论文检索

Social Aspects · Accountability, Transparency and Interpretability

Brian Christian, Jessica Thompson, Elle Michelle Yang, Vincent Adam, Hannah Kirk, Christopher Summerfield, Tsvetomira Dumbalska

Reward models (RMs) are central to aligning large language models (LLMs) with human values but have received less attention than pretrained and post-trained LLMs themselves. Because RMs are initialized from LLMs, they inherit representations that shape their behavior, but the nature and extent of this influence remain understudied. In a comprehensive study of 10 leading open-weight RMs using validated psycholinguistic corpora, we show that RMs exhibit significant differences along multiple dimensions of human value as a function of their base model. Using the "Big Two" psychological axes, we show a robust preference of Llama RMs for "agency" and a corresponding robust preference of Gemma RMs for "communion." This phenomenon holds even when the preference data and finetuning process are identical, and we trace it back to the logits of the respective instruction-tuned and pretrained models. These log-probability differences themselves can be formulated as an implicit RM; we derive usable implicit reward scores and show that they exhibit the very same agency/communion difference. We run experiments training RMs with ablations for preference data source and quantity, which demonstrate that this effect is not only repeatable but surprisingly durable. Despite RMs being designed to represent human preferences, our evidence shows that their outputs are influenced by the pretrained LLMs on which they are based. This work underscores the importance of safety and alignment efforts at the pretraining stage, and makes clear that open-source developers' choice of base model is as much a consideration of values as of performance.

Social Aspects · Accountability, Transparency and Interpretability

Gonçalo Paulo, Nora Belrose

Sparse autoencoders (SAEs) are a useful tool for uncovering human-interpretable features in the activations of large language models (LLMs). While some expect SAEs to find the true underlying features used by a model, our research shows that SAEs trained on the same model and data, differing only in the random seed used to initialize their weights, identify different sets of features. For example, in an SAE with 131K latents trained on a feedforward network in Llama 3 8B, only 30\% of the features were shared across different seeds. We observed this phenomenon across multiple layers of three different LLMs, two datasets, and several SAE architectures. While ReLU SAEs trained with the L1 sparsity loss showed greater stability across seeds, SAEs using the state-of-the-art TopK activation function were more seed-dependent, even when controlling for the level of sparsity. Our results suggest that the set of features uncovered by an SAE should be viewed as a pragmatically useful decomposition of activation space, rather than an exhaustive and universal list of features ``truly used'' by the model.

Social Aspects · Accountability, Transparency and Interpretability

Gonçalo Paulo, Nora Belrose

Sparse autoencoders (SAEs) and transcoders have become important tools for machine learning interpretability. However, measuring how interpretable they are remains challenging, with weak consensus about which benchmarks to use. Most evaluation procedures start by producing a single-sentence explanation for each latent. These explanations are then evaluated based on how well they enable an LLM to predict the activation of a latent in new contexts. This method makes it difficult to disentangle the explanation generation and evaluation process from the actual interpretability of the latents discovered. In this work, we adapt existing methods to assess the interpretability of sparse coders, with the advantage that they do not require generating natural language explanations as an intermediate step. This enables a more direct and potentially standardized assessment of interpretability. Furthermore, we compare the scores produced by our interpretability metrics with human evaluations across similar tasks and varying setups, offering suggestions for the community on improving the evaluation of these techniques.

Social Aspects · Accountability, Transparency and Interpretability

Sayan Saha, Sanghamitra Bandyopadhyay

Model-level explanations for Graph Neural Networks (GNNs) aim to identify class-discriminative motifs that capture how a classifier recognizes a target class. Because the true motifs relied on by the classifier are unobservable, most approaches evaluate explanations by their target class score. However, class score alone is not sufficient as high-scoring explanations may be pathological or may fail to reflect the full range of motifs recognized by the classifier. To bridge this gap, this work introduces sufficiency risk as a formal criterion for whether explanations adequately represent the classifier’s reasoning, and derives distribution-free certificates that upper-bound this risk. Building on this foundation, three metrics are introduced: Coverage, Greedy Gain Area (GGA), and Overlap which operationalize the certificates to assess sufficiency, efficiency, and redundancy in explanations. To ensure practical utility, finite-sample concentration bounds are developed for these metrics, providing confidence intervals that enable statistically reliable comparison between explainers. Experiments on synthetic data and with three state-of-the-art explainers on four real-world datasets demonstrate that these metrics reveal differences in explanation quality hidden by class scores alone. Designed to complement class score, they constitute the first theoretically certified framework for evaluating model-level explanations of GNNs.

