Representation Geometry in Language Models

Sparse AutoEncoders (SAEs) offer a promising way to inspect language model representations, but it is still unclear what kind of linguistic structure their latents expose. We use part-of-speech (PoS) categories as a controlled test case to study whether morpho-syntactic information is encoded by individual latents or by structured groups of features. We find that PoS distinctions are highly recoverable from SAE activations, but do not align with one-to-one latent / category mappings. This recoverability is not reducible to lexical memorisation, and Open and Closed PoS classes differ substantially. Categories are supported by compact groups of sparse latents, with substantial variation across tags. These groups remain stable on held-out data, while also showing overlap between related categories. Our results show that SAEs localise morpho-syntactic information in a distributed and category-dependent form rather than through atomic grammatical features.

Reordering a set of mathematical rules without changing its meaning should preserve the correct answer, but must a model's internal representations stay invariant too? We investigate this question using synthetic multi-step function-composition problems, each presented under multiple rule orderings with the same correct answer. We measure accuracy and permutation signal-to-noise ratio (SNR), which quantifies how distinctly ordering patterns are represented relative to variation across problem instances. Across 16 language models ranging from 1B to 8B parameters, we find a pattern: models that solve reordered problems more accurately represent different rule orderings more distinctly. Layer-averaged permutation SNR is positively rank-correlated with accuracy in every synthetic setting we evaluate, with Spearman correlations reaching 0.86. These findings highlight a distinction between answer invariance and representation invariance: successful mathematical rule composition can accompany distinct internal representations between equivalent rule orderings. This motivates distinguishing answer invariance from representation invariance, and offers a representational perspective on mathematical reasoning beyond answer accuracy alone.

Meaning identity (whether two sentences say the same thing after wording changes) is treated in retrieval and RAG as a geometric fact about independently encoded sentence vectors. We show that, for frozen off-the-shelf encoders and language models, it is not: identity is computed when both sentences share one forward pass, and is not a property of the embedding geometry those systems ship. On overlap-matched PAWS-X, purpose-built encoders (BGE, E5, GTE, MiniLM, E5-Mistral-7B) reach English confirm AUC only 0.55-0.65 (dense peak 0.70). Independently encoded last-token states of Llama 3, Mistral, and Qwen do no better; late fusion of the two vectors stays near chance. The same probe on a joint forward pass reaches 0.90-0.96 from 1.5B to 32B, collapses under partner shuffle, is mid-depth, saturates near 0.94 by 3B, and appears more weakly in GPT-2 XL (0.76). The gap holds beyond Llama-style models on other causal LMs, bidirectional encoders (DeBERTa, RoBERTa), and encoder-decoders (Flan-T5, T5, BART). Fixed or linear readers over frozen independent encodings never unlock identity; nonlinear pair readers recover part of it only on the full 49k-pair PAWS train split (0.68-0.87). Off-the-shelf rerankers split: BGE-reranker-large reaches 0.94, while MS-MARCO and Jina stay at 0.55-0.64. Independently trained families compute the same relation and a 1.5B joint reader can distill it from unlabelled teacher scores, while no linear function of the teachers own independent vectors can. Bi-encoders can be fine-tuned to fit PAWS (0.87-0.93), but transfer and STS-B suffer. Cosine compares wording neighbourhoods; identity is a cheap computed operator, not a property of either sentence vector.

The Linear Representation Hypothesis associates high-level concepts with directions in language models, but it remains unclear how these concept-related linear structures are organized within the model. We propose the Answer-Basin Representation Hypothesis: the probability measure induced over answers by the model's continuation distribution organizes these linear structures, with its statistics represented along linear directions shared across questions. All continuations yielding the same answer form an answer basin, whose mass is their total probability. These basin masses define the pushforward probability measure over answers. We posit that concept-related linear structure emerges from differences in the answer measure rather than being determined by changes in concept labels. Experiments across models and tasks link concept-consistent effects and their reversals in probing and steering to the alignment between concept labels and the answer measure.

