Representation Geometry in Language Models

Follow this topic Ask about this topic

Momentum

48 papers in the last four weeks, up 118% on the four weeks before. 0.5% of all new papers.

Jul 13Week of Sep 28

Latest papers 221

Aug 31, 2026cs.CL

Lingua Franca or Probing Artifact? Rethinking Latent Language in Multilingual LLMs

Latent language identification is often used to argue that multilingual language models route computation through language-specific states, such as English pivots. However, existing probes infer latent language from different signals, such as the geometry of hidden states or what can be decoded from intermediate representations. Since such claims shape conclusions about how models share and route information across languages, we ask whether these probes measure the same phenomenon or expose distinct aspects of multilingual computation. We study this question across model families, training regimes, domains, tasks, checkpoints, and up to 27 languages. We find that identification probes systematically disagree: the GMM-based representation probe, which draws evidence from hidden state geometry, shows earlier cross-lingual mixing, whereas decoding-based probes, which rely on output-space decodability, retain sharper language-specific and more English-biased signals. These differences track model multilinguality and training progression, but are comparatively stable across domains. Our results suggest a more cautious interpretation of latent language identification, where current probes expose different aspects of multilingual processing, rather than directly revealing a single internal lingua franca.

Deniz Bayazit, Badr AlKhamissi, Antoine Bosselut

EPFL

Multilingual Language Models Multilingual Language Model Evaluation Representation Geometry in Language Models Language Model Probing

Aug 31, 2026cs.AI

Geometry of Divergence: Tracking Hidden-State Trajectories for Adaptive Multi-Turn Reasoning

LLM agents need to sustain goal-consistent reasoning across long multi-turn interactions under strict resource constraints. However, as the multi-turn context accumulates, it can destabilize the underlying LLM's internal representation of task-relevant information from earlier turns, blurring the boundary between constructive reasoning and representation drift. We formulate multi-turn reasoning as a hidden-state trajectory of the underlying LLM that is characterized via two complementary signals: temporal curvature that captures the directional consistency of turn-to-turn updates, and variance slope which measures the expansion or contraction of the exploration space. Across four tasks and three underlying LLMs, we observed that these geometric signals distinguish between correct and incorrect episodes prior to completion. We further decompose each episode into three-action chains formed from four actions (Read, Write, Respond, Transfer) and show that separability is action-dependent, with different signals distinguishing various chain patterns. Our experiments demonstrate that trajectory geometry can identify critical turns in the reasoning process, increasing task success rates on

τ

-Bench from 24.1% to 39.6% while reducing token cost by 11.2%.

Jie Liang, Zhengxin Yu, Hamid Nasiri +1

School of Computing and Communications, Lancaster University, UK · School of Computing and Engineering, University of Huddersfield, UK

Long-Horizon LLM Agents Representation Geometry in Language Models

Aug 31, 2026cs.LG

Context Staircase: Signature-Aligned Dynamics of Token Embeddings under Small Initialization

Token embeddings are the basic representational units that connect discrete tokens with continuous computation in language models. Although modern language models learn embeddings from random initialization through gradient-based training, the dynamical mechanism by which meaningful embedding structures emerge remains unclear. In this work, we identify that the evolving embedding structures are closely related to token-conditioned label and contextual distributions, which we formalize as probability signatures. We observe a progressive learning process, which we term Context Staircase: embeddings learn the low-order statistic signatures of the data before the high-order ones. More specifically, we observe that early in training they align with the simplest, context-free signature linking a token to its label, and as training proceeds, they progressively reflect signatures involving more and more context tokens. We then analyze the gradient flow of embeddings under small initialization to explain this phenomenon, deriving embedding evolution equations for feed-forward and self-attention architectures. We further extend these observations to real language-model training. Finally, we show that these embedding structures play an important role in both task learning and the incorporation of semantic structure into the embedding space. Overall, our results provide a dynamic explanation of how data statistics and architecture jointly shape token embeddings in language models, and reveal an implicit bias in the space of data statistics: training proceeds from simpler, low-order statistical relations toward increasingly complex, context-dependent ones.

Junjie Yao, Liangkai Hang, Zhi-Qin John Xu

School of Mathematical Sciences Shanghai Jiao Tong University Shanghai 200240, China · Institute of Natural Sciences, School of Mathematical Sciences Shanghai Jiao Tong University Shanghai 200240, China

Language Model Pretraining Word Embeddings Neural Network Training Dynamics Representation Geometry in Language Models LLM Training

Aug 30, 2026stat.ML

Learning Representations through Token Prediction: Geometry, Approximation, and Downstream Guarantees

Token prediction is a central pre-training objective for modern language models. Despite its empirical success, why token prediction learns broadly useful representations remains incompletely understood. We develop a statistical framework connecting token prediction with representation geometry, encoder approximation, and downstream performance. Under a softmax prediction head, we show that accurate token prediction organizes token embeddings according to similarities between the distributions of contexts in which different token types appear, as measured by Hellinger distance, with explicit errors governed by prediction accuracy and token frequency. Meanwhile, the contextual representation provides a low-dimensional coordinate for the conditional distribution of the target token relative to these embeddings. We further introduce a self-consistency principle showing that repeated applications of a shared representation block can progressively refine the contextual representation without introducing additional block parameters. Among representations with the same prediction accuracy, this recurrent construction favors those that can be stably reconstructed from their contexts. Finally, we establish downstream guarantees for token generation, token community recovery, and classification by a linear probe, showing how prediction accuracy and recovered geometry translate into performance beyond the pre-training objective. Together, these results explain how the simple objective of predicting tokens can recover semantic geometry and produce broadly useful representations. A controlled simulation illustrates the theoretical mechanisms.

