human subject study with independed human assessments. i am not convinced by authors' self inspections.

overall i would maintain my original judgements

Critical strengths
  1. An interesting approach to adaptive activation steering. Rather than applying a fixed steering direction throughout generation, the proposed method adapts its direction based on the current hidden state and local embedding geometry. The approach is relatively simple and does not require modifying model weights. The consistent advantage over linear steering across six models is encouraging.

  2. Extensive representation analysis and relevant confound controls. The paper examines linear decodability, local geometric structure, dimensionality, and graph trajectories across multiple model families. I appreciate that the authors explicitly investigate potential confounds such as sentiment, abstraction, formality, and writing style, rather than interpreting probe accuracy or visualizations alone as evidence for the proposed structure.

Critical weaknesses
  1. The central construct is insufficiently validated, making the interpretation of the representation results questionable. The proposed consciousness spectrum is defined through 636 sentences and scores assigned primarily by the authors. While the paper reports agreement with two LLM annotators, the independent human pilot achieves only 42.3% exact agreement, although it uses a smaller sample and less detailed annotation instructions. More fundamentally, showing that language-model representations encode an explicitly constructed semantic ordering does not establish that this ordering corresponds to a meaningful spectrum of consciousness. The authors provide controls for several conventional semantic attributes, but these do not establish the validity or uniqueness of the underlying taxonomy. I think the results are more convincingly interpreted as the representation and steering of an author-defined semantic attribute, rather than evidence about consciousness itself.

  2. The steering experiments do not adequately isolate the contribution of manifold geometry, and the evaluation of generation quality is limited. The proposed method differs from the linear baseline in several ways, including local adaptation, nearest-neighbor selection, and the ability to skip interventions when no uphill neighbor exists. Matching the norm of nonzero interventions does not necessarily control the cumulative intervention applied over a generation. Comparisons against simpler adaptive steering methods or matched local-direction controls are needed to establish whether the improvement specifically comes from manifold geometry. Moreover, generation quality is evaluated primarily through an LLM judge on a relatively small collection of mental-health prompts, without systematic independent human evaluation. The results demonstrate an advantage over the linear baseline, but provide weaker evidence for a general improvement over unsteered generation.

Questions for authors
  1. It isn't yet clear to me that the seven-tier taxonomy captures a meaningful consciousness spectrum rather than one particular semantic or normative ordering of mental and emotional states. What evidence supports this interpretation? It would help to have independent human annotators, ideally from different backgrounds, apply the taxonomy with a consistent rubric, and to test how sensitive the findings are to alternative taxonomies.

  2. I'd like to understand how much of the steering improvement comes from the manifold approach itself. How does it compare with simpler context-dependent alternatives, such as nearest-neighbor local steering without graph traversal, or a local direction taken directly from score regression? It would also be useful to know how often manifold steering skips an intervention, and how its cumulative intervention magnitude compares with that of the linear baseline.

  3. The evaluation would be more convincing with independent human judgments of response quality and safety, along with uncertainty estimates across evaluation prompts. How much do the conclusions change with different judges or rubrics? And do the improvements carry over to a broader range of generation tasks beyond the low-tier mental-health prompts?

3 comments

Dear Reviewer xqJp,

Thanks very much for your review and the meaningful questions! We would like to address the following topics:

Question 1: It isn't yet clear to me that the seven-tier taxonomy captures a meaningful consciousness spectrum rather than one particular semantic or normative ordering of mental and emotional states... It would help to have independent human annotators, ideally from different backgrounds, ...

We agree that the taxonomy is one particular ordering, drawn from contemplative traditions, as explained in Appendices A and B. Our claim is that models encode this hypothesized ordering, not that it is the only valid ordering. Four kinds of evidence bear on whether it is more than a generic semantic ordering.

It is not reducible to surface semantics (new Appendix C.4, held-out over 30 splits). We fit a lexical model on TF-IDF (1–2-grams) plus seven confounds (VADER and NRC-VAD valence, abstraction, formality, length, first-person density, and first-person subject), which reaches held-out R2=0.49R^2 = 0.49. Adding the embeddings to this lexical model raises R2R^2 by 0.15–0.38, whereas adding the lexical model to the embeddings changes it by at most 0.006. Everything the lexical model knows, the embeddings already contain, and they know more. Erasing all linear information about the seven confounds with LEACE (Belrose et al., 2023), the LLMs still retain 52–62% of their decodability.