Social Aspects · Accountability, Transparency and Interpretability

Jingtong Su, Julia Kempe, Karen Ullrich

Transformers have achieved state-of-the-art performance across diverse language and vision tasks. This success drives the imperative to interpret their internal mechanisms with the dual goals of enhancing performance and improving behavioral control. Attribution methods help advance interpretability by assigning model outputs associated with a target concept to specific model components. Current attribution research primarily studies multi-layer perceptron (MLP) neurons and addresses relatively simple concepts such as factual associations (e.g., Paris is located in France). This focus tends to overlook the impact of the attention mechanism and lacks a unified approach for analyzing more complex concepts. To fill these gaps, we introduce Scalable Attention Module Discovery (SAMD), a concept-agnostic method for mapping arbitrary, complex concepts to specific attention heads of general transformer models. We accomplish this by representing each concept as a vector, calculating its cosine similarity with each attention head, and selecting the TopK-scoring heads to construct the concept-associated attention module. We then propose Scalar Attention Module Intervention (SAMI), a simple strategy to diminish or amplify the effects of a concept by adjusting the attention module using only a single scalar parameter. Empirically, we demonstrate SAMD on concepts of varying complexity, and visualize the locations of their corresponding modules. Our results demonstrate that module locations remain stable before and after LLM post-training, and confirm prior work on the mechanics of LLM multi-lingualism. Through SAMI, we facilitate jailbreaking on HarmBench (+72.7%) by diminishing “safety” and improve performance on the GSM8K benchmark (+1.6%) by amplifying “reasoning”. Lastly, we highlight the domain-agnostic nature of our approach by suppressing the image classification accuracy of vision transformers on ImageNet.

Social Aspects · Accountability, Transparency and Interpretability

Aideen Fay, Inés García-Redondo, Qiquan Wang, Haim Dubossarsky, Anthea Monod

Existing interpretability methods for Large Language Models (LLMs) predominantly capture linear directions or isolated features. This overlooks the high-dimensional, relational, and nonlinear geometry of model representations. We apply persistent homology (PH) to characterize how adversarial inputs reshape the geometry and topology of internal representation spaces of LLMs. This phenomenon, especially when considered across operationally different attack modes, remains poorly understood. We analyze six models (3.8B to 70B parameters) under two distinct attacks, indirect prompt injection and backdoor fine-tuning, and show that a consistent topological signature persists throughout. Adversarial inputs induce topological compression, where the latent space becomes structurally simpler, collapsing the latent space from varied, compact, small-scale features into fewer, dominant, large-scale ones. This signature is architecture-agnostic, emerges early in the network, and is highly discriminative across layers. By quantifying the shape of activation point clouds and neuron-level information flow, our framework reveals geometric invariants of representational change that complement existing linear interpretability methods.

Social Aspects · Accountability, Transparency and Interpretability

Martina G. Vilas, Safoora Yousefi, Besmira Nushi, Eric Horvitz, Vidhisha Balachandran

Reasoning models improve their problem-solving ability through inference-time scaling, allocating more compute via longer and multiple token budgets. Identifying which reasoning traces are likely to succeed remains a key opportunity: reliably predicting productive thinking paths can substantially reduce wasted computation and improve overall efficiency. We introduce _Latent-Trajectory_ signals that characterize the temporal evolution of a model's internal representations during the generation of intermediate reasoning tokens. By analyzing both the extent and temporal course of latent representational change, as well as its alignment with the final state, we show that these signals are strong predictors of solution accuracy, outperforming conventional output-based confidence measures. We use latent-trajectory signals to guide answer selection across multiple sampled generations, demonstrating that they make test-time scaling more effective and efficient, reducing token usage by up to 70% while preserving and even improving accuracy on average in comparison with majority voting. Finally, we show that these signals often emerge early in the reasoning trace, which enables early selection and allocation of compute to the most promising answer candidates during generation. Our findings contribute not only practical strategies for inference-time efficiency, but also a deeper interpretability perspective on how reasoning processes are represented and differentiated in latent space.