Shulei Wang

Department of Statistics, University of Illinois at Urbana-Champaign, 605 E. Springfield Ave., Champaign, IL 61820

Language Model Pretraining Next-Token Prediction Predictive Representation Learning Representation Geometry in Language Models

Aug 30, 2026cs.CL

The Depth Flow of Token Representations Is Nonlinear and Does Not Descend Its Own Density

A token's representation is carried through the network layer by layer. The whole vocabulary carried together forms a flow. We fit this flow's equation of motion as a discrete Langevin model over corpus-mean trajectories of Pythia-160M and Pythia-410M, and score the predicted steps on held-out tokens. Linear maps are often used as cheap surrogates for a layer. The flow they summarize is not linear: a quadratic drift beats the linear linear map at every transition of both models, and the Kramers--Moyal estimator agrees wherever its neighborhoods stay local. We then characterize the flow further. First, we show that it does not descend its own log-density. The drift instead descends a potential that is not the density. Second, the rotational component is not negligible,

4

to

45\%

of the explainable drift, and the circulation shows in what the flow preserves: a token keeps its angular rank across all thirteen layers while its norm rank is shuffled and its concentration rank is reversed by the last block.

Alexandre Quemy

Hother Labs

Dynamical Systems Representation Geometry in Language Models

Aug 30, 2026cs.CL

A Hub of Short Rows Inflates Intrinsic Dimension Estimation of Token Embeddings

A token-embedding table holds a hub of short rows near its origin, and we show that this cluster biases what nearest-neighbor intrinsic-dimension (ID) estimators report. Because of the concentration of measure, a token is closer to the central cluster than to any other token, so its first two neighbors are both hub rows at nearly the same distance. As a result, the ID estimators such as TwoNN return a dimension far above the real ID. Measured one token at a time, dimension is a heavy-tailed distribution. Measured on the full vocabulary, it grows with the model's parameter count. However, when we remove the hub, the heavy tail disappears and the measured dimension collapses to a narrow range for eleven models, from GPT-2 to models such as K3 and GLM-4.7. The hub acts as a switch: a few hundred rows are enough to fully inflate the estimate. We reproduced an experiment stating that the intrinsic dimension (ID) of Pythia's token-embedding table grows with the parameter count, from

27

to

122

between 160M and 12B parameters. We show that this result disappears when the hub is removed: the table then reads

10

to

17

at every size. The hub contains a subset of the population that under-trained-token detectors flag, but on Pythia the hub that we detected and removed as a whole was updated during training: what seem to characterize these rows is simply their length, not an absence of updates. Finally, we show that normalizing the rows instead of removing them gives the same lower reading.

Alexandre Quemy

Hother Labs

Intrinsic Dimensionality Neural Representation Geometry Representation Geometry in Language Models

Aug 27, 2026cs.CL

How Language Models Organize and Structure Moral Knowledge

How do large language models (LLMs) organize moral knowledge? Models detect moral content broadly, but detection is a low bar. We ask whether they go further, distinguishing moral foundations from one another and organizing the relationships between them geometrically. We train six independent linear probes on open-weight language models, one per Moral Foundations Theory (MFT) category (care/harm, fair/cheat, lib/oppress, loy/betray, auth/subv, sanc/degrade), and examine how the resulting directions relate to each other in representation space. We find the directions neither collapse into a single moral detector nor isolate from one another. Rather, they span a near-maximal number of independent dimensions while sharing a positive common component. The shared component is the signature of integration, and it is moral-specific relative to a matched non-moral concept battery built identically (mean pairwise cosine 0.26 vs. 0.013). The geometry is consistent across architectures and scale and reaches its integration regime early in pre-training, well before probe accuracy saturates. The structure the model discovers shows no evidence of the individualizing/binding distinction predicted by Moral Foundations Theory (an underpowered test: only 10 distinct splits exist, so it cannot reject at the 0.05 level) but rather reflects corpus statistics. Extending to moral dilemmas, each dilemma direction partially composes from its component foundations, at 2.7x a mismatched-pair baseline, while the majority of its variance encodes conflict-specific structure. The model represents moral tension itself, not a pre-resolved judgment.

Orion Reblitz-Richardson

Distiller Labs

Moral Reasoning Representation Geometry in Language Models LLM Interpretability

Aug 27, 2026cs.CL

Double Trouble: Bilingual Pretraining Leaves Language-Conditioned Effects in Shared-Language Representations

A concept can carry different associations across languages, while modern language models learn English alongside many other languages during pretraining. Yet comparisons among existing models cannot easily isolate how any one language changes the way these models represent English concepts because their training corpora, compute, architectures, and random seeds all differ. We study this question through a controlled experiment with 40 matched 310M-parameter decoder-only models that share an architecture, tokenizer, training recipe, and English data source. Each bilingual condition adds one of eight languages, while four experimental comparisons separately account for English exposure, total training, and English-document overlap. We align each model pair using 3,000 common English words, then measure where 1,000 held-out English concepts fall along 50 fixed semantic contrasts, such as red versus white. Across 32 experimental comparisons, English concept positions differ more between bilingual and English-only conditions than between English-only runs with different random seeds. These differences are larger in contextual states than in token embeddings and peak in middle layers. The language learned alongside English can therefore change how a model represents English concepts even when its English input representations are explicitly aligned.