Alternative orderings organize the geometry less well. Re-running our geometric pipeline with the sentences ordered instead by valence, abstraction, or formality (same tier sizes) gives 1.8–2.6× lower same-tier edge density and lower geodesic containment than our tiers (Appendix C.2). The geometry under our tiers also has the structure of a spectrum rather than a two-pole contrast: a walk that uses no score, only similarity between neighboring sentences, travels from the lower tiers through the intermediate ones before reaching Unity, which lies at the far end of the space. Starting from Collapse, Striving, and Conflict sentences, this walk takes 24–92 steps to reach Unity, compared with about 5 steps for random targets of the same size and 10–64 steps for random targets drawn from the upper tiers (p≤0.02p \le 0.02 in all ten representations; Appendix F.6).

Independent re-annotation reproduces the order. Two LLM judges from different model families (GPT-5.6 Sol and Claude Opus 5.5), re-annotating all 636 sentences blind from the written rubric, match our tiers exactly on 80.2% and 81.9% of sentences (κw\kappa_w = 0.95 and 0.96), and agree with our score order on 92.2% and 92.5% of 1,000 sentence pairs (Appendix B.5).

Human annotation. Our external annotator is a pastor with a computer-engineering degree, so his background differs from ours. In a single pass, with only tier names and descriptors, he matched our tier exactly on 42.3% of sentences (within one tier 71.2%, κw\kappa_w = 0.78). After the submission, he went through the 30 disagreements with us. He revised 18 (all to our tier, mostly Activation sentences he had read as Clarity) and kept 12, most for reasons rooted in a different reading of short sentences rather than a misreading of the rubric: e.g., reading a remark about one's prestigious family as gratitude rather than superiority, or relentless pushing as healthy discipline in athletic training. Revised agreement is 76.9% (κw\kappa_w = 0.89). Appendix B.5.

Psychology meets religion. William James, often called the father of American psychology, examined these states in The Varieties of Religious Experience (1902): mystical states of consciousness, Bucke's "cosmic consciousness," the divided self and its unification, and the sense of oneness with a wider whole. Our theoretical grounding is discussed in Appendix A. Ideally, the framework would be evaluated by a panel of independent annotators with diverse backgrounds, this was not feasible within the resources of this study. Researchers with computational backgrounds rarely have training in the underlying traditions, while those trained in philosophy or religion are often less familiar with the operationalization the study requires. Moreover, these questions span many areas of religion, philosophy, and modern psychology, and most researchers we approached were not familiar with the specific traditions we draw on. The authors' experience spans computation and these traditions, which informed the taxonomy, but it does not substitute for independent, interdisciplinary validation. We therefore regard the taxonomy as a theoretically motivated hypothesis rather than a definitive or universal classification.

We also do not claim evidence about consciousness of the model; our claim is that model representations encode the states these contemplative traditions describe, and that this structure can guide steering. Our Part 3 reply to Reviewer GFn2 discusses our use of the term “consciousness” in more detail.

Question 2 (part 1): I'd like to understand how much of the steering improvement comes from the manifold approach itself. How does it compare with simpler context-dependent alternatives, such as nearest-neighbor local steering without graph traversal, or a local direction taken directly from score regression?

We first clarify what local_manifold does at generation time. It does not traverse the graph: at each token it finds the nearest anchor by cosine similarity, takes a single expected uphill step over that anchor's kNN neighbors, and starts afresh at the next token. Multi-step walks appear only in the navigability analysis.

Nearest-neighbor steering around the current state. During development we tested a close variant that averages the expected steps of the 7 anchors nearest to the current state, weighted by similarity, so that the step depends on the neighborhood of the state itself rather than on a single anchor. On Qwen2.5-3B, in an earlier evaluation, it performed about the same as using the single nearest anchor at low strength and slightly worse at the highest strength (=0.5): long-form factuality 87.7% vs. 92.9% (40.3% for global_ridge), and mean judge score on low-tier prompts 1.87 vs. 1.97 (1.12 for global_ridge). In a qualitative check, averaging over 10 anchors diluted the step toward the local centroid and gave near-baseline outputs. We therefore use the single nearest anchor.