The quadratic scaling of Transformer self-attention has driven the adoption of sub-quadratic Selective State Space Models (SSMs) like Mamba, which compress past context into a fixed-size recurrent hidden state. This strict informational bottleneck raises a foundational question for mechanistic interpretability: do SSMs and Transformers learn fundamentally distinct latent representations? In this work, we employ Sparse Autoencoders (SAEs) to conduct a large-scale, feature-level correspondence analysis between Mamba-130m and Pythia-70m over a 10-million token corpus. Contrary to hypotheses predicting widespread architectural divergence, we find no evidence of systematic representational divergence between architectures: across the observed Jaccard distribution, 99.98% of Mamba features cluster toward the upper alignment boundary, providing preliminary feature-level support for the Universality Hypothesis. We further identify and qualitatively characterize this microscopic fraction (0.02%) of diverging features, finding patterns consistent with the hypothesis that the recurrent bottleneck selectively limits the parsing of rigid syntax rather than broad semantic ontology. We demonstrate that while Pythia's unconstrained attention permits the monosemantic decomposition of distinct formatting edge-cases, Mamba is forced to compress unrelated syntactical anomalies into polysemantic "junk drawer" neurons to preserve state capacity. Collectively, these results suggest that architectural routing mechanisms may have negligible impact on core semantic understanding, with representational divergence confined to extreme structural margins.

The Platonic Representation Hypothesis (PRH) claims that independently trained models converge on a shared statistical model of reality, yet recent work finds only weak pointwise similarity between models. In this paper, we show that what models share is not the location of samples in representation space, but the directions (displacement vectors) between them. Under a single orthogonal alignment--rotation and reflection only--these displacement vectors are substantially preserved across 44 independently trained vision and language encoders spanning modalities and asymmetric capability pairs, consistent with the PRH evidence. The samples' absolute positions are not, consistent with recent counter-evidence. Both arise from a single decomposition: representations split into a shared semantic component that is linearly aligned across models, and a private capability component that is not. We trace this geometry to concept-level structure: within a model, parent concepts are orthogonal to their child variation vectors; across models, concept displacements are parallel. Our theory falsifiably predicts (and experiments confirm) that fine-tuning preserves pointwise similarity but collapses displacement, and that relational distillation does the opposite. A major implication is that, because semantics align linearly but capabilities do not, capabilities can be imported from one model to another using a single cached forward pass through the source. We call this Shadow Casting. As a proof of concept, our SHADOWCLIP instantiation outperforms strong fine-tuned baselines at orders of magnitude less compute. A cache can be released alongside open model weights, letting one model's capabilities be downloaded and imported into any number of other models without fine-tuning.

Sparse attention reduces the quadratic cost of global self-attention while retaining strong empirical performance, but how its restricted interactions shape the evolution of token representations remains theoretically underexplored. Modeling tokens as particles on the unit sphere, we introduce opinion leader dynamics, a framework that identifies two mechanisms through which token groups converge internally while maintaining distinct limiting directions. In the explicit model, fixed representatives induce a potential that attracts tokens toward distinct local maxima. In the implicit model, disconnected interaction groups evolve toward separate consensus directions. We formulate both models as reverse Wasserstein gradient flows and establish exponential convergence under suitable conditions. We further connect these theoretical predictions to token evolution in frontier sparse-attention LLMs that motivate our framework. Across four benchmarks, Kimi-K3, MiniMax-M3, and DeepSeek-V4-Flash consistently exhibit clearer cluster separation and higher clustering scores than the dense-attention model GLM-4.7-Flash in projected token representations. These observations support the relevance of the predicted multiple-group structure to trained frontier LLMs, while finite-particle simulations illustrate the theoretical convergence behavior. Together, our results connect restricted token interactions to distinct group-level attractors, providing a dynamical account of how sparse attention can support alignment within groups while preserving separation between them.

Claims about what an internal direction in an LLM represents need evidential constraints beyond an observer's prior beliefs about the system. Decodability and successful activation steering do not, by themselves, establish which construct the direction tracks. This gap is especially consequential for welfare-relevant interpretations, where a proposed functional state must be distinguished from correlated features of the extraction contrast. We treat the question as one of construct validity and study a good-bad outcome direction in a maze task, using controlled interventions that separate the realised outcome from the informational history through which it became known. Across multiple LLM checkpoints, directions fitted on one explicit outcome encoding transfer well to another, indicating that the readout is not tied to surface form. When the same realised outcome is reached through announced and unannounced histories, however, transfer degrades substantially: even after both histories receive the same explicit outcome, the post-event readout remains strongly conditioned on the earlier announcement. In the base model, steering along the direction changes actions, yet removing it leaves the natural cue effect almost intact. In a matched maze-RL run, the post-RL direction becomes substantially more predictive of reference-MDP remaining return and the policy becomes more dependent on it at the tested sites, while the history dependence persists. These dissociations support a functional, value-related interpretation of the direction, but not its identification with a history-invariant scalar valence state.