Anjishnu Mukherjee, Ziwei Zhu, Antonios Anastasopoulos

George Mason University

Language Model Pretraining Multilingual Language Models Representation Geometry in Language Models

Aug 13, 2026cs.CL

Falsehood and Impossibility Are Different Directions in an AI's Representation of Language

Language can describe states of affairs that are false and states of affairs that could not be the case at all. Whether an AI model internally distinguishes these failures remains unclear. I report an exploratory activation study of the multimodal open-weight model Gemma 3 4B IT using 85 prompts from 17 philosophical families and a topic-matched modality set of 15 topics, each expressed as a truth, contingent falsehood, improbable claim, semantic anomaly, and necessary falsehood. In its answers, the model conflates contingent falsehood with contradiction, labeling 12 of 15 false statements "contradiction." Its activations show a different pattern. A linear truth probe separates impossible from true statements (AUC 0.93) but not impossible from false statements (AUC 0.20). An impossibility probe evaluated on held-out topic families separates necessary from contingent falsehood at AUC 1.00, peaking at layer 15 with balanced accuracy 0.97 (Bonferroni-adjusted P=0.018). The truth and impossibility directions are close to orthogonal, whereas the impossibility direction partially overlaps a semantic anomaly direction while remaining distinguishable from it. Sparse autoencoder features at the same layer repeat this geometry. Features selective for impossibility also fire on anomalous sentences but rarely on contingent falsehoods. In this model's activation space, necessary falsehoods are not extreme cases of contingent falsehood but lie closer to the experimentally defined category of semantic anomaly. This representational proximity does not imply that impossible statements are intrinsically meaningless. These correlational observations from one small model offer an empirical footnote to an old philosophical distinction.

Yoon Pyo Lee

Department of Nuclear, Plasma, and Radiological Engineering, University of Illinois Urbana-Champaign, Urbana, Illinois, USA.

Representation Geometry in Language Models

Aug 12, 2026cs.LG

Geometric and Behavioral Stratification in Transformer Residual Streams

Trained transformer models develop privileged bases: coordinate axes whose statistics differ from the rest of the residual stream. But what kind of direction does such a basis select? We investigate the prediction direction, the unembedding direction of the token a model currently predicts, and find that it functions as a content-defined privileged anchor. Measured with respect to this anchor, residual-stream variation is geometrically and behaviorally stratified by proximity to the prediction. The stratification holds in all eighteen models tested (dense and mixture-of-experts, 7B-120B, base and instruction-tuned). A narrow, scale-invariant prediction interface concentrates readout-relevant structure, while the vast prediction-distal complement expands with model scale. Because the prediction direction sits nearly orthogonal to the principal variance axes, variance-based analyses recover this organization only partly, and the shortfall grows with prompt heterogeneity. Anchoring reveals a steep geometric gradient: prediction-proximal regions are highly structured and cluster related prompts, while the complement is flatter and anti-discriminates among prompt groups. The interface is a narrow slice but functionally decisive. Disrupting the variance directions closest to the prediction causes immediate divergence and frequent task-frame shifts; disrupting the next level down delays divergence and preserves framing. The complement is weakly readout-aligned per direction yet causally and temporally load-bearing, and behavior is driven by direction rather than magnitude. These results establish the prediction direction as a privileged anchor distinct from previously described coordinate axes, and give a geometric account of how high-dimensional computation coexists with linear readout.

Nelson Guda

Transformer Interpretability Transformer Representation Geometry in Language Models

Aug 12, 2026cs.LG

Reading the Gate, Not the Interference: Output-Side Interference Measurement Does Not Track Merge Collapse

Task-arithmetic merging works until it doesn't, and the field diagnoses why by measuring interference inside the merged model. We take the most direct such measure, the exact layerwise activation cross-term of a factorial ledger, establish its causal anatomy, and then ask what it tracks. The anatomy is clean: each block mostly transports and amplifies the cross-term rather than generating it; erased, it is regenerated by the untouched marginal paths to 99% of its norm unless removed late; its output effect varies monotonically with the displacement's angle (orthogonal displacements make interference worse), and a two-assumption model derives the angle law and retro-dicts the dose curve (R^2 >= 0.99). What the measure tracks is not what the field assumes. Behavioural expert-likeness is decoupled from it across four instruments. Its cross-condition behaviour is denominator-dominated: an instruction template pins the main effect to within 1% while the absolute interaction grows 111x from two to six merged tasks, suppressing expressed interference at k=2 and amplifying it at k=6. And where merging actually collapses, the cross-term is a bystander, not the carrier: across two collapse parameterizations at two scales, even erased persistently at every position, removing it entirely repairs none of the collapse. There the output-side ratio carries no method information under a common counterfactual, while two state-space measures the field already uses rank methods correctly at both scales. All 81 predictions were frozen before their data; falsifications are reported as such. Output-side interference measurement reads the gate, the denominator, and the displacement budget, not the interference. What fails a merge is the carrier-bystander split: collapse rides in the marginal displacements while the cross-term merely accompanies it, and only state space sees the carrier.

Chencheng Zhu

UNSW Sydney

Task Vector Merging Task Arithmetic Language Model Merging Representation Geometry in Language Models

Aug 12, 2026cs.LG

GCPO: Diagnosing and Constraining Subspace Geometry in Rollout RL for LLMs

On-policy rollout methods such as GRPO are central to post-training of large language models, yet they frequently suffer from training instabilities, cross-task capability degradation, and response-length inflation. Although prior work has characterized the subspace geometry of aggregate updates, the stepwise variation of this geometry and its relationship to model performance remain unclear. We introduce Principal-Subspace Overlap, a dimension-corrected measure of individual rollout updates relative to the dominant singular subspaces of pretrained weights. Despite low average overlap, transient spikes often precede performance degradation. To address this, we propose GCPO (Geometrically Constrained Policy Optimization), which applies hard bilateral orthogonal projections to constrain updates to the complementary subspaces, preventing such excursions by construction. Across mathematical reasoning, code generation, and tool-use tasks on Qwen3-8B and GLM4-9B, GCPO consistently outperforms GRPO and recent variants, including DAPO and GSPO, improving over the base models and the strongest baseline by up to 27.69 and 2.37 points, respectively. Furthermore, GCPO preserves general capabilities, eliminates response-length inflation, and stabilizes policy entropy. Our findings provide a new diagnostic lens and a principled design perspective for stable reinforcement learning post-training.