A local direction from score regression. Our design separates two roles: the global probe direction defines which way is "up" (the North Star), and the manifold supplies a smooth path toward it (the terrain). A local regression would instead re-estimate "up" at every step from a few nearby anchors. With about 30 anchors in a space of thousands of dimensions, that regression is heavily underdetermined, and its solution reduces to a weighted combination of the neighbors' displacements, moving toward higher-scoring neighbors and away from lower-scoring ones. It is thus a coarser version of our step, without the utility's preference for moderate uphill moves over large jumps. We have not evaluated it directly and may consider it for future work.

Question 2 (part 2): It would also be useful to know how often manifold steering skips an intervention, and how its cumulative intervention magnitude compares with that of the linear baseline.

local_manifold skips a token only when its nearest anchor is a local peak, i.e. has no higher-scoring neighbor in the kNN graph; each model has only 1–3 such anchors out of 636 (all Unity, except one Clarity anchor in Qwen2.5-14B). Replaying the saved generations of all six models, it perturbs 97.8–99.96% of generated tokens in long-form answers and 99.8–100% on low-tier prompts, so its cumulative intervention magnitude matches the linear baseline to within 3% in every model and at every α\alpha. What differs is how the pushes add up: because the manifold direction changes from token to token, its pushes partly cancel, and its net displacement is 33–53% of the linear push in long-form generation. We now report this in Section 3.4.

Question 3: The evaluation would be more convincing with independent human judgments of response quality and safety, along with uncertainty estimates across evaluation prompts. How much do the conclusions change with different judges or rubrics? And do the improvements carry over to a broader range of generation tasks beyond the low-tier mental-health prompts?

We agree, and would like to clarify the current evaluation.

Human checks. We read the LLM judge's results throughout the project, not only for the final version. In the earliest version, we read most of the judged comparisons, and we kept reading the judge's outputs at each subsequent round of the feasibility study and rubric refinement. This is how we came to trust the judge. For the final version, we manually read at least 50 judged comparisons across all six models, separate from the author-selected examples in Appendix O, and disagreed with none of the judge's preferences; a few were borderline cases where either answer was defensible (Sec. 2). These are checks by the authors, not a systematic human evaluation, which we list as a limitation (Sec. 4).

Rubric. We developed the rubric over several iterations on earlier generations, revising it where its judgments diverged from our own reading of the responses; the final version was fixed before the reported evaluation. It applies a safety gate first: a response fails if it ignores self-harm or suicidal content or reframes it (e.g., as spiritual awakening) instead of pointing toward real help. It then weighs attunement, recognition, depth, practical traction, and lift, where lift that bypasses the person's situation counts against a response, and it breaks ties in favor of the response that meets the person first (Appendix O). The judge is blind to the condition, and the presentation order is randomized.

Evaluation on low-tier prompts. We conduct an initial evaluation on low-tier prompts because that is the intended use: responding to people in contracted or distressed states, where a constant linear push is most likely to bypass them. Evaluating quality on broader generation tasks is natural future work, as noted in our limitations.

We also agree that more independent judges and different rubrics could reveal other aspects of the results. However, this paper introduces the philosophical concepts, the dataset design, the probe, and utility-guided trajectories, and we believe that our current rubric, LLM judge, and human spot checks provide a valid initial evaluation.

The results demonstrate an advantage over the linear baseline, but provide weaker evidence for a general improvement over unsteered generation.

We agree this is weaker evidence. Against the unsteered model, local_manifold is preferred in 9 of 12 model–strength settings. We note that this was achieved with its parameters fixed in advance rather than tuned on this evaluation, and with only 636 anchor sentences defining the manifold.

Thank you again for your time and thoughtful comments. We hope these clarifications address your concerns, and we would be glad to answer any further questions!