Pronouns, adverbs and other functional words (such as they, her, somewhere, there) are often used in language to replace concrete nouns or phrases, when their properties - such as gender, grammatical number - provide sufficient information for the given context. Do pretrained transformer models encode such functional words in a manner that allows them to be used like humans do? Can language models recognize the syntactic and semantic parallelism of sentences such as "The researchers wrote the paper" and "They wrote it", which relies on such lexical abstraction? We map these linguistic questions into the embedding space of a pretrained transformer model, and compare representations of nouns, with the representations of the pronouns and adverbs that can replace these nouns, in isolation and in parallel lexicalized and functional sentences. We then probe for shared syntactic and semantic structure in the embeddings of parallel lexicalized and functional sentences. We find that functional words are located centrally compared to nouns, but are also distinct, which is congruent with their behaviour as place-holders in a wide variety of contexts. The analysis of the embeddings of parallel (lexicalized and functional) sentences show them inhabiting different subspaces of the embedding space. Experiments that distil the structural information of the sentence show that training on either type of data does not reveal the shared structure - because of the over-consistency of the vocabulary (in case of the functional data), and the too much variety (in case of the lexicalized versions). However, training with a mix of functional and lexicalized sentences, the shared structure emerges.

Large language models (LLMs) trained on next-token prediction exhibit remarkable in-context learning (ICL) abilities, yet the representations that support ICL remain poorly understood. We consider such representations in a controlled setting: prompting LLMs with data emitted from hidden Markov models (HMMs) and probing for the corresponding belief state -- the posterior distribution over the HMM's hidden states given the observed token history. Across six open-source LLMs prompted with data from 40 HMMs selected for non-trivial belief structure, we find that belief states are linearly decodable from residual stream activations, with peak probe $R^2$-values from 0.83-0.99 across HMM and LLM combinations, ranging from early to late layers. To establish functional relevance, we intervene directly on the probe-identified subspace via patching and steering, resulting in downstream prediction quality on the order of the untampered model, while controls degrade performance substantially. Together, these results provide representation-level evidence that ICL in open-source LLMs approximates optimal Bayesian prediction over a context-inferred generative model. More broadly, our findings extend prior results linking input-distribution structure to activation geometry: from toy networks trained explicitly on HMM data to production-scale LLMs.

A language model's representation geometry is not predetermined; it evolves as the model runs. A faithful account of that geometry must capture that dynamic process, and so cannot be based solely on model-independent statistics such as co-occurrence. Here we introduce a mean-field analysis of attention. The average attention from one token to another defines a kernel that carries representations layer to layer and can be iterated through the network to model how the geometry is transformed. We condition this average two ways. Conditioned on a whole corpus, the kernel predicts the average-case evolution of representation geometry. Conditioned instead on a single context, it predicts the expected geometry for that context. A head's departure from that prediction, its \emph{mean-field deviation}, isolates the context-specific computation that the mean field misses. Under the corpus-conditional reading, the kernel yields an open-loop model: from the input embeddings and the frozen weights alone, we can iterate the kernel and the model's own MLPs over token representations, never consulting a measured deviation at any layer. The resulting prediction is highly accurate. In early training the model and its corpus mean field are indistinguishable. Replace every attention head with its mean field, and the substitution leaves the loss on real text unchanged. Around the onset of induction, the two diverge, and the gap widens as representations become contextualized. Under the context-conditional reading, deviation from the mean field is a task-agnostic measure of context-specific computation. The residual decomposes additively into unusual attention routing and contextualization of the transported values. Across controlled induction and few-shot settings, greater deviation tracks greater reliance on in-context information.

