Representational Probing

• Representational probing is a suite of methods that evaluates the information encoded in fixed neural representations using linear probes, RSA, and intervention techniques.
• Techniques include training simple classifiers on frozen embeddings, computing dissimilarity matrices, and using control tasks to distinguish genuine informational content.
• Applications span NLP, vision, audio, reinforcement learning, and biomedical imaging, demonstrating its value in assessing model behavior and alignment with external structures.

to=arxiv_search 和天天中彩票json {"^^^^^^^^8query8 probing arXiv RSA probing representations probe diagnostic classification framework8^^^^^^^^", "^^^^^^^^8max_results8 ^^^^8representational probing arXiv RSA probing representations probe diagnostic classification framework8query8^^^^} to=arxiv_search 大发快三官网json {"^^^^^^^^8query8 My Representation Capture X? Probe-Ably^^^^^^^^8\8 OR ^^^^^^^^8\8 OR ^^^^^^^^8\8 as Quantifying Inductive Bias^^^^^^^^8\8 OR ^^^^^^^^8\8 Similarity Analysis^^^^^^^^8\8 "^^^^^^^^8max_results8 ^^^^8representational probing arXiv RSA probing representations probe diagnostic classification framework8query8^^^^} to=arxiv_search 彩神争霸官方ోజన్ {"^^^^^^^^8query8 &&&^^^^8representational probing arXiv RSA probing representations probe diagnostic classification framework8^^^^&&&, &&&^^^^8max_results8^^^^&&&, &&&^^^^8query8^^^^&&&, &&&^^^^8\8^^^^&&&, &&&^^^^8 OR \8^^^^&&&)", "^^^^^^^^8max_results8 ^^^^8representational probing arXiv RSA probing representations probe diagnostic classification framework8query8^^^^} Representational probing is a family of methods for characterizing what information is encoded in learned neural representations. In the most common formulation, an encoder is frozen, representations are extracted, and a lightweight external model is trained to predict a target property from those representations; related formulations compare representational geometries through representational dissimilarity matrices, quantify the structure of version spaces without training classifiers, or intervene directly on latent states to test causal relevance. Across NLP, vision, audio, reinforcement learning, computational neuroscience, and biomedical imaging, representational probing is used to ask whether properties are linearly decodable, geometrically organized, statistically aligned with external structure, or behaviorally active under intervention (&&&^^^^8max_results8^^^^&&&, &&&^^^^8query8^^^^&&&, Lepori et al., 2020, Chen et al., 4 Jun 2026).

^^^^8representational probing arXiv RSA probing representations probe diagnostic classification framework8^^^^. Formal definition and analytical target

A standard probing setup begins with a fixed representation map. If PRESERVED_PLACEHOLDER_^^^^8query8^^^^ is the embedding of an input PRESERVED_PLACEHOLDER_^^^^8representational probing arXiv RSA probing representations probe diagnostic classification framework8^^^^, a probing classifier or regressor predicts a property PRESERVED_PLACEHOLDER_^^^^8max_results8^^^^ from PRESERVED_PLACEHOLDER_^^^^8query8^^^^ without back-propagating into PRESERVED_PLACEHOLDER_^^^^8\8^^^^. In document-level event extraction, this is written as a probe PRESERVED_PLACEHOLDER_^^^^8 OR \8^^^^ operating on frozen embeddings, with a linear form PRESERVED_PLACEHOLDER_^^^^8 OR \8^^^^ or a two-layer MLP PRESERVED_PLACEHOLDER_^^^^8 OR \8^^^^ trained by cross-entropy on probing examples $\{(z_i,y_i)\}$ (&&&^^^^8max_results8^^^^&&&). Probe-Ably presents the same basic abstraction as $g_\theta:\mathcal{R}\rightarrow Y$ over a fixed representation space PRESERVED_PLACEHOLDER_^^^^8representational probing arXiv RSA probing representations probe diagnostic classification framework8query8^^^^, again emphasizing that the object of study is the representation rather than the end-to-end model (&&&^^^^8query8^^^^&&&).