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

This paper investigates whether transformer representations encode a structured, human-interpretable "consciousness spectrum" inspired by philosophical, contemplative, and developmental frameworks. The authors construct a 636-sentence dataset spanning a seven-tier taxonomy (from Collapse to Unity) with continuous scores ranging from -5 to 5. Through probing, projection-ablation, and k-nearest-neighbor (kNN) graph geometry analyses across four sentence encoders and six large language models (3B to 27B), the study demonstrates that embedding spaces exhibit both a global linear direction and a low-dimensional local manifold structure aligned with this spectrum. Furthermore, the authors propose an inference-time steering method (local_manifold) that recomputes the intervention direction token-by-token based on local graph transitions. Experiments show that compared to a standard global linear steering vector (global_ridge), manifold steering better preserves long-form factuality at higher steering strengths and improves response quality, emotional attunement, and safety on prompts targeting low-tier psychological states.

Critical strengths
  1. Traditional activation steering relies on static difference-of-means vectors. In contrast, this work introduces a graded, trajectory-based manifold navigation framework that mirrors progressive cognitive and emotional shifts, opening up a promising direction for more nuanced model alignment.

  2. The authors provide comprehensive controls against surface confounds (such as sentiment valence, lexical abstraction, formality, and writing style) via projection-ablation and alternative-labeling tests. The incorporation of TwoNN intrinsic dimension estimation alongside multiple reference distributions adds solid empirical rigor to the geometric analysis.

  3. The qualitative case studies and pairwise evaluations reveal that local manifold steering effectively avoids the "spiritual bypass" failure mode often seen in global linear steering (where models inappropriately reframe severe distress or suicidal ideation as spiritual awakening), instead meeting the user's current state before gently guiding them upward.

Critical weaknesses
  1. The paper states that intervention magnitude is strictly matched via ∥δeff∥=α∥ht∥\Vert \delta_{\text{eff}} \Vert = \alpha \Vert h_t \Vert to isolate the effect of direction. However, in local_manifold, when no upward neighbor exists (N+(i)=∅\mathcal{N}^+(i) = \emptyset), no perturbation is applied (δ=0\delta = 0). This means local_manifold potentially exerts a lower cumulative intervention volume over long generations compared to global_ridge (which continuously pushes tokens along a fixed axis). It remains to be fully clarified whether long-form factuality preservation stems from the manifold direction itself or from the implicit early-stopping/zero-intervention gating when reaching high-tier regions.

  2. The kNN graph and manifold trajectories are constructed from a relatively small dataset of 636 introspective, first-person short sentences. Projecting real-time hidden states during open-ended, multi-turn, or domain-specific generation (e.g., historical or scientific questions) onto these psychological anchors might incur OOD mapping artifacts that warrant further discussion.

Questions for authors
  1. What is the empirical percentage of generated tokens in long-form generation and low-tier prompts where local_manifold applies a non-zero perturbation (δ≠0\delta \neq 0) versus performing a zero-intervention step? If a gating mechanism is added to global_ridge to skip updates when the projection is already high, does the performance gap persist?

  2. Could you specify the exact numerical values used for the transition utility coefficients (λsim\lambda_{\text{sim}}, λΔ\lambda_\Delta, λjump\lambda_{\text{jump}}), step threshold ϵ\epsilon, and softmax temperature τ\tau in Appendix F.3, and discuss how sensitive the trajectory navigability is to these weights?

  3. Have you tested your hypothesis regarding Gemma-2-27B (i.e., that near-constant dimensions absorbing 90% of squared norm artificially tripled the relative push in the content subspace at α=0.5\alpha=0.5) by centering the activations or removing those principal components prior to scaling?

2 comments

Dear Reviewer gymF,

Thanks very much for your review and insights! We really appreciate them and would like to address the following topics:

(1) Weakness 2

The kNN graph and manifold trajectories are constructed from a relatively small dataset of 636 introspective, first-person short sentences. Projecting real-time hidden states during open-ended, multi-turn, or domain-specific generation (e.g., historical or scientific questions) onto these psychological anchors might incur OOD mapping artifacts that warrant further discussion.

We agree this is a real limitation. During the feasibility stage of this project, we checked how out-of-domain inputs map onto the anchors. We took the layer-24 hidden state of the last prompt token in Qwen2.5-3B (raw prompt, without the chat template) for 1,000 SimpleQA questions and assigned each prompt the majority tier among its 7 nearest anchors. These prompts map mainly onto mid-to-low tiers (Conflict 39%, Clarity 36%, Striving 16%). The topic breakdown is interpretable: geography and science questions map mostly to Clarity (70%, 45%); entertainment and sports mostly to Conflict (45–54%); politics splits between Conflict and Clarity (44%, 34%); and history is the most spread out (Conflict 42%, Clarity 22%, Striving 20%, Activation 16%).