Transformer representations evolve through learned additive transformations that either preserve their current direction or redirect it. We study this evolution as a functional geometry, decomposing learned updates into parallel and perpendicular components. Across pretrained models, we find substantial parallel components beyond the residual identity path. We then apply the decomposition in two spaces: to attention and MLP updates relative to the hidden state, and to attention value aggregation relative to the current token's value. Targeted edits reveal a strongly space-dependent asymmetry: exclude-self value-space parallel manipulation is markedly more robust than residual-space and perpendicular counterparts, preserving the direct self message while scaling only the non-self aggregate. The same decomposition gives a component-resolved description of compression-induced update error: perpendicular error separates compression methods more clearly than parallel error. Extensive experiments further demonstrate that full-aggregate parallel suppression during from-scratch pretraining lowers validation-loss trajectories and improves downstream averages, with the value-space variant strongest. Together, these results connect representation geometry to editing robustness, compression diagnosis, and training-time intervention. Code is available in the project repository.

Large language models (LLMs) with human-like performance on linguistic tasks have transformed the study of language in neurodiverse conditions. LLMs provide representations of linguistic input in the form of high-dimensional vectors (embeddings), and next-token predictions computed from these embeddings. Previous crosslinguistic evidence suggests a complexity reduction in the form of both lower intrinsic dimensionality (ID) of LLM representations and higher mean surprisal (prediction error) in psychosis. We hypothesized that these metrics reflect a general deficit of information compression in psychosis, linked to grammatical organization as what enables predictions in language.We operationalized surprisal difference as the difference between surprisal as estimated from word frequency and surprisal as based on a contextual LM, which is sensitive to grammatical organization over and above lexical concepts. Using a dataset of 144 Turkish speakers, including 106 patients with schizophrenia-spectrum disorders (SSD) - 56 with chronic schizophrenia (SZH), 33 with first-episode psychosis (FEP), and 17 with schizoaffective disorder (SZA) - and 38 healthy controls. We report: (1) Surprisal difference is attenuated in all clinical groups relative to controls, independently of word count; (2) Compressibility (intrinsic dimension) is reduced in SZH and FEP; (3) Syntactic complexity and compressibility both predict surprisal difference. These results, further refining an alteration in the geometry of the semantic space in psychosis as previously attested, suggest a broader deficit in information compression in this disorder, with a mechanistic underpinning in the operations of grammar.

A century of psychology has found that the trait words people use to describe one another vary, but the relational structure among those traits, which ones go together and which oppose, is strikingly consistent across raters and cultures. We test whether the LLM (Qwen 2.5-7B-Instruct) reproduces this structure in its internal trait representations. From millions of crowd-sourced personality ratings of fictional characters, we build a human implicit-personality matrix over hundreds of traits; from contrastive model activations, we build a matching matrix over the same traits. The two relational structures align strongly (Mantel r = 0.77), and the agreement holds trait by trait as well as in aggregate. Two dominant axes of the model's trait representations recover the social and intellectual dimensions long known to organize human personality impressions, social warmth and intellectual competence. On held-out dialogue, projecting model activations onto these directions yields personality profiles that agree with human ratings. This work establishes a framework that enables comprehensive, human-grounded comparison between internal model trait geometry and the shared structure of human personality impressions.

Fine-grained control of language model behaviors (e.g., steering) is among the more actionable outcomes of interpretability research. For binary concepts such as refusal, a single direction in activation space often suffices for steering. However, many concepts are not binary: Animals and Countries contain many subcategories, each with multiple instances. For these concepts, the search space over possible representation geometries is far larger than for binary concepts; it is thus not clear what geometries are most appropriate, nor what methods are most effective at recovering them. In this work, we introduce MAxBench, a geometry-agnostic evaluation framework for multinomial concept representations based on sampling from the recovered concept representation. We use MAxBench to compare 10 localization methods (covering 5 geometry types) across 6 concepts and 4 models. Using this framework, we find that (i) affine subspaces steer more reliably and have greater recall than rank-one or linear subspaces; (ii) much of this advantage is due to better non-zero offsets rather than the choice of bases; (iii) manifold steering is competitive with the best methods when applicable; and (iv) no method consistently outperforms prompting, in alignment with prior findings on binary concepts. These findings underscore the importance of expanding the scope of interpretability research and meta-evaluation to concepts with more varied structure.