This formulation is deliberately narrow: the probe is meant to measure what is extractable from a representation, not to improve the representation. In reinforcement learning, for example, a frozen encoder PRESERVED_PLACEHOLDER_^^^^8representational probing arXiv RSA probing representations probe diagnostic classification framework8representational probing arXiv RSA probing representations probe diagnostic classification framework8^^^^ is evaluated by training only a linear head for reward prediction or expert-action classification; the encoder remains fixed throughout, and the probe serves as a low-variance surrogate for expensive downstream RL evaluation (&&&^^^^8representational probing arXiv RSA probing representations probe diagnostic classification framework8max_results8^^^^&&&). In spatial audio, frozen scene embeddings are probed with a simple linear head to predict azimuth, elevation, distance, event class, RT^^^^8 OR \8query8^^^^, room volume, and room shape, making the benchmark explicitly about representational content rather than full task learning (Chen et al., 4 Jun 2026).

The target of probing varies by domain. Surface-level probes test counts or lengths; semantic probes test coreference, argument detection, or category information; event-level probes test template structure; cross-modal probes test whether language embeddings align with visual regions; neuroscientific probes test whether model geometry resembles fMRI or EEG structure (&&&^^^^8max_results8^^^^&&&, &&&^^^^8representational probing arXiv RSA probing representations probe diagnostic classification framework8 OR \8^^^^&&&, &&&^^^^8query8^^^^&&&, &&&^^^^8representational probing arXiv RSA probing representations probe diagnostic classification framework8 OR \8^^^^&&&). This diversity of targets has led to a broad notion of representational probing: any method that systematically evaluates what a fixed latent space contains about a specified variable, structure, or relation.

^^^^8max_results8^^^^. Decoder-based probing and its variants

The classical probe is a simple supervised predictor trained on frozen features. In NLP, linear probes and small MLPs are standard because they keep the readout simple enough that positive results can be interpreted as evidence that the relevant information is accessible in the embedding. Document-level event extraction uses eight such probes, spanning Word Count, Sentence Count, Coreference, Argument Detection, Argument Typing, Event Count, Co-Event, and Event Typing, with classification accuracy averaged over five random seeds (&&&^^^^8max_results8^^^^&&&). Similar linear probing is used in graph representation analysis, where a single affine map with softmax or linear regression is trained on frozen graph embeddings and evaluated with ROC-AUC or PRESERVED_PLACEHOLDER_^^^^8representational probing arXiv RSA probing representations probe diagnostic classification framework8max_results8^^^^ (&&&^^^^8representational probing arXiv RSA probing representations probe diagnostic classification framework8^^^^9&&&).

Some probe families are specialized to structured outputs. DepProbe learns two linear subspaces from frozen contextual embeddings: a structure subspace PRESERVED_PLACEHOLDER_^^^^8representational probing arXiv RSA probing representations probe diagnostic classification framework8query8^^^^ that scores dependency proximity via PRESERVED_PLACEHOLDER_^^^^8representational probing arXiv RSA probing representations probe diagnostic classification framework8\8^^^^, and a relation subspace PRESERVED_PLACEHOLDER_^^^^8representational probing arXiv RSA probing representations probe diagnostic classification framework8 OR \8^^^^ that predicts dependency labels with a token-level softmax. Combined with an PRESERVED_PLACEHOLDER_^^^^8representational probing arXiv RSA probing representations probe diagnostic classification framework8 OR \8^^^^ tree-building algorithm, this yields directed, labeled dependency trees using approximately PRESERVED_PLACEHOLDER_^^^^8representational probing arXiv RSA probing representations probe diagnostic classification framework8 OR \8^^^^ parameters, and the resulting probe identifies the best source treebank for transfer parsing PRESERVED_PLACEHOLDER_^^^^8representational probing arXiv RSA probing representations probe diagnostic classification framework88^^^^ of the time across ^^^^8representational probing arXiv RSA probing representations probe diagnostic classification framework8query8^^^^ languages (&&&^^^^8max_results8query8^^^^&&&).