Kai Yang, Jingwei Xu, Wanyu Wang +4

School of Computer Science, Shanghai Jiao Tong University · Shanghai Artificial Intelligence Laboratory · NovaCore +3

RL for Language Model Reasoning Constrained RL RL Post-Training Representation Geometry in Language Models

Aug 11, 2026cs.LG

Beyond a Bag of Features: Set-Level Instability in Sparse Autoencoders

Shani et al. (2026) show that LLM representations broadly recover human category boundaries, while failing to reflect fine-grained typicality structure. Their analysis uses cosine similarity over dense model representations. We revisit their approach using overlap over active sparse autoencoder (SAE) latent sets as a more interpretable similarity measure. We first verify that this set-level measure is meaningful: SAE latent sets can recover union-like compositional structure in controlled toy models and induce semantically coherent neighborhoods in natural text. Extending the human-concepts analysis to SAE set similarities, we find that SAE activation sets do not recover human category boundaries or within-category typicality more faithfully than dense embeddings or residual-stream states, but instead track model-internal similarity structure. To probe this gap further, we study active latent sets under well-controlled semantic modifications, revealing a substantial mismatch between human judgements of conceptual change and change in the SAE active set. We interpret this as evidence that, outside idealised settings, SAE features do not compose via simple bag-of-features semantics.

Nikolai Bolik, Lennart Stöpler, Artur Andrzejak

Heidelberg University

Representation Learning LLM Interpretability Sparse Autoencoders Representation Geometry in Language Models

Aug 11, 2026cs.LG

Mapping and Measuring the Behavioral Evolution of Large Language Models

Benchmark leaderboards summarize how well a language model performs, but not how its behavior relates to that of other models or changes across generations. We characterize the output behavior of 32 models from six families using their responses to a shared bank of 10{,}000 prompts. After embedding each response, we construct three complementary sentence-level dissimilarities: an aligned mean per-prompt distance, which is a pseudometric on observed model responses; a PCA-compressed summary of prompt-wise disagreement; and an alignment-free Gromov--Wasserstein discrepancy between models' internal response geometries. We use these constructions to study static organization and temporal change on a release-date axis through behavioral maps, family-wise drift, hierarchical clustering, cross-family convergence, and response-cloud dispersion. Across the three constructions, model families form coherent clusters, with \texttt{gpt-2} as a global outlier; cross-family distances decrease over time; and several recent reasoning-oriented models have comparatively compact response clouds. A token-level cross-check based on per-prompt Maximum Mean Discrepancy closely agrees with the sentence-level mean distance (Spearman

ρ=0.98

) and recovers the same qualitative findings. We organize these comparisons through a measure-theoretic lens making their alignment and invariance assumptions explicit. We also establish an architecture-agnostic sufficient condition linking behavioral similarity to inference-prompt coverage, small excess population log-loss, and similar effective target distributions---a possible training-side account rather than an empirical explanation of the observed trends. Our pipeline is label-free, and re-encoding every response with three further encoders---down to one

73\times

smaller---preserves the rank geometry, the outliers, and the sign of the time trend.

Dong Qiao, Chris Ding, Jicong Fan

School of Data Science, The Chinese University of Hong Kong, Shenzhen, China

LLM Evaluation Representation Geometry in Language Models Drifting Models

Aug 10, 2026cs.CL

Off-Axis, On Purpose: Where a Transformer Computes Concepts and Why it Does So

A transformer's answer lives on one axis: the direction its unembedding reads. Its intermediate states largely do not, and that off-axis position is usually treated as an obstacle to interpretation. We show it is functional. A 12-layer model computes in two phases. Through the first, every sublayer writes into a subspace held near-orthogonal to the read-out, attention 75 to 96 degrees off it at every depth. Moving attention's values onto the read-out is 64 to 84 times more damaging than a matched random rotation, and the damage is entirely in cross-token mixing: the subspace insulates composition from the vocabulary. Beneath it the frame itself turns rigidly with depth. In the second phase the answer arrives on-axis, late, and by addition rather than by turning accumulated content onto the read-out. Pressing every layer onto the read-out instead, as training for early exit does, matches the baseline on perplexity, LAMBADA and BLiMP while cutting the concept-phase workspace from about twenty-five effective dimensions to fourteen, a change none of those benchmarks register. The geometry can also be imposed, though not by asking for it. Prescribing it through the loss is a lottery: six of eight seeds collapse, because a model told to null its read-out projection obeys most cheaply by discarding dimensions. Inserting one fixed rotation at the phase boundary lands it instead, at baseline quality. A sparse rotation the surrounding weights can absorb converges on all nine seeds, against five of nine for ordinary training. Which rotation is immaterial: twenty-five runs across thirteen distinct ones reach the same quality, and two baselines from different seeds hold their concepts in near-orthogonal frames while agreeing on their read-outs. That freedom is usable: a basis drawn at random and prescribed before training is adopted across the concept phase, with quality unchanged.

Mark Oskin

School of Computer Science and Engineering University of Washington

Transformer Interpretability Transformer Attention Representation Geometry in Language Models

Aug 10, 2026cs.AI

One Adapter Pair per Model: A Universal Activation Interface for Language Models

Activation-based tools are usually tied to one model's native hidden space, requiring probes, sparse autoencoders, and natural-language interpreters to be rebuilt or rediscovered for each new language model. We present a Universal Activation Bus, a framework that provides a common activation interface across compatible language models. Using a small set of source models, we learn a shared dense space together with one lightweight linear encoder--decoder adapter pair per model. After source training, the interface is frozen; a new model joins by fitting only its adapter pair on unlabeled matched text. The resulting interface allows activation-based tools to be shared across connected models, including common probes and SAE features as well as access to an NLA originally trained for a different model. Across five models, semantically related texts form consistent neighborhoods in the shared space, and an onboarded model reuses these tools effectively without retraining them. We further show that an intermediate activation from one model can be used by another model's frozen upper layers to produce predictions. These results establish a stable, model-wise activation contract for reusable tools across compatible language models.