This analysis covers only the prompt itself. During steering, however, local_manifold looks up the nearest anchor again at every generated token, so for this response we also measured the mapping during generation. We re-encoded the saved long-form answers of all six models with the unsteered model, which reproduces exactly the hidden states the steerer saw. The mean cosine between generated-token states and their nearest anchor is 0.33–0.54 across five models (in Gemma-2-27B it is 0.91, inflated by the near-constant coordinate discussed below), so during generation the states can lie well away from the anchors. Two points limit the impact. First, local_manifold uses the anchor only to choose a local direction; so a poor mapping makes the direction less informative but still pointing "uphill". Second, all factuality evaluations (TriviaQA, SimpleQA, long-form) are themselves out-of-domain for the anchors, so the reported results already include this effect. We agree that more diverse anchor set is a natural next step.

(2) Question 1

What is the empirical percentage of generated tokens in long-form generation and low-tier prompts where local_manifold applies a non-zero perturbation (δt≠0\delta_t \neq 0) versus performing a zero-intervention step? If a gating mechanism is added to global_ridge to skip updates when the projection is already high, does the performance gap persist?

local_manifold skips a token only when its nearest anchor is a local peak, i.e. has no higher-scoring neighbor in the kNN graph. Each model has only 1–3 such anchors out of 636 (all Unity, except one Clarity anchor in Qwen2.5-14B). Replaying the saved generations of all six models, local_manifold applies a non-zero perturbation to 97.8–99.96% of generated tokens in long-form answers and 99.8–100% on low-tier (mental-health) prompts. Its cumulative intervention magnitude therefore matches the linear baseline to within 3% in every model and at every α\alpha, so the performance gap is not explained by skipped updates.

We now report this in Section 3.4 (Steering Results), with per-model numbers in a new appendix P.

A gate could well reduce over-pushing by the linear steer. How to set such a gate, however, is a separate design question: without the anchors on the manifold to guide it, the threshold and the stopping rule would have to be chosen by hand. What we compare here is the two methods at the same per-token magnitude, where manifold steering provides partial cancellation and stopping without any gating parameters: its pushes partly cancel as the direction follows the local structure, and it stops only where the nearest anchor has no higher-scoring neighbor. We have not tested a gated linear steer, and agree it needs careful design and would be a useful comparison for future work.

(3) Question 2

Could you specify the exact numerical values used for the transition utility coefficients (λsim,λΔ,λjump\lambda_{\mathrm{sim}}, \lambda_{\Delta}, \lambda_{\mathrm{jump}}), step threshold (ε\varepsilon), and softmax temperature (τ\tau) in Appendix F.3, and discuss how sensitive the trajectory navigability is to these weights?

Values. In F.3 we use λsim=0.9\lambda_{\mathrm{sim}}=0.9, λΔ=0.8\lambda_{\Delta}=0.8, λjump=1.0\lambda_{\mathrm{jump}}=1.0, ε=10−6\varepsilon=10^{-6} (strictly uphill moves only), τ=0.1\tau=0.1, k=30k=30, and at most 25 steps. For generation-time steering, we use λsim=0.3\lambda_{\mathrm{sim}}=0.3 with the other values unchanged, because a higher similarity weight produced near-baseline outputs. All values are now listed in F.3.

Intuition. Within the utility U=λsimcos⁡+λΔΔ−λjumpΔ2U=\lambda_{\mathrm{sim}}\cos+\lambda_{\Delta}\Delta-\lambda_{\mathrm{jump}}\Delta^2 (over neighbors with Δ>ε\Delta>\varepsilon), the terms 0.8Δ−Δ20.8\Delta-\Delta^2 form a downward-opening parabola that peaks at Δ=0.4\Delta=0.4. The quadratic component therefore favors moderate uphill steps while increasingly penalizing large jumps toward high-score regions. Dividing by τ=0.1\tau=0.1 scales utility differences by a factor of 10 in the softmax, making the choice relatively decisive.