Other variants change the direction of the prediction problem. Quantized reverse probing measures the mutual information between concept labels and a quantized representation PRESERVED_PLACEHOLDER_^^^^8representational probing arXiv RSA probing representations probe diagnostic classification framework8^^^^9, estimating PRESERVED_PLACEHOLDER_^^^^8max_results8query8^^^^ through K-means clustering and a reverse linear probe from concept space to cluster indices (&&&^^^^8max_results8representational probing arXiv RSA probing representations probe diagnostic classification framework8^^^^&&&). Sparse linear probing, as used for binomial ordering preferences, fits a Lasso regressor to predict a continuous preference-strength target PRESERVED_PLACEHOLDER_^^^^8max_results8representational probing arXiv RSA probing representations probe diagnostic classification framework8^^^^ from z-score standardized hidden states, revealing whether a behavioral statistic is encoded in a low-dimensional subspace (&&&^^^^8max_results8max_results8^^^^&&&). Sparse autoencoders constitute another probe-like family: they reconstruct activations through a sparse bottleneck and are then evaluated for semantic meaning, out-of-distribution generalization, ontology recovery, and controllable generation (&&&^^^^8max_results8query8^^^^&&&).

A recurring pattern is that the probe need not match the downstream task. RL reward probes and action probes are intentionally generic; common-ground probing maps contextual noun embeddings to visual region features with an InfoNCE objective rather than a linguistic label; computational pathology probing compares whole representational spaces across foundation models rather than predicting pathology endpoints directly (&&&^^^^8representational probing arXiv RSA probing representations probe diagnostic classification framework8max_results8^^^^&&&, &&&^^^^8representational probing arXiv RSA probing representations probe diagnostic classification framework8 OR \8^^^^&&&, &&&^^^^8max_results8 OR \8^^^^&&&). Representational probing is therefore broader than auxiliary classification alone.

^^^^8query8^^^^. Representational similarity analysis and geometric probing

Representational Similarity Analysis (RSA) treats probing as a second-order comparison between geometries. Given a set of items and their embeddings, one computes a representational dissimilarity matrix (RDM), often using

PRESERVED_PLACEHOLDER_^^^^8max_results8max_results8^^^^

and then compares vectorized off-diagonal entries of two RDMs with a similarity statistic such as Spearman’s PRESERVED_PLACEHOLDER_^^^^8max_results8query8^^^^ (Lepori et al., 2020). In the color-qualia benchmark, nine colors produce a PRESERVED_PLACEHOLDER_^^^^8max_results8\8^^^^ representational similarity matrix, which is converted to an RDM by PRESERVED_PLACEHOLDER_^^^^8max_results8 OR \8^^^^ and compared across brain and model systems with

PRESERVED_PLACEHOLDER_^^^^8max_results8 OR \8^^^^

The same study also tests a Gaussian squared-Euclidean control distance (&&&^^^^8query8^^^^&&&).

RSA is especially useful when the systems being compared differ in dimension, scale, or modality. It has been used to compare BERT verb and pronoun embeddings with hypothesis models for subject and antecedent sensitivity, to compare CodeBERT code representations with BERT representations of natural-language descriptions as an intrinsic measure of semantic grounding, to compare computational pathology foundation models over H^^^^^^^^8\8 patches, and to compare Audio LLM layers with EEG over time (Lepori et al., 2020, &&&^^^^8query8query8^^^^&&&, &&&^^^^8max_results8 OR \8^^^^&&&, &&&^^^^8representational probing arXiv RSA probing representations probe diagnostic classification framework8 OR \8^^^^&&&). This model-agnostic property is one of RSA’s main attractions.