Su-Hyeon Kim, Jiwan Mun, Yo-Sub Han

Department of Artificial Intelligence, Yonsei University · Department of Computer Science, Yonsei University

Representation Geometry in Language Models Language Model Probing

Aug 9, 2026cs.AI

Reproducing and Stress-Testing Two Approaches to LLM Reasoning Reliability: Test-Time Probability Aggregation and Logic-Representation Editing

We independently reproduce two recent methods for making large language model (LLM) reasoning more reliable, and stress-test them across domains and models (RPC across four new task domains with Qwen3-8B, LCF across four 7-8B models). The first, RPC, aggregates token probabilities and self-consistency at inference; the second, LCF, trains projectors that split hidden states into "content" and "logic" and edits the logic part toward a valid region. Validating such reliability claims matters because the original evaluations are run by each method's own authors and were never independently reproduced or stress-tested across models and domains, and LCF shipped no public code. We re-run RPC's published-path aggregation and re-implement LCF's projector, contrastive, and intervention pipeline, then extend both to text-to-SQL, legal extraction, fallacy identification, and precedent grading, and probe LCF's representation directly. RPC reproduces the original grid exactly on the authors' released reasoning paths; on four new domains its edge over self-consistency is never significant (ties or small mixed differences, paired p >= 0.28), and on BIRD, the one domain where we vary the budget, the edge grows with K as predicted but its largest gap (+2.5 accuracy at K=32, p=0.16) reverses to -0.25 when we enlarge the sample to n=200. LCF's logic-validity direction is real but weak (0.82 separability at the single best sub-layer versus 0.95 for a semantic-attribute control); its one positive effect (Qwen3

Δ

Prob) is not significant (p=0.56), while it significantly reduces

Δ

Prob on two of the other three models.

Minhan Cho, Jimin Kweon

Remember & Company AI Lab · DSAIL, Dept. of Applied AI · Sungkyunkwan University +2

LLM Evaluation Reasoning Consistency in Language Models LLM Reliability Representation Geometry in Language Models LLM Reasoning

Aug 9, 2026cs.AI

HoloAegis: Frozen Representation, Topological Inference --- Minimally Parametric Safety Manifolds and Their Capability Boundaries for LLM Guardrails

Current LLM safety guardrails face a fundamental tension: fine-tuning distorts pre-trained representations while generative judges incur prohibitive inference costs. We ask a complementary question: how far can safety be achieved through pure geometric reasoning over frozen representations, and where does it fail? We present HoloAegis, a minimally parametric topological inference framework that decouples representation from reasoning: an un-fine-tuned encoder maps text to the unit sphere S^{d-1}, and all decisions reduce to Gibbs-Boltzmann free-energy differences over pre-computed anchor centroids. We contribute a boundary-mapping study rather than a leaderboard claim. On a frozen three-benchmark protocol, HoloAegis (3.2 MB) statistically matches WildGuard-7B (14 GB) on toxicity (0.96 vs. 0.96), exceeds it on harmful behaviors (0.99 vs. 0.79), and cedes oversafety detection (0.62 vs. 0.98) -- while ShieldGemma-2B fails on indirect harms (0.34). These failure modes are complementary and mechanistically traceable: potential-difference scoring senses manifold clustering, whereas policy-conditioned LLM judging requires explicit taxonomy matching. We restate our Topological Boundary Stability conjecture in ratio form and validate it via reference-set bootstrap: anchor banks reduce score variance 4-15x and boundary displacement to approximately 0.44 + 0.23 sqrt(k/K) of the full-space estimator. Per-domain analysis further reveals that geometric separability tracks within-domain semantic homogeneity. Our results chart where geometric guardrails substitute for, and where they must defer to, LLM judges.

Tak Ho Alex Li, Kaijie Liu, Lik-Hang Lee +3

Department of Mathematics, Hong Kong Baptist University, Hong Kong 999077 · Guangdong-Hong Kong-Macao Institute of ESG and New Quality Productive Forces, School of Optoelectronic Engineering, Guangdong Polytechnic Normal University, Guangzhou 510665 · Guangdong Institute of Digital Industry, Guangzhou 510000 +3

Language Model Safety Evaluation LLM Guardrails LLM Safety Representation Geometry in Language Models

Aug 7, 2026cs.IR

Preserving Item Semantics for Free: Rethinking Token Initialization in LLM-Based Generative Recommendation

Recent advances in generative recommendation (GR) leverage large language models (LLMs) as recommender backbones, enabling LLMs to directly generate recommendations conditioned on item-interaction histories. In these systems, items are often represented through semantic IDs (SIDs) added to the LLM vocabulary as special tokens. Ideally, SIDs imbue item token representations with semantic priors, thereby improving model generalization. However, standard vocabulary expansion typically initializes these tokens as random Gaussian vectors, discarding the SIDs' underlying continuous geometry and forcing the LLM to relearn token relationships from interaction data. To demonstrate the consequences of this design, we first show that training from this initialization tends to organize SID embeddings around item popularity rather than semantics. We further show that, despite partially reducing the reliance on popularity and improving cold item performance, the computationally expensive process of continual pretraining (CPT) fails to reliably recover the original semantic geometry. To address these findings, we propose a simple, parameter-free intervention that initializes SID token embeddings directly from their corresponding centroids in the semantic embedding space. Requiring only a few lines of code and no additional training or inference overhead, this drop-in approach improves pure-SFT Recall@5 by up to 16%, reaches peak performance with up to 40% fewer SFT steps, and improves cold-item Recall@5 by up to 60%. Moreover, on datasets that benefit from additional CPT, centroid initialization reaches comparable performance while requiring half as many CPT epochs. Together, our findings show that preserving SID geometry, beyond shared-prefix structure, provides a simple and effective semantic prior for LLM-based GR.