A deployed language model may refuse a harmful request in English yet comply with its faithful translation, revealing a cross-lingual safety failure that cannot be characterized reliably by output behavior alone. We formalize this phenomenon through an audited equivalence relation and show that, for a declared quotient, representation, metric, feature dictionary, scoring head, threshold, and contrast model, the resulting safety drift admits an exact linear-algebraic characterization. Specifically, the drift is a cross-Gram functional of the within-fiber contrast; its worst admissible value is a support function, while margin invariance is characterized by an annihilator condition. We introduce an intrinsic calibrated exposure measure, governed by the leverage duality $χ^2=1/\ell-1$, which separates observed drift into three diagnostically distinct regimes: a reader fault removable by recalibration, an exact correction that is too ill-conditioned to be reliable, and a representation-level collision that no readout-only intervention can remove. Thus, identical observed exposure can lead to fundamentally different remediation verdicts. The framework also extends to cone-valued safety heads. An untied order-swap identity provides a diagnostic for the linear control interface; its calibration-state residual predicts a distinct three-control composition error on unseen states and targets, achieving median Spearman correlation $0.964$, compared with $0.269$ for a static cross-Gram baseline. etc.....

Continuous reasoning has emerged as a promising way to improve reasoning in large language models (LLMs). Yet we still lack a clear principle for deciding what a latent state should preserve. Reasoning by superposition shows that a single latent state can encode several search alternatives and expand them in parallel. We ask how those states should be weighted as reasoning proceeds. A natural choice is to preserve only the states active at the frontier step, since keeping every reached state appears to spread a limited hidden width too thin. We show that the opposite can hold. When later computation draws on several reached states, a cumulative state can guide attention correctly at a smaller hidden width than a frontier state that stores fewer states. At the same width, the cumulative state therefore keeps more intermediate states available for later reasoning. More generally, equal cumulative weights are optimal when future queries are unknown and remain close to the best task-specific weights when those queries are known. Experiments with two-layer and GPT-2 Transformers reproduce the predicted width advantage and show that unequal weights fail first on the states that receive the least weight. This suggests a important principle: keep reached states equally weighted, and restore equal weights as computation proceeds.

Activation steering provides a lightweight way to control large language models (LLMs) by modifying their hidden activations at inference time. Among these approaches, norm-preserving steering aims to change model behavior without altering the activation norm, reducing the risk of representation collapse and degradation. However, existing norm-preserving methods are limited by predefined steering trajectories and by their reliance on one-step updates, which may fail to capture the complex structure of activation distributions. We propose GeoSteer, an optimization-based method for norm-preserving activation steering. GeoSteer formulates steering as a Riemannian optimization problem and updates activations through a sequence of small geodesic steps on the representation manifold. To avoid fixed steering directions, GeoSteer learns a nonlinear activation-space objective that distinguishes desired from undesired activations, and uses this function to adaptively guide each steering step. This multistep formulation yields smoother, more stable, and more consistent steering behavior while preserving the activation norm. Across TruthfulQA, RealToxicityPrompts, and UltraFeedback benchmarks, GeoSteer consistently improves over state-of-the-art activation steering baselines. These results suggest that norm-preserving steering can be made more effective by replacing predefined one-step edits with adaptive, geometry-aware optimization.

We study K/V-cache interventions -- transplanting a target-conditioned K/V trajectory into a source-persona generation -- as a structured surface for persona control in decoder-only language models. Across 13 intervention configurations applied to Llama-3.1-8B for a fixed source-to-target persona pair, we report two consistent dissociations between representation-level alignment and behavioral expression, plus a common failure under position perturbations. First, all layer-band K/V replacements (early, mid, late) achieve strong local V-space alignment (V-gap 0.91, 0.89, 0.84), but only mid-layer replacement (layers 9-20) combines substantial target-marker expression with preserved lexical diversity. Second, full and mid-layer replacement induce comparable alignment (V-gap 0.94 vs. 0.89) yet produce different lexical-diversity profiles (TTR 0.65 vs. 0.77). Third, position perturbations (lag and shuffle) apply distinct operations yet uniformly suppress target-persona expression -- a common behavioral failure rather than a strict dissociation. Representation-level similarity metrics alone are thus not sufficient predictors of downstream persona expression in the regimes we study; the K/V cache emerges as a controllable but structurally constrained intervention surface. Because the transplanted trajectory carries the target's own generated token history, we characterize the intervention as trajectory-level transplantation rather than isolated persona-representation injection; a same-token-sequence control, decoding an identical token sequence under source vs. target conditioning, reproduces the sign and layer localization of the L28 representational shift, indicating the shift is not explained solely by imported token history. These findings characterize representation-behavior dissociation in a high-signal setting rather than establishing universality across models or persona pairs.