Several studies use RSA to dissociate different kinds of representational structure. In color qualia, almost all tested models align better with the “no-report” condition than with the “report” condition, with approximately PRESERVED_PLACEHOLDER_^^^^8max_results8 OR \8^^^^ of models showing higher alignment to the no-report RDM and mean scores of PRESERVED_PLACEHOLDER_^^^^8max_results88^^^^ versus PRESERVED_PLACEHOLDER_^^^^8max_results8^^^^9 (&&&^^^^8query8^^^^&&&). In code models, pre-training alone yields RS values near PRESERVED_PLACEHOLDER_^^^^8query8query8^^^^–PRESERVED_PLACEHOLDER_^^^^8query8representational probing arXiv RSA probing representations probe diagnostic classification framework8^^^^, whereas semantic fine-tuning can raise deep-layer RSA substantially, with bimodal NL+PL inputs outperforming unimodal code-only inputs (&&&^^^^8query8query8^^^^&&&). In Audio LLMs, layer-wise EEG alignment depends strongly on the metric: rank-based measures such as Spearman RSA and Kendall’s PRESERVED_PLACEHOLDER_^^^^8query8max_results8^^^^ need not agree with dependence-based measures such as distance correlation, RV, or CKA, producing what the study terms a rank-dependence split (&&&^^^^8representational probing arXiv RSA probing representations probe diagnostic classification framework8 OR \8^^^^&&&).

Geometric probing also includes non-RSA methods. DirectProbe analyzes the geometry of an embedding space through the version-space perspective, clustering examples into label-pure convex-hull cells and measuring margins between clusters of different labels. Its central claim is that representational analysis need not always proceed through a trained classifier; the geometry itself can anticipate classifier performance (&&&^^^^8 OR \8^^^^&&&). In a related but more intervention-oriented direction, the Manifold Probe generalizes linear regression probes by jointly learning feature functions and directions that parameterize a low-dimensional concept manifold embedded in a high-dimensional representation (&&&^^^^8query8 OR \8^^^^&&&).

^^^^8\8^^^^. Methodological controls, caveats, and competing interpretations

A central criticism of naive probing is that auxiliary-task accuracy may reflect probe expressivity rather than genuine encoding. Probe-Ably summarizes this concern explicitly: increasing probe complexity can raise auxiliary accuracy even on random labels, so reliable probing should include control tasks, varying probe complexity, selectivity, and optionally information-theoretic measures such as minimum description length (&&&^^^^8query8^^^^&&&). In that framework, selectivity is defined as

PRESERVED_PLACEHOLDER_^^^^8query8query8^^^^

where the control task uses random labels (&&&^^^^8query8^^^^&&&).

A stronger critique reframes the problem itself. “Probing as Quantifying Inductive Bias” argues that probing should measure the inductive bias a representation provides for a task, not simply how separable labels are under a chosen probe. The proposed Bayesian framework compares representations through marginal likelihood, defining

PRESERVED_PLACEHOLDER_^^^^8query8\8^^^^

Within that framework, random or uninformative representations receive low evidence because no probe can generalize well, while overly flexible probes are penalized by Occam’s razor through the evidence term (&&&^^^^8\8^^^^&&&).

Classifier-free alternatives pursue the same goal from a geometric angle. DirectProbe studies version spaces, cluster counts, and margin distributions, arguing that the number of disjoint label regions and their separation can reveal whether a representation supports linear or nonlinear readout before any classifier is trained (&&&^^^^8 OR \8^^^^&&&). Quantized reverse probing similarly departs from the standard forward probe by imposing a decoding bottleneck through clustering; it is explicitly designed to capture joint concepts such as “red apple,” which may be semantically organized in cluster space even when individual attributes are not linearly separable in the original representation (&&&^^^^8max_results8representational probing arXiv RSA probing representations probe diagnostic classification framework8^^^^&&&).

