arXiv subjects

Shauli Ravfogel

Publications and source records attributed to Shauli Ravfogel.

At least 19 recordsLinked to original sources

arXiv2026

What LLMs explain is not what they believe: Evaluating explanation sufficiency under models' own input beliefs

Large language models (LLMs) are increasingly deployed in high-stakes domains, where free-text explanations such as chain-of-thought and post-hoc rationales are used to justify model outputs. Yet it remains unclear whether these explanations are sufficient, i.e., if they contain enough information to explain the model's output-generating process. We generalize classical sufficiency from feature attributions to arbitrary explanations and prove that explanation sufficiency can change depending on the input distribution, which must be explicitly defined for LLM explanations. We propose using the LLM itself to generate alternative inputs conditioned on an explanation, capturing its beliefs about possible inputs. We formalize self-consistent sufficiency as a goal for free-text explanations and introduce an information-theoretic metric, SCSuff, that enables evaluation of free-text explanations without relying on predefined biases or shortcuts. Our experiments show that SCSuff agrees with targeted perturbation tests where applicable and demonstrate that explanation sufficiency can vary with the input distribution. We find LLM explanations are generally insufficient and weakly correlated with model size, accuracy, or output entropy. Analysis of final-token hidden states shows that top and bottom SCSuff scores can be predicted from internal representations, suggesting that SCSuff can guide detection and improvement of sufficient LLM explanations. The code for this paper is available at https://github.com/rajesh-lab/self-consistent-sufficiency .

cs.LG↗

arXiv2026

Can LLMs Introspect? A Reality Check

Can large language models detect and report their own internal states? A number of recent studies have argued that they can. Drawing on lessons from human metacognition research, we argue that this conclusion may be premature. We identify two conditions that a paradigm needs to meet in order to establish introspection. First, the test needs to require privileged access: it should not be solvable using cues available in the input. Second, it needs to require second-order computation: second-order, meta-representations of first-order, task-related representations. This condition cannot be satisfied by task performance alone: it requires designs under which second-order and first-order accounts make divergent predictions. We re-examine two paradigms that have been used to argue for model introspection in light of these conditions. In the first, models must predict labels derived from their own hidden states; we find that classifiers that can only access the input match the models' in-context predictions, indicating that the original results do not demonstrate privileged access to internal representations. In the second paradigm, models must detect whether their internal states have been tampered with; we find they cannot reliably distinguish such interventions from manipulations of the input, suggesting that their success reflects generic anomaly detection rather than sensitivity to internal interventions in particular. We conclude that current evidence is insufficient to establish metacognitive monitoring in LLMs.

cs.AI↗

arXiv2026

Can LLM Agents Infer World Models? Evidence from Agentic Automata Learning

We propose agentic automata learning to evaluate the extent to which tool-calling LLM agents can uncover hidden environments through interaction. In our setup, an agent should uncover a hidden deterministic finite automaton (DFA) by interacting with an oracle through (1) membership queries ("Does this string belong to the target language?") and (2) equivalence queries ("Is this the target DFA?"). This yields a scalable testbed with controlled task complexity, measurable interaction efficiency, and strong baselines (classic automata-learning algorithms). Evaluating state-of-the-art LLMs, we find that performance drops sharply as DFA size increases. Reasoning models are markedly stronger than non-reasoning models, yet trajectory analyses reveal recurring failures in query planning, evidence integration, and hypothesis construction. Overall, our results show that current LLM agents can sometimes perform non-trivial interactive discovery, but remain far less robust and efficient than classic algorithms for the task.

cs.CL↗

arXiv2026

From Benchmarks to Skills: Low-Rank Factors for LLM Evaluation

Current evaluations of large language models (LLMs) rely heavily on a growing collection of benchmarks and on aggregate benchmark scores, yet it remains unclear what this comparison actually captures, and what these scores reveal about models' underlying capabilities. Here, we propose a new paradigm for LLM evaluation, by asking whether benchmark performance reflects many independent abilities, or rather relies on a small number of shared dimensions. To answer this, we apply Factor Analysis (FA) to a massive performance matrix of LLMs versus benchmarks $(60\times44)$ revealing an \emph{intrinsically low-rank} structure of that matrix. That is, a small number of latent factors captures most of the structure in the full task space. This low-rank geometry reveals substantial redundancy across existing tasks and explains why many benchmarks appear to be measuring overlapping abilities. We further show that these latent factors correspond to coherent, skill-like, dimensions of LLM behavior. Leveraging this latent skill-space, we deliver three practical tools for LLM evaluation and downstream users: (i)~identifying redundant tasks, (ii)~profiling new models using a small subset of tasks, and (iii)~selecting models aligned with desired skill profiles. Our method provides a solid alternative to the de-facto standard of a single aggregate score, and establishes an interpretable and practical framework for understanding and benchmarking LLM core capabilities.