Donald Loveland, Liam Collins, Bhuvesh Kumar +2

Snap Inc. Bellevue, WA, USA · University of Michigan Ann Arbor, MI, USA

Large Language Model-Based Recommendation Representation Geometry in Language Models Generative Recommendation

Aug 7, 2026cs.AI

Transformers Struggle to Use Their Emergent World Models: Revisiting the Tower of Hanoi, and the Illusion of Thinking

The Tower of Hanoi is a simple planning puzzle that in prior work has proven challenging for large reasoning models (LRMs). Current models solve the standard formulation of the puzzle, but still struggle with the flat-to-flat variant (where initial and goal states are not restricted to have all rings on a single peg). This paper presents an in-depth study of how both small, in-house Transformers and large, third-party LRMs solve this task. To understand the failures mechanistically, we first train small Transformers from scratch on precomputed solution traces. Using a variety of interpretability techniques, we show that these Transformers develop an emergent world model: a linearly decodable, geometrically faithful representation of the puzzle's state space (the Sierpinski triangle), that is causally involved in solving the puzzles. Second, we return to the large LLMs and apply our techniques to two frontier reasoning models, Qwen3.6-27B and DeepSeek-R1-Distill-Qwen-32B, that attempt to solve the task through extended chain-of-thought. Surprisingly, we find that both models encode the Sierpinski world model near-perfectly at the end of the prompt, and yet fail at the majority of tasks when there are more than 3 rings. We locate the source of this failure in the decaying representation of the world model. We probe for the representation at different stages during planning, and establish causality by showing that performance can be improved by injecting the prompt-time representation at inference. The failure of the models is thus one of maintenance of the required representations, not their absence, and performance is at least partially recoverable. These results thus reframe the reported collapse in performance from prior work: current Large Reasoning Models build a world model, and then lose it.

Devin Pereira, Willem Zuidema

1Artificial Intelligence Program, University of Amsterdam · ELLIS Unit Amsterdam · Institute for Logic Language and Computation, University of Amsterdam

LLM World Models LLM Planning World Models World Model-Based Planning Representation Geometry in Language Models Large Reasoning Models

Aug 6, 2026cs.CL

Lost in Interpolation: Why Predictive Feedback Fails in Diffusion Language Models

Soft-masking accelerates the convergence of Masked Diffusion Language Models (MDLMs). Existing formulations build this blend with linear interpolation (LERP) in the raw embedding space, which implicitly treats that space as Euclidean. We analyze the embedding space of MDLMs and find that the mask and predicted-token embeddings maintain a near-constant angle of (\approx 73^\circ) throughout training, while embedding norms remain essentially flat across vocabulary-frequency rank. These indicate a hyperspherical geometry, for which LERP is the wrong interpolation primitive. We introduce Spherical Soft-Masking (S-SM), a drop-in replacement that aggregates the top-(k) predictions with a Fr'echet mean on the hypersphere and blends this mean with the mask direction using spherical linear interpolation (SLERP), then restores the native mask norm. We evaluate S-SM on continued pre-training of a released 169M-parameter MDLM checkpoint across a wide range of inference-time step budgets, SLERP feedback avoids the training degradation that LERP feedback induces and delivers MAUVE gains of up to 2x over the vanilla MDLM baseline and 27.5-56.1% over TopK/LERP at various sampling budgets, alongside consistently lower generative perplexity (16.9-19.6% over the baseline), while leaving output entropy and convergence essentially unchanged.

Lavanya Nigam, Ishaan Bansal, Aryan Sood +2

Indian Institute of Technology Roorkee, Roorkee, Uttarakhand, India

Masked Diffusion Models Diffusion Language Model Decoding Diffusion Language Models Representation Geometry in Language Models

Aug 6, 2026cs.LG

Reasoning Errors Have a Region and a Direction in the Residual-Stream Trajectory of LLMs

As language models are increasingly used for tasks that require verifiable reasoning, reliably distinguishing sound reasoning from flawed reasoning has become an important practical problem. Recent trajectory-based methods seek this signal in layerwise residual-stream displacements, which capture how representations change while attenuating some stable, token-specific information. However, displacement omits the state from which an update originates, whereas restoring the full state risks reintroducing shortcut-prone information. We identify this trade-off and propose a three-stream detector that combines motion with two restricted views of location. A coarse region reader based on vector quantization and a fine direction reader over normalized multi-layer states. This design restores enough state context to interpret the motion without returning to full-state probing. On reasoning benchmarks unseen during training, our method improves selection accuracy by up to 12% over the displacement-only state of the art and 21% over single-layer probing baselines. Although trained only on reasoning benchmarks, it also reads factual completion and fact verification, ahead of every detector we compare against, which places the signal on correctness rather than on a kind of reasoning. Ablations further show that motion, region, and direction provide complementary signals. These results suggest that reasoning validity is better read from state-conditioned motion than from either static states or decontextualized trajectories alone.

Hamed Damirchi, Ignacio Meza De la Jara, Damith Ranasinghe +2

Australian Institute for Machine Learning · Adelaide University · Naval Group Pacific +1

Language Model Error Detection LLM Interpretability LLM Reliability Representation Geometry in Language Models LLM Reasoning

Aug 6, 2026cs.CL

Scaling Inherently Interpretable Language Models

Interpretability is often treated as a tax on capability: language models are trained as opaque systems, then explained after the fact, with methods whose reliability is difficult to establish. In this work, we challenge this premise. Rather than reverse-engineering a model, we make interpretability a constraint of the training pipeline, optimized alongside the language modeling objective. Across three orders of magnitude of compute, on both autoregressive and diffusion language models, interpretability scales with capability rather than against it. Surprisingly, model representations become more disentangled and aligned with human-understandable concepts with scale. We instantiate the training-time recipe with Steerling-8B, a diffusion language model with a causal attention mask. For any group of generated tokens, Steerling-8B attributes the output to relevant input tokens, human-understandable concepts, and training data. This enables closed-loop intervention: diagnose an output through its concept or feature attribution, retrieve similar training data, and correct the behavior through concept steering without retraining. Steerling-8B remains competitive with open peer models trained on substantially 2-16x more compute, suggesting a different scaling paradigm: interpretability can be designed into training, and it improves with scale.