Another methodological fault line concerns what a positive or negative probe result means. Vernier addresses this directly in causal reasoning under lexical renaming: a performance gap between original prompts and placeholder prompts could reflect information loss or representational misalignment. The study’s paired-view LoRA update, variable-name probe, and activation patching diagnostics favor representational misalignment in the working regimes, not simple destruction of task-relevant content (&&&^^^^8representational probing arXiv RSA probing representations probe diagnostic classification framework8^^^^&&&). This is an important conceptual distinction: probe failure does not necessarily imply absence of information.

^^^^8 OR \8^^^^. Interventionist and causal probing

Intervention-based probing asks not only whether a variable is decodable, but whether a recovered subspace or state is causally involved in model behavior. Vernier provides a canonical example. It trains a small LoRA adapter with a paired-view loss over original and placeholder prompts, then inspects the remaining mechanism through linear probes and activation patching. On Qwen-^^^^8 OR \8^^^^B after ^^^^8 OR \8query8query8^^^^ steps, accuracy on the original view rises from PRESERVED_PLACEHOLDER_^^^^8query8 OR \8^^^^ to PRESERVED_PLACEHOLDER_^^^^8query8 OR \8^^^^, accuracy on the placeholder view rises from PRESERVED_PLACEHOLDER_^^^^8query8 OR \8^^^^ to PRESERVED_PLACEHOLDER_^^^^8query88^^^^, and the signed lexical gap PRESERVED_PLACEHOLDER_^^^^8query8^^^^9 shifts from PRESERVED_PLACEHOLDER_^^^^8\8query8^^^^ to PRESERVED_PLACEHOLDER_^^^^8\8representational probing arXiv RSA probing representations probe diagnostic classification framework8^^^^ percentage points. Decision-token activation patching from the original view into the placeholder view raises placeholder accuracy from approximately PRESERVED_PLACEHOLDER_^^^^8\8max_results8^^^^ to approximately PRESERVED_PLACEHOLDER_^^^^8\8query8^^^^, with PRESERVED_PLACEHOLDER_^^^^8\8\8^^^^ agreement with the original answer identity and a PRESERVED_PLACEHOLDER_^^^^8\8 OR \8^^^^ random-donor control (&&&^^^^8representational probing arXiv RSA probing representations probe diagnostic classification framework8^^^^&&&).

The Manifold Probe extends linear probing to curved, multidimensional structure. Under the superposition model PRESERVED_PLACEHOLDER_^^^^8\8 OR \8^^^^, it learns both feature functions PRESERVED_PLACEHOLDER_^^^^8\8 OR \8^^^^ and representational directions PRESERVED_PLACEHOLDER_^^^^8\88^^^^, producing an estimated manifold

PRESERVED_PLACEHOLDER_^^^^8\8^^^^9

Applied to time and space in Llama ^^^^8max_results8^^^^-^^^^8 OR \8^^^^b, the method recovers linearly predictable manifold features and supports steering experiments: for release-year prompts, adding PRESERVED_PLACEHOLDER_^^^^8 OR \8query8^^^^ with PRESERVED_PLACEHOLDER_^^^^8 OR \8representational probing arXiv RSA probing representations probe diagnostic classification framework8^^^^ can increase the probability of completions within PRESERVED_PLACEHOLDER_^^^^8 OR \8max_results8^^^^ years of the target by up to PRESERVED_PLACEHOLDER_^^^^8 OR \8query8^^^^ percentage points at optimal layers (&&&^^^^8query8 OR \8^^^^&&&).