cs.CL↗

arXiv2026

Geometric Factual Recall in Transformers

How do transformer language models memorize factual associations? A common view casts internal weight matrices as associative memories over pairs of embeddings, requiring parameter counts that scale linearly with the number of facts. We develop a theoretical and empirical account of an alternative, \emph{geometric} form of memorization in which learned embeddings encode relational structure directly, and the MLP plays a qualitatively different role. In a controlled setting where a single-layer transformer must memorize random bijections from subjects to a shared attribute set, we prove that a logarithmic embedding dimension suffices: subject embeddings encode \emph{linear superpositions} of their associated attribute vectors, and a small MLP acts as a relation-conditioned selector that extracts the relevant attribute via ReLU gating, and not as an associative key-value mapping. We extend these results to the multi-hop setting -- chains of relational queries such as ``Who is the mother of the wife of $x$?'' -- providing constructions with and without chain-of-thought that exhibit a provable capacity-depth tradeoff, complemented by a matching information-theoretic lower bound. Empirically, gradient descent discovers solutions with precisely the predicted structure. Once trained, the MLP transfers zero-shot to entirely new bijections when subject embeddings are appropriately re-initialized, revealing that it has learned a generic selection mechanism rather than memorized any particular set of facts.

cs.CL↗

arXiv2026

RELIC: Evaluating Complex Reasoning via the Recognition of Languages In-Context

Large language models (LLMs) are increasingly used to solve complex tasks where they must retrieve and compose many pieces of in-context information in long reasoning chains. For many real-world tasks it is hard to accurately gauge how model performance and strategy change as task complexity grows. To evaluate models' complex reasoning capability in a scalable and verifiable way, we introduce RELIC (Recognition of Languages In-Context), a framework that evaluates an LLM's ability to decide whether a given string belongs to the context-free language (CFL) generated by a grammar presented in-context. CFL recognition allows us to modulate the intrinsic complexity of the problem by varying grammar size and string length and translate this asymptotic complexity into predictions for ideal LLM performance. We find that even the most advanced reasoning models perform poorly on RELIC, not only failing to appropriately scale their inference compute to keep pace with task difficulty, but even reducing the number of reasoning tokens they use as task complexity increases. We find that these decreases in compute accompany changes in reasoning strategy, as models move from identifying and implementing algorithmic solutions to guessing. For models whose full completions go uninspected, this manifests as ``quiet quitting'' on hard tasks.

cs.CL↗

arXiv2026

Discrete Diffusion Models Exploit Asymmetry to Solve Lookahead Planning Tasks

While Autoregressive (AR) Transformer-based Generative Language Models are frequently employed for lookahead tasks, recent research suggests a potential discrepancy in their ability to perform planning tasks that require multi-step lookahead. In this work, we investigate the distinct emergent mechanisms that arise when training AR versus Non-Autoregressive (NAR) models, such as Discrete Diffusion Models (dLLMs), on lookahead tasks. By requiring the models to plan ahead to reach the correct conclusion, we analyze how these two paradigms fundamentally differ in their approach to the problem. We identify a critical asymmetry in planning problems: while forward generation requires complex lookahead at branching junctions, reverse generation is often deterministic. This asymmetry creates an opportunity for NAR models. Through mechanistic analysis of training and inference dynamics, we demonstrate that NAR models learn to solve planning tasks by utilizing future tokens to decode backwards, avoiding the need to learn complex traversal mechanisms entirely. Consequently, we report that both AR and NAR models are able to achieve perfect accuracy on the lookahead task. However, NAR models require exponentially fewer training examples and shallower architectures compared to AR models, which often fail to converge without specific curriculum adjustments.

cs.LG↗

arXiv2026

The Truthfulness Spectrum Hypothesis

Large language models (LLMs) have been reported to linearly encode truthfulness, yet recent work questions this finding's generality. We reconcile these views with the truthfulness spectrum hypothesis: the representational space contains directions ranging from broadly domain-general to narrowly domain-specific. To test this hypothesis, we systematically evaluate probe generalization across five truth types (definitional, empirical, logical, fictional, and ethical), sycophantic and expectation-inverted lying, and existing honesty benchmarks. Linear probes generalize well across most domains but fail on sycophantic and expectation-inverted lying. Yet training on all domains jointly recovers strong performance, confirming that domain-general directions exist despite poor pairwise transfer. The geometry of probe directions explains these patterns: Mahalanobis cosine similarity between probes near-perfectly predicts cross-domain generalization (R^2=0.98). Concept-erasure methods further isolate truth directions that are (1) domain-general, (2) domain-specific, or (3) shared only across particular domain subsets. Causal interventions reveal that domain-specific directions steer more effectively than domain-general ones. Finally, post-training reshapes truth geometry, pushing sycophantic lying further from other truth types, suggesting a representational basis for chat models' sycophantic tendencies. Together, our results support the truthfulness spectrum hypothesis: truth directions of varying generality coexist in representational space, with post-training reshaping their geometry. Code for all experiments is provided in https://github.com/zfying/truth_spec.