Guide Labs Team, Andreas Madsen, Aya Abdelsalam Ismail +7

Guide Labs Team

Language Model Steering LLM Interpretability Diffusion Language Models Representation Geometry in Language Models LLM Training

Aug 4, 2026cs.CL

Logic Before Language: Pre-pretraining on Formal Derivations Fosters Skill Acquisition and Compressibility

Pre-pretraining language models (LMs) on symbolic data can accelerate and improve natural language acquisition. However, existing pre-pretraining tasks, such as Dyck and procedural algorithms, rely on narrow primitives that fail to capture the expressive capacity of natural language. Moreover, prior studies remain restricted to relatively small token budgets, offering limited insight into skill emergence and representational dynamics. To address these limitations, we propose logic pre-pretraining (Logic-PPT) as a principled initialization strategy, leveraging formal derivations to impart richer structural and linguistic biases. Formal derivations require abstract mechanisms that are central to natural language, simultaneously binding variables, connecting quantifiers and relational dependencies, and composing predicate-argument structures over long contexts. Scaling our evaluation to a 100B-token regime, logic pre-pretraining substantially accelerates skill acquisition in LMs, achieving 80% accuracy on linguistic tasks with 36B fewer tokens than standard initialization, and outperforming alternative pre-pretraining baselines. Mechanistically, formal derivations induce persistent structural reorganization, distinctively characterized by a lower-rank, spectrally concentrated representation space. Crucially, we show that this internal geometry enables improved model compressibility via pruning, matching the dense baseline performance even at

\approx

33% sparsity.

Jo-Ku Cheng, Nikolaos Aletras, Marco Valentino

School of Computer Science, University of Sheffield, United Kingdom

Language Model Pretraining Representation Geometry in Language Models

Aug 4, 2026cs.LG

The Ignition Is Real, and It Lives at the Readout: Latent composition, difficulty-clocked ignition, and the interface-constituted commit in a recurrent-depth reasoner

We test whether the "compositional ignition" reported in latent-reasoning models is real computation, an instrument artifact, or inherited from verbal training data. We grow an independent realization of a published 30M-parameter recurrent-depth reasoner from scratch (same recipe and seed), film its development, certify fidelity through a pre-registered whole-signature gate, and measure resolution in two channels at once: the vocabulary readout and the hidden state. The ignition is real and lives at the readout: arrival time rises lawfully with problem depth, resolution is sharp and holds, and the signature reproduces across two same-seed realizations with divergent training trajectories. At commitment the decision margin jumps 5.8-8.0 logits in one iteration, exceeding the 90th percentile of near-threshold non-event steps in 96% of cases; the signed margin's zero-crossing there is definitional and carries no evidential weight, so the evidence is that conditioned magnitude. The hidden-state direction snaps in raw geometry, meeting its pre-registered criterion (in the decoder's LayerNorm coordinates it attenuates just below our bar, so the composite decoder-coordinate claim is not confirmed), and then freezes in both (descriptively so in decoder coordinates; angular steps 52.9 to 1.2 degrees over eight iterations), while subsequent displacement is predominantly radial (0.961 of squared-norm) and readout-null to a measured bound (radial logit effect <=5.7e-6). An earlier velocity-trough claim is withdrawn: pre-registered normalization controls showed it coordinate-dependent. Intermediates were never recoverable through the tied readout (relay 0.00). All criteria were frozen before their data; the predictions ledger, including this paper's own withdrawn headline, ships in the companion repository.

Simon Lam-Muir

Prime Calibre, Australia

Compositional Reasoning Representation Geometry in Language Models Latent Reasoning

Aug 4, 2026cs.CL

Language Models Encode the Contextual Truth of Propositions

Prior work has shown that LLMs encode the truth of factual propositions along linear directions in activation space. It's unclear how these representations extend to contextual truth: propositions whose truth is determined by in-context evidence rather than world knowledge. We show that LLMs maintain a linear representation of contextual truth that persists across structurally different output policies, even when the output doesn't require the model to determine a proposition's truth, and show causal evidence via steering experiments. Using the transcripts from a collaborative vision-language task that requires two LLMs to maintain a shared common ground, we show that truth representations of a proposition are significantly swayed by partner assertions about that proposition, even when the LLM has enough evidence to determine its truth. We find evidence that propositions near the decision boundary are more susceptible to having their truth shifted through partner assertions. Separating representation from output distinguish two forms of sycophancy that output behavior alone cannot: the model may accommodate a false proposition while continuing to represent it as false, or shift its representation across the boundary. The latter is 2.59x more common when the model agrees by restating the false claim explicitly than when it agrees implicitly.