Sparse probing can likewise support causal tests. In multilingual binomial ordering, sparse linear probes recover preference strength most strongly in middle-to-late layers, with a peak Spearman correlation of PRESERVED_PLACEHOLDER_^^^^8 OR \8\8^^^^ at layer ^^^^8representational probing arXiv RSA probing representations probe diagnostic classification framework8\8^^^^ for last-token representations under PRESERVED_PLACEHOLDER_^^^^8 OR \8 OR \8^^^^, using on average about ^^^^8representational probing arXiv RSA probing representations probe diagnostic classification framework8max_results8^^^^ nonzero coefficients out of ^^^^8max_results8^^^^,^^^^8 OR \8 OR \8query8^^^^ dimensions. Injecting the normalized probe direction back into hidden states systematically alters ordering probabilities: positive intervention strengths flatten preference curves, whereas negative strengths accentuate them (&&&^^^^8max_results8max_results8^^^^&&&).

Sparse autoencoders supply another interventionist route. In diffusion models, SAE features learned from penultimate text-encoder activations enable semantic steering by amplifying or suppressing individual sparse features before the U-Net. Validation via CLIP similarity shows a PRESERVED_PLACEHOLDER_^^^^8 OR \8 OR \8^^^^ increase in similarity to the attribute-augmented caption, with PRESERVED_PLACEHOLDER_^^^^8 OR \8 OR \8^^^^ in a paired t-test over PRESERVED_PLACEHOLDER_^^^^8 OR \88^^^^ (&&&^^^^8max_results8query8^^^^&&&). This suggests that some probe-discovered features are not merely descriptive but operationally connected to generation.

^^^^8 OR \8^^^^. Applications, empirical regularities, and benchmarked evaluation

Across application areas, representational probing often reveals trade-offs invisible to downstream accuracy alone. In document-level event extraction, IE fine-tuning improves Argument Detection and Argument Typing to approximately PRESERVED_PLACEHOLDER_^^^^8 OR \8^^^^9–PRESERVED_PLACEHOLDER_^^^^8 OR \8query8^^^^ accuracy, but Coreference degrades from approximately PRESERVED_PLACEHOLDER_^^^^8 OR \8representational probing arXiv RSA probing representations probe diagnostic classification framework8^^^^ in untrained BERT to approximately PRESERVED_PLACEHOLDER_^^^^8 OR \8max_results8^^^^–PRESERVED_PLACEHOLDER_^^^^8 OR \8query8^^^^ after fine-tuning, and Event Typing drops from approximately PRESERVED_PLACEHOLDER_^^^^8 OR \8\8^^^^ in BERT to approximately PRESERVED_PLACEHOLDER_^^^^8 OR \8 OR \8^^^^–PRESERVED_PLACEHOLDER_^^^^8 OR \8 OR \8^^^^ after IE training (&&&^^^^8max_results8^^^^&&&). In common-ground probing between text and vision, text-only contextual models already retrieve correct object categories strongly on unseen categories, with BERT-base reaching PRESERVED_PLACEHOLDER_^^^^8 OR \8 OR \8^^^^, but instance retrieval remains much lower than human performance, with BERT-base at PRESERVED_PLACEHOLDER_^^^^8 OR \88^^^^ versus PRESERVED_PLACEHOLDER_^^^^8 OR \8^^^^9 for human subjects in a ^^^^8representational probing arXiv RSA probing representations probe diagnostic classification framework8query8query8^^^^-way forced-choice setting (&&&^^^^8representational probing arXiv RSA probing representations probe diagnostic classification framework8 OR \8^^^^&&&).

In reinforcement learning, probe scores can track downstream utility unusually well. Reward probing on Atari^^^^8representational probing arXiv RSA probing representations probe diagnostic classification framework8query8query8^^^^k achieves a Spearman rank correlation of PRESERVED_PLACEHOLDER_^^^^8 OR \8query8^^^^ with IQM-HNS, action probing reaches PRESERVED_PLACEHOLDER_^^^^8 OR \8representational probing arXiv RSA probing representations probe diagnostic classification framework8^^^^, and the probing workflow is reported as up to PRESERVED_PLACEHOLDER_^^^^8 OR \8max_results8^^^^ faster than full RL evaluation (&&&^^^^8representational probing arXiv RSA probing representations probe diagnostic classification framework8max_results8^^^^&&&). In graph learning, transformer-style graph models consistently encode more chemically relevant information than message-passing baselines under linear probes, while architectural changes such as residual connections and virtual nodes materially alter what remains recoverable in the hidden states (&&&^^^^8representational probing arXiv RSA probing representations probe diagnostic classification framework8^^^^9&&&).