Large language models learn similar behaviours, yet it remains unclear what structure they share or how to change one behaviour without disturbing others. The Fisher-Rao geometry of next-token probabilities connects these questions: behaviour determines this geometry up to output-preserving symmetries, whereas activation geometry depends on coordinates. Across transformer, state-space and recurrent models, output geometries agree more strongly than activation geometries, and shared geometry supports semantic-category transfer. Agreement with human word choices increases with predictive accuracy, scale and training, and improves further after model-only calibration. Token probabilities and read-out geometry jointly predict the spectrum and its effective dimension. Controlled language assignments show that geometry follows the language law across architectures. Pretraining corpus statistics predict held-out fact acquisition without recalibration, while randomised experiments show that deeper evidence substantially delays acquisition across every tested architecture and evidence construction. Finally, the geometry prescribes minimum-disturbance local interventions, predicts their relative cost, and supports reusable control: updates learned on donor prompts transfer to unseen prompts while better preserving behaviour on reference prompts than Euclidean control. The same geometric correction improves steering, editing, attribution, dictionary learning and fine-tuning.

Recent state-space models (SSMs) such as Mamba achieve language modeling performance comparable to transformers despite relying on fundamentally different architectures. This raises an important question: how do these structural differences influence the geometry and functional nature of their internal representations? We study this question through a multi-scale analysis of representations in transformers, SSMs, and hybrid architecture. First, we find that SSMs distribute their representational information evenly across all dimensions, whereas transformer representations are heavily dominated by a single principal direction. By evaluating hybrid architectures, we observe that the representation space becomes increasingly skewed toward a single dominant direction after each attention layer. Next, we explore how the different geometric spread of representations impacts representational capacity through compressibility. Surprisingly, we find that despite their contrasting geometric structures, both architectures exhibit tightly matched effective capacities. We further investigate whether this skewed geometry affects how concepts are encoded. Using rank-constrained probes, we demonstrate that both architectures encode concepts in subspaces of surprisingly similar dimensionality. Furthermore, we demonstrate that the transformers' dominant principal direction does not inherently encode more conceptual information. Finally, we zoom in and examine the alignment between manifolds, either by analyzing representations of specific topics or by looking at the nearest neighborhoods of tokens, and find that they are highly aligned. Ultimately, our analysis suggests that while transformers and SSMs induce different usage of latent space, they display a striking functional convergence at the level of local semantic manifolds.

Recent methods in language model interpretability employ techniques such as sparse autoencoders to decompose residual stream contributions into linear, semantically meaningful features. Such methods are commonly interpreted as identifying features that persist in the residual stream and that subsequent layers build upon. We challenge this view by identifying the Transformer Layer Correction Mechanism (TLCM), wherein adjacent transformer layers systematically counteract portions of each other's contributions. TLCM appears in 5 out of 7 major open-source model families and activates across nearly all tokens in diverse texts. We show that TLCM emerges during pretraining, operates most strongly on contextually dependent tokens, and adaptively calibrates its correction strength based on the preceding layer's output. Using the layer Jacobian, we further show that TLCM selectively corrects specific subspaces while reinforcing others, which we interpret through a ``propose-and-reject'' framework in which layers propose candidate features and subsequent layers selectively remove inappropriate ones. This dynamic suggests that the residual stream at any layer contains transient proposals alongside persistent features, helping explain why SAE feature descriptions often have low specificity, why effective model steering requires extreme feature amplification, and why transcoders hold a theoretical advantage over SAEs.