Rupak Sarkar, Pritika Ramu, Rachel Rudinger

University of Maryland, College Park

LLM Sycophancy Knowledge Conflicts in Language Models Factual Knowledge in Language Models Representation Geometry in Language Models

Aug 4, 2026cs.LG

A Graph Signal Processing Perspective on Numerical Sequence Representations in LLM In-Context Learning

Pretrained large language models (LLMs) have demonstrated in-context learning (ICL) capabilities for numerical inference over sequences serialized as text. Prior work has identified and characterized this form of numerical inference primarily through output-level evaluations such as prediction error. However, how numerical information is organized within LLM representations remains much less understood. To study this internal organization, we adopt a graph signal processing perspective in which attention induces a weighted graph over tokens, while token hidden states define signals on its nodes. Quantitative graph-spectral diagnostics and qualitative token-graph visualizations reveal that representations become more clearly differentiated by input dynamical complexity as context length increases. Simpler inputs produce attention-induced token graphs with stronger global connectivity and smoother, spectrally concentrated hidden-state signals, whereas more complex inputs produce more localized graphs and hidden-state signals with broader spectral support and greater high-frequency energy. Together, these findings point to systematic, context-dependent internal signatures associated with numerical ICL that are conserved across model families.

Jiajun Bao, Zihao Qi, Toni J. B. Liu +4

1Cornell University · 2Goodfire AI · 3Imperial College London

Numerical Reasoning in Language Models In-Context Learning Representation Geometry in Language Models

Aug 3, 2026cs.LG

ChaosProbe: A Neurochaotic Lens on Frozen Transformer Input-Embedding Spaces

Transformer models are most often understood through what they do: their benchmark performance, generation quality, or behavior on downstream tasks. Yet frozen transformer input-embedding spaces may also be examined through their responses to a controlled deterministic probe before contextual computation or task-specific adaptation. Guided by this response-based view, we introduce ChaosProbe, a deterministic neurochaos-inspired method for constructing response-based fingerprints of frozen transformer input-embedding spaces. For each prompt-level embedding matrix, ChaosProbe applies a chaotic trajectory-based transformation and summarizes its Firing Rate and Entropy channel responses with complementary representation-level measures, producing a fixed-length signature for each model. In a bounded proof-of-concept study of

80

neutral prompts and four pretrained models---GPT-2, DistilGPT2, BERT-base-uncased, and RoBERTa-base---Pearson correlation, Spearman correlation, and cosine similarity each recover all four same-family nearest-neighbor assignments and both expected mutual family pairs. Euclidean distance recovers three of the four assignments and one of the two mutual family pairs. Paired bootstrap resampling supports the stability of the Pearson and Spearman pairings over the observed prompt set, and signature-validity checks show that constant or collapsed responses do not dominate the reported fingerprints. These results provide a cohort-dependent proof of concept that deterministic neurochaotic response signatures can expose broad structure among frozen transformer input-embedding spaces.

Kunal Kumar Pant, Nithin Nagaraj

Department of Electronics and Communication Engineering (ECE) Amrita Vishwa Vidyapeetham Kollam, Kerala, India · Complex Systems Programme National Institute of Advanced Studies (NIAS) IISc Campus, Bengaluru Karnataka, India

Language Model Fingerprinting Representation Probing Representation Geometry in Language Models Language Model Probing

Aug 3, 2026cs.AI

Rewriting or Reweighting? A Geometric Account in Language Models

Post-training can substantially alter language-model behavior, yet aggregate behavior rates do not reveal whether training removes an existing mechanism, creates a new one, or changes how an inherited mechanism is used. We study this question through two mechanistically distinct failures, repetition as a decoding-attractor pathology and sycophancy as a preference-related alignment failure. We introduce behavioral manifold analysis, which isolates behavior-specific geometry by selecting sparse behavior-associated coordinates and lifting them into low-dimensional local charts. We construct these charts in two complementary spaces. ACT captures runtime activation states, while NOC quantifies how strongly the model routes functional information flow through the shared behavior-associated subspace. Across multiple model families, the resulting charts are highly compressed and partially alignable across architectures. Contribution-space charts expose a more architecture-robust shared core, whereas activation-space charts retain stronger family-specific structure. Tracking these charts through controlled post-training reveals a consistent asymmetry. Supervised fine-tuning substantially alters the inherited behavioral geometry, whereas reward optimization changes behavior while largely preserving the underlying chart. This geometric perspective provides a unified framework for understanding the mechanistic distinction between the two objectives. SFT tends to rewrite behavioral geometry, whereas reward optimization primarily reweights it. Code is available at https://github.com/ronglingze/Manifold-Analysis

Juntong Wang, Shengkun Yang, Xiyuan Wang +1

Institute for Artificial Intelligence, Peking University, Beijing, China · School of Intelligence Science and Technology, Peking University, Beijing, China · School of Computer Science and Technology, Beijing Jiaotong University, Beijing, China

Language Model Post-Training Representation Geometry in Language Models Data Repetition in Language Model Training

Aug 3, 2026cs.CL

Divergent large language model predictions from convergent representations in ambiguous word pairs

In this work we investigate how decoder-only transformers resolve lexical ambiguity through layer-by-layer analysis of three models spanning three parameter sizes (GPT-2-Small-117M, Llama-3.2-3B, Qwen2.5-32B). For both homonyms and polysemes, we find that representations become maximally distinct in middle layers, then partially reconverge in late layers, while the KL divergence between their next-token predictions reaches its maximum in the final layers. The activation patching experiment provides causal evidence that late-layer representational differences directly determine outputs despite apparent increased similarity in embedding space. Our single-layer ablation experiment indicates that models achieve equivalent disambiguation despite qualitatively different layer-wise vulnerabilities. These findings offer a mechanism for recent observations where models' internal embedding similarities show low correlation with their behavioural outputs despite strong performance. The semantic distinctions therefore remain present but become increasingly invisible to similarity measures over the embeddings, with implications for embedding-based methods such as semantic search, retrieval, and clustering that rely on late-layer cosine similarity.

K. Jack Scott, Narun Pat, Veronica Liesaputra

Department of Psychology, University of Otago, Dunedin, New Zealand · School of Computing, University of Otago, Dunedin, New Zealand

Transformer Interpretability Decoder-Only Language Models Representation Geometry in Language Models