cs.LG↗

arXiv2026

From Directions to Regions: Decomposing Activations in Language Models via Local Geometry

Activation decomposition methods in language models are tightly coupled to geometric assumptions on how concepts are realized in activation space. Existing approaches search for individual global directions, implicitly assuming linear separability, which overlooks concepts with nonlinear or multi-dimensional structure. In this work, we leverage Mixture of Factor Analyzers (MFA) as a scalable, unsupervised alternative that models the activation space as a collection of Gaussian regions with their local covariance structure. MFA decomposes activations into two compositional geometric objects: the region's centroid in activation space, and the local variation from the centroid. We train large-scale MFAs for Llama-3.1-8B and Gemma-2-2B, and show they capture complex, nonlinear structures in activation space. Moreover, evaluations on localization and steering benchmarks show that MFA outperforms unsupervised baselines, is competitive with supervised localization methods, and often achieves stronger steering performance than sparse autoencoders. Together, our findings position local geometry, expressed through subspaces, as a promising unit of analysis for scalable concept discovery and model control, accounting for complex structures that isolated directions fail to capture.

cs.CL↗

arXiv2025

State over Tokens: Characterizing the Role of Reasoning Tokens

Large Language Models (LLMs) can generate reasoning tokens before their final answer to boost performance on complex tasks. While these sequences seem like human thought processes, empirical evidence reveals that they are not a faithful explanation of the model's actual reasoning process. To address this gap between appearance and function, we introduce the State over Tokens (SoT) conceptual framework. SoT reframes reasoning tokens not as a linguistic narrative, but as an externalized computational state -- the sole persistent information carrier across the model's stateless generation cycles. This explains how the tokens can drive correct reasoning without being a faithful explanation when read as text and surfaces previously overlooked research questions on these tokens. We argue that to truly understand the process that LLMs do, research must move beyond reading the reasoning tokens as text and focus on decoding them as state.

cs.CL↗

arXiv2025

Preserving Task-Relevant Information Under Linear Concept Removal

Modern neural networks often encode unwanted concepts alongside task-relevant information, leading to fairness and interpretability concerns. Existing post-hoc approaches can remove undesired concepts but often degrade useful signals. We introduce SPLINCE-Simultaneous Projection for LINear concept removal and Covariance prEservation - which eliminates sensitive concepts from representations while exactly preserving their covariance with a target label. SPLINCE achieves this via an oblique projection that 'splices out' the unwanted direction yet protects important label correlations. Theoretically, it is the unique solution that removes linear concept predictability and maintains target covariance with minimal embedding distortion. Empirically, SPLINCE outperforms baselines on benchmarks such as Bias in Bios and Winobias, removing protected attributes while minimally damaging main-task information.

cs.LG↗

arXiv2025

Beyond Single Embeddings: Capturing Diverse Targets with Multi-Query Retrieval

Most text retrievers generate \emph{one} query vector to retrieve relevant documents. Yet, the conditional distribution of relevant documents for the query may be multimodal, e.g., representing different interpretations of the query. We first quantify the limitations of existing retrievers. All retrievers we evaluate struggle more as the distance between target document embeddings grows. To address this limitation, we develop a new retriever architecture, \emph{A}utoregressive \emph{M}ulti-\emph{E}mbedding \emph{R}etriever (AMER). Our model autoregressively generates multiple query vectors, and all the predicted query vectors are used to retrieve documents from the corpus. We show that on the synthetic vectorized data, the proposed method could capture multiple target distributions perfectly, showing 4x better performance than single embedding model. We also fine-tune our model on real-world multi-answer retrieval datasets and evaluate in-domain. AMER presents 4 and 21\% relative gains over single-embedding baselines on two datasets we evaluate on. Furthermore, we consistently observe larger gains on the subset of dataset where the embeddings of the target documents are less similar to each other. We demonstrate the potential of using a multi-query vector retriever and open up a new direction for future work.

cs.CL↗

arXiv2025

Emergence of Linear Truth Encodings in Language Models

Recent probing studies reveal that large language models exhibit linear subspaces that separate true from false statements, yet the mechanism behind their emergence is unclear. We introduce a transparent, one-layer transformer toy model that reproduces such truth subspaces end-to-end and exposes one concrete route by which they can arise. We study one simple setting in which truth encoding can emerge: a data distribution where factual statements co-occur with other factual statements (and vice-versa), encouraging the model to learn this distinction in order to lower the LM loss on future tokens. We corroborate this pattern with experiments in pretrained language models. Finally, in the toy setting we observe a two-phase learning dynamic: networks first memorize individual factual associations in a few steps, then -- over a longer horizon -- learn to linearly separate true from false, which in turn lowers language-modeling loss. Together, these results provide both a mechanistic demonstration and an empirical motivation for how and why linear truth representations can emerge in language models.