Chain-of-Thought (CoT) improves the reasoning ability of Large Language Models (LLMs) but incurs substantial computation and context costs. Existing methods either lose intermediate information through hard pruning or lack a principled criterion for continuous compression. We present A*-Thought-V2, a geometric dynamics of LLM guided framework that models CoT as a hidden-state trajectory and replaces hard deletion with an explicit-implicit interleaved latent architecture. After projecting question, step, and solution representations into a 3D PCA space, it measures alignment between each local transition and global question-to-solution direction. Aligned steps remain explicit text, whereas deviating steps are compressed into continuous latent tokens. Directional angles capture both local semantics and reasoning dynamics: small angles indicate direct execution and answer formation, while large angles more frequently involve checking, correction, and branch exploration; their temporal variation reveals exploration, convergence, and refinement stages. To train this architecture, we introduce stepwise embedding forcing, which pools each redundant step into a single latent embedding, and label forcing, which supervises that latent token with a soft multi-modal vocabulary distribution instead of a hard one-hot label. Experiments on Qwen3.5-9B and Qwen3.6-27B across six in-domain and out-of-domain benchmarks show that A*-Thought-V2 improves average accuracy by up to 2.6% while reducing response length by up to half, increasing Accuracy per Computation Unit by 2.29$\times$, and reducing preprocessing and training time by 94.6% and up to 80.3%, respectively. Representation analyses suggest that latent states form a compact region distinct from textual states, while higher entropy at latent-token positions reflects broader soft targets that encourage richer step-level feature learning.

Chain-of-thought (CoT) reasoning improves the reasoning ability of large language models by introducing intermediate computation, but explicit rationales increase decoding length, latency, and context cost. Implicit CoT offers a more efficient alternative by moving intermediate reasoning into continuous latent states. However, latent reasoning can be unstable: successive latent states may become overly similar and collapse toward a shared dominant direction, reducing the diversity of the reasoning trajectory. In this work, we identify $\textit{latent rank collapse}$ and propose $\textbf{WIDER}$, a lightweight spectral regularizer for implicit CoT. During training, WIDER estimates the shared direction of each latent trajectory and penalizes projections onto this direction, encouraging latent states to span a broader representational subspace. The method is plug-and-play and leaves the backbone model, latent schedule, and inference-time decoding procedure unchanged. We further formulate this collapse as a geometric bottleneck in implicit reasoning, casting its mitigation as a training-time regularization problem rather than an inference-time decoding change. Extensive experiments show that WIDER improves matched implicit CoT baselines, while mechanistic analyses reveal higher effective rank, lower dominant-direction energy, and reduced redundancy among latent steps. These results highlight latent subspace utilization as an important factor for efficient continuous reasoning, providing a geometric perspective for analyzing and improving implicit CoT. Code is available at https://github.com/whitesweater/WIDER.

Continuous chain-of-thought models compress reasoning into latent tokens. Matrix-valued variants, which route each latent token through a d x d matrix bottleneck, introduce rank as a single-sample structural observable on the latent matrix Z. If matrix latents carry parallel reasoning paths via superposition, rank should track them, and truncating Z to low rank should hurt accuracy on tasks whose solutions plausibly require multiple components. Across four training regimes of a matrix-CODI model (three on ProsQA, one on GSM8K-Aug below the learning threshold), the rank-k projection ablation curve is flat to within 0.6 percentage points. A three-seed replication yields 81.0 +/- 2.0 percentage points accuracy while the final effective rank of Z spans {4, 12, 13}; the loss does not reward any particular rank. To test whether rank-blindness arises from the flatten-then-project readout alone, we trained four readouts: a bilinear reparametrization, a bilinear-plus-GELU readout nonlinear in Z, an SVD-augmented readout feeding singular values through an MLP, and a quadratic readout in Z Z^T. All four rank-k curves remain flat (Spearman p-values 0.63, 0.14, 0.82, 0.46). The flat curves persist for readouts nonlinear in Z. A linear probe on Z underperforms a raw pretrained hidden state at target prediction (AUC 0.673 vs. 0.846). A negative control on vanilla GPT-2 SFT (no matrix bottleneck, no Z, three seeds, n=500) reproduces a flat rank-k curve under the same intervention paradigm with pooled-mean range 0.20pp, and a random-h sensitivity floor lands at the same accuracy: the rank-k ablation alone conflates rank-blindness with position-irrelevance.

LLMs latent-state reasoning methods replace discrete intermediate tokens with continuous states, such as weighted mixtures of token embeddings, to retain multiple possible reasoning directions rather than committing to one. Yet pretrained language models often fail to preserve these mixtures. We study why through a combination of theoretical analysis and controlled empirical investigations on a variety of models. We identify three independent, distinct sources of failure. First, transformer architectures already distort mixture geometry, and training substantially amplifies this effect. Moreover, the failure can occur even if the model transports mixtures perfectly linearly: the softmax readout and autoregressive feedback form a dynamical system that either amplifies small differences until one component of the mixture dominates or contracts different mixtures until they become indistinguishable. We verify this theoretical prediction empirically: the observed transition between contraction and amplification occurs near the theoretical threshold derived by our analysis, and pretrained-model rollouts lie predominantly on the amplifying side. Finally, we generalize to mixtures of many components and show that exact preservation generally requires context-dependent correction, whose required dimensionality can grow with the number of components.