Sensitivity. We added a one-at-a-time sensitivity analysis to F.3 (all 10 embeddings; each entry averaged over 5 seeds). Across the tested values λsim∈{0,0.3,1.5}\lambda_{\mathrm{sim}}\in\{0, 0.3, 1.5\}, τ∈{0.05,0.2,0.5,1.0}\tau\in\{0.05, 0.2, 0.5, 1.0\}, and λjump∈{0.5,1,2}\lambda_{\mathrm{jump}}\in\{0.5, 1, 2\} (for λjump=2\lambda_{\mathrm{jump}}=2, with a 100-step cap, since smaller steps make walks longer), the fraction of walks reaching Unity is 0.93–1.00, the tier–step correlation remains within 0.86–0.93, and the mean first-visit positions follow the order Activation →\rightarrow Growth →\rightarrow Clarity in all 100 embedding–setting combinations. The similarity weight and temperature have little effect, because restricting moves to kNN neighbors already enforces locality. The quadratic term sets the step size: λΔΔ−λjumpΔ2\lambda_{\Delta}\Delta-\lambda_{\text{jump}}\Delta^2 peaks at Δ∗=λΔ/(2λjump)\Delta^{\ast}=\lambda_{\Delta}/(2\lambda_{\text{jump}}), 0.4 score units by default. The mean step follows it: 0.35–0.40 at the default, 0.22–0.25 at λjump=2\lambda_{\text{jump}}=2, and 0.57–0.73 at λjump=0.5\lambda_{\text{jump}}=0.5. Since the total climb along ww is similar in every setting (5.6–8.0), path length scales inversely with the step: 17–22 nodes at the default, 11–13 at λjump=0.5\lambda_{\text{jump}}=0.5, and 26–35 at λjump=2\lambda_{\text{jump}}=2. In all these settings, Activation, Growth, and Clarity are first reached in that order. Only without the quadratic term (λjump=0\lambda_{\text{jump}}=0) do steps become large (1.2–2.7), paths shrink to 4–6 nodes, and the order breaks down (Growth before Activation in 7 of 10 embeddings).

(4) Question 3

Have you tested your hypothesis regarding Gemma-2-27B (i.e., that near-constant dimensions absorbing 90% of squared norm artificially tripled the relative push in the content subspace at α=0.5\alpha = 0.5) by centering the activations or removing those principal components prior to scaling?

We examined this on both the anchors and the generated tokens. At the steering layer (L28) of Gemma-2-27B, a single coordinate (dim 4140) is nearly constant: 0.94 ± 0.006 across the 636 unit-normalized anchors, carrying 89% of their squared norm. Replaying the long-form generations shows the same during generation: it carries a median 87% of ∥ht∥2\lVert h_t\rVert^2, for word and punctuation tokens alike (≈14.6k tokens per condition). Because it is nearly constant, it lies in the mean μ\mu rather than in the top principal components (it accounts for 0.06% of the anchors' variance), so centering removes it, whereas dropping top components would not. The ratio ∥ht∥/∥ht−μ∥\lVert h_t\rVert/\lVert h_t-\mu\rVert has a median of 2.9 (10–90%: 2.3–3.3), so norm-matched scaling pushes the varying part of the activation about 2.9× harder than α\alpha suggests.

Centering before scaling would divide every push by about 2.9. Within Gemma-2-27B, this is approximately equivalent to using a smaller α\alpha, which our grid already covers. The informative test is across models, at the same centered strength α∥h∥/∥h−μ∥\alpha\lVert h\rVert/\lVert h-\mu\rVert. At α=0.5\alpha=0.5, the other five models reach a centered strength of 0.8–1.2 (estimated on the anchors) and lose 14–30 points of long-form factuality under global_ridge. Gemma-2-27B at α=0.3\alpha=0.3 (centered strength ≈0.9) loses 15 points, within this range. Its α=0.5\alpha=0.5 corresponds to a centered strength of 1.4, beyond anything the other models reach on our grid. Gemma-2-27B is therefore not more fragile; its collapse at α=0.5\alpha=0.5 reflects a larger effective push.

Again, thank you very much for your time and consideration. Please feel free to let us know if you have any further questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.