cs.CL↗

arXiv2025

The Medium Is Not the Message: Deconfounding Document Embeddings via Linear Concept Erasure

Embedding-based similarity metrics between text sequences can be influenced not just by the content dimensions we most care about, but can also be biased by spurious attributes like the text's source or language. These document confounders cause problems for many applications, but especially those that need to pool texts from different corpora. This paper shows that a debiasing algorithm that removes information about observed confounders from the encoder representations substantially reduces these biases at a minimal computational cost. Document similarity and clustering metrics improve across every embedding variant and task we evaluate -- often dramatically. Interestingly, performance on out-of-distribution benchmarks is not impacted, indicating that the embeddings are not otherwise degraded.

cs.CL↗

arXiv2025

Intrinsic Test of Unlearning Using Parametric Knowledge Traces

The task of "unlearning" certain concepts in large language models (LLMs) has attracted immense attention recently, due to its importance in mitigating undesirable model behaviours, such as the generation of harmful, private, or incorrect information. Current protocols to evaluate unlearning methods largely rely on behavioral tests, without monitoring the presence of unlearned knowledge within the model's parameters. This residual knowledge can be adversarially exploited to recover the erased information post-unlearning. We argue that unlearning should also be evaluated internally, by considering changes in the parametric knowledge traces of the unlearned concepts. To this end, we propose a general evaluation methodology that leverages vocabulary projections to inspect concepts encoded in model parameters. We use this approach to localize "concept vectors" - parameter vectors that encode concrete concepts - and construct ConceptVectors, a benchmark dataset containing hundreds of common concepts and their parametric knowledge traces within two open-source LLMs. Evaluation on ConceptVectors shows that existing unlearning methods minimally impact concept vectors and mostly suppress them during inference, while directly ablating these vectors demonstrably removes the associated knowledge and significantly reduces the model's susceptibility to adversarial manipulation. Our results highlight limitations in behavioral-based unlearning evaluations and call for future work to include parameter-based evaluations. To support this, we release our code and benchmark at https://github.com/yihuaihong/ConceptVectors.

cs.CL↗

arXiv2025

Representation Surgery: Theory and Practice of Affine Steering

Language models often exhibit undesirable behavior, e.g., generating toxic or gender-biased text. In the case of neural language models, an encoding of the undesirable behavior is often present in the model's representations. Thus, one natural (and common) approach to prevent the model from exhibiting undesirable behavior is to steer the model's representations in a manner that reduces the probability of it generating undesirable text. This paper investigates the formal and empirical properties of steering functions, i.e., transformation of the neural language model's representations that alter its behavior. First, we derive two optimal, in the least-squares sense, affine steering functions under different constraints. Our theory provides justification for existing approaches and offers a novel, improved steering approach. Second, we offer a series of experiments that demonstrate the empirical effectiveness of the methods in mitigating bias and reducing toxic generation.

cs.LG↗

arXiv2025

LEACE: Perfect linear concept erasure in closed form

Concept erasure aims to remove specified features from an embedding. It can improve fairness (e.g. preventing a classifier from using gender or race) and interpretability (e.g. removing a concept to observe changes in model behavior). We introduce LEAst-squares Concept Erasure (LEACE), a closed-form method which provably prevents all linear classifiers from detecting a concept while changing the embedding as little as possible, as measured by a broad class of norms. We apply LEACE to large language models with a novel procedure called "concept scrubbing," which erases target concept information from every layer in the network. We demonstrate our method on two tasks: measuring the reliance of language models on part-of-speech information, and reducing gender bias in BERT embeddings. Code is available at https://github.com/EleutherAI/concept-erasure.

cs.LG↗

arXiv2025

GRADE: Quantifying Sample Diversity in Text-to-Image Models

We introduce GRADE, an automatic method for quantifying sample diversity in text-to-image models. Our method leverages the world knowledge embedded in large language models and visual question-answering systems to identify relevant concept-specific axes of diversity (e.g., ``shape'' for the concept ``cookie''). It then estimates frequency distributions of concepts and their attributes and quantifies diversity using entropy. We use GRADE to measure the diversity of 12 models over a total of 720K images, revealing that all models display limited variation, with clear deterioration in stronger models. Further, we find that models often exhibit default behaviors, a phenomenon where a model consistently generates concepts with the same attributes (e.g., 98% of the cookies are round). Lastly, we show that a key reason for low diversity is underspecified captions in training data. Our work proposes an automatic, semantically-driven approach to measure sample diversity and highlights the stunning homogeneity in text-to-image models.

cs.CV↗