In biomedical and neuroscientific settings, geometric analyses often reveal systematic nuisance structure. Computational pathology foundation models show high slide-dependence and low disease-dependence: slide-specificity measured by Cliff’s Delta exceeds PRESERVED_PLACEHOLDER_^^^^8 OR \8query8^^^^ for all models, disease-specificity lies around PRESERVED_PLACEHOLDER_^^^^8 OR \8\8^^^^–PRESERVED_PLACEHOLDER_^^^^8 OR \8 OR \8^^^^, and Macenko stain normalization reduces slide-dependence by PRESERVED_PLACEHOLDER_^^^^8 OR \8 OR \8^^^^ to PRESERVED_PLACEHOLDER_^^^^8 OR \8 OR \8^^^^ depending on the model (&&&^^^^8max_results8 OR \8^^^^&&&). Mouse V^^^^8representational probing arXiv RSA probing representations probe diagnostic classification framework8^^^^ digital twins exhibit positive correlations between neural-response prediction quality and linear probe performance, with contrast AUC versus neural PRESERVED_PLACEHOLDER_^^^^8 OR \88^^^^ at PRESERVED_PLACEHOLDER_^^^^8 OR \8^^^^9, $\{(z_i,y_i)\}$^^^^8query8^^^^, and hidden-population eigenspectrum exponent $\{(z_i,y_i)\}$^^^^8representational probing arXiv RSA probing representations probe diagnostic classification framework8^^^^ versus neural prediction at $\{(z_i,y_i)\}$^^^^8max_results8^^^^, $\{(z_i,y_i)\}$^^^^8query8^^^^ (&&&^^^^8 OR \8query8^^^^&&&).

Audio and brain-alignment studies expose similar structure at different timescales. SARL finds that source-level factors are consistently easier to decode than room-level factors, that input configuration and training paradigm shape spatial encoding, and that sensitivity analysis under controlled perturbations reveals heterogeneous responses to source and room variation (Chen et al., 4 Jun 2026). Audio LLM–EEG alignment shows depth-dependent peaks and a pronounced increase in RSA within the ^^^^8max_results8 OR \8query8^^^^–^^^^8 OR \8query8query8^^^^ ms time window, consistent with N^^^^8\8query8query8^^^^-related neural dynamics, while negative prosody decreases rank-based geometric similarity but increases covariance-based dependence (&&&^^^^8representational probing arXiv RSA probing representations probe diagnostic classification framework8 OR \8^^^^&&&). In the color-qualia benchmark, almost all tested vision models align better with pure perception than with task-modulated perception, and high-performing AI models exhibit an anti-diagonal symmetry in their RDMs that human RDMs do not share, suggesting a divergence between artificial and biological inductive biases (&&&^^^^8query8^^^^&&&).

Benchmarking and tooling have become a distinct subtheme. The color-qualia task is packaged in a Brain-Score compatible format with separate no-report and report metrics (&&&^^^^8query8^^^^&&&). SARL is released as an open-source benchmark for spatial audio probing (Chen et al., 4 Jun 2026). Probe-Ably automates data processing, control-task generation, hyperparameter search, probe training, and evaluation with standardized outputs (&&&^^^^8query8^^^^&&&). Taken together, these developments suggest that representational probing has matured from isolated diagnostic classifiers into a broader experimental methodology: one that combines simple readouts, geometric comparison, control tasks, information-theoretic criteria, and causal interventions to determine not only whether a representation contains a property, but how that property is organized and whether it matters for behavior.