Patients with depression present with diverse symptom profiles, yet clinical practice routinely reduces this variation to a single severity score. Large language models (LLMs) can potentially capture various symptoms and their severity from patient speech. However, how depressive symptoms are represented inside LLMs remains poorly understood, limiting clinical trust. To examine whether internal model activations match clinician judgment, we analyzed the residual stream of Gemma-3-27B-PT using mechanistic interpretability techniques. Recording activations across symptom descriptions drawn from validated clinical instruments, we found that symptom groups geometrically separated the most at layer 21 across multiple distance metrics. Using Semantic Projection, we then projected held-out naturalistic text onto Symptom Vectors constructed from these instruments. The resulting per-symptom coefficients preserved clinician-annotated rank ordering across mood, somatic, and suicidality axes. Furthermore, a single depression vector in Layer 21 separates held-out depressive from non-depressive text (AUC = 0.789), which can be used as an emotional valence gate that restricts symptom projection to depressive speech. These results reveal a decorrelated, clinician-aligned symptom signal readable directly from internal activations, offering a mechanistic foundation for interpretable depression-assessment tools.

Neural networks are increasingly employed to identify both well-defined and ambiguous concepts, yet output-level metrics reveal little about how those concepts are represented internally. Our study asks if these networks exhibit \textit{conceptual separation}: if examples of the same concept form coherent representations, and whether related concepts lie closer together in the representation space. We examine this conceptual organisation in Convolutional Neural Networks (CNNs) and Large Language Models (LLMs) through geometric and distributional analysis of their internal activations. In CNNs, familiar ImageNet concepts form coherent and semantically ordered representations, while this coherence weakens for unseen concepts and suffers within-class domain shift. In LLMs, clearly distinct domains remain well separated, related subdomains move closer together, and the distinction between ambiguous topics collapses at both the mean and covariance level. These results suggest that conceptual separation can reveal structure that output accuracy alone cannot, and may serve as a useful diagnostic of how robustly a model represents the concepts it is asked to identify. Code and data available on GitHub.

A transformer carries each token's state from layer to layer, and the whole vocabulary carried together forms a cloud that moves with depth. We ask whether a trained network moves this cloud the way optimal transport would: at the cheapest cost, and along the map that pairs each token with its optimal destination. We measure both on Pythia-160m and Pythia-410m, with an exact assignment between consecutive layer clouds, a measured sampling floor, calibration on couplings known to be optimal, and a split of the cost into the common shift of the cloud and the token-specific moves. At the last layer, both models move their tokens where the optimal-transport map sends them, at the optimal cost for Pythia-410m and slightly above it for Pythia-160m. At the first layer they do not. In between, single layers can be judged on cost at only two of ten transitions, and blocks of several layers move the cloud at close to the optimal cost. The agreement at the last layer is much weaker at initialisation (0.64 against 0.86) and grows with training.

Probing studies have established that syntactic information is decodable in early and middle transformer layers, but what happens to that information in later layers remains poorly understood. We apply a cross-layer generalisation analysis to three Greek-tuned large language models evaluated on tightly controlled minimal pairs: object-relative constructions in Modern Greek, where canonical (Subject-Verb-Object; SVO) and non-canonical (Verb-Subject-Object; VSO) orders differ only in within-clause word order, while preserving propositional meaning. When a probe trained on late layers (20-31) is tested on each early layer individually, it produces below-chance transfer (cluster-corrected, p<0.01), classifying 99.3% of non-canonical sentences as canonical. Probe coefficients reverse sign around layer 22, indicating a directional recoding toward the canonical form rather than simple information loss. These findings characterise a representational format change in late transformer layers that goes beyond the well-established decline in syntactic decodability, and they generate a directly testable prediction for human EEG and MEG decoding studies using the same stimuli. Code and stimuli are publicly available on OSF.