Social Aspects · Accountability, Transparency and Interpretability

Tim Hua, Andrew Qin, Samuel Marks, Neel Nanda

Large language models (LLMs) can sometimes detect when they are being evaluated and adjust their behavior to appear more aligned, compromising the reliability of safety evaluations. In this paper, we show that adding a steering vector to an LLM's activations can suppress evaluation-awareness and make the model act like it is deployed during evaluation. To study our steering technique, we train an LLM to exhibit evaluation-aware behavior using a two-step training process designed to mimic how this behavior could emerge naturally. First, we perform continued pretraining on two sets of documents describing its behavior. The first says that our model uses Python type hints during evaluation but not during deployment. The second says that our model can recognize that the presence of a certain evaluation cue always means that it is being tested. Then, we train the model with expert iteration to use Python type hints in evaluation settings. The resulting model is evaluation-aware: it writes type hints in evaluation contexts more than deployment contexts. We find that activation steering can suppress evaluation awareness and make the model behave during evaluation as it would during deployment. Importantly, we constructed our steering vector using the original model before our additional training. Our results suggest that AI evaluators could improve the reliability of safety evaluations by steering models to act like they are deployed.

Social Aspects · Accountability, Transparency and Interpretability

Robert Graham, Edward Stevinson, Leo Richter, Alexander Chia, Joseph Miller, Joseph Bloom

Identifying inputs that trigger specific behaviours or latent features in language models could have a wide range of safety use cases. We investigate a class of methods capable of generating targeted, linguistically fluent inputs that activate specific latent features or elicit model behaviours. We formalise this approach as *context modification* and present ContextBench - a benchmark with tasks designed to assess the capabilities of context modification methods across core capabilities and potential safety applications. Our evaluation framework measures both elicitation strength (the degree to which latent features or behaviours are successfully elicited) and linguistic fluency, highlighting how current state-of-the-art methods struggle to balance these objectives. We develop two novel enhancements to Evolutionary Prompt Optimisation, a gradient-based token-editing method: LLM-assistance and diffusion model inpainting, achieving strong performance in balancing elicitation and fluency. We release our benchmark here: https://github.com/lasr-eliciting-contexts/ContextBench.

Social Aspects · Accountability, Transparency and Interpretability

Oscar Hill, Mateo Espinosa Zarlenga, Mateja Jamnik

Modern deep neural networks remain challenging to interpret due to the opacity of their latent representations, impeding model understanding, debugging, and debiasing. Concept Embedding Models (CEMs) address this by mapping inputs to human-interpretable concept representations from which tasks can be predicted. Yet, CEMs fail to represent inter-concept relationships and require concept annotations at different granularities during training, limiting their applicability. In this paper, we introduce *Hierarchical Concept Embedding Models* (HiCEMs), a new family of CEMs that explicitly model concept relationships through hierarchical structures. To enable HiCEMs in real-world settings, we propose *Concept Splitting*, a method for automatically discovering finer-grained sub-concepts from a pretrained CEM’s embedding space without requiring additional annotations. This allows HiCEMs to generate fine-grained explanations from limited concept labels, reducing annotation burdens. Our evaluation across multiple datasets, including a user study and experiments on *PseudoKitchens*, a newly proposed concept-based dataset of 3D kitchen renders, demonstrates that (1) Concept Splitting discovers human-interpretable sub-concepts absent during training that can be used to train highly accurate HiCEMs, and (2) HiCEMs enable powerful test-time concept interventions at different granularities, leading to improved task accuracy.

Social Aspects · Accountability, Transparency and Interpretability

Nikhil Prakash, Natalie Shapira, Arnab Sen Sharma, Christoph Riedl, Yonatan Belinkov, Tamar Shaham, David Bau, Atticus Geiger

How do language models (LMs) represent characters’ beliefs, especially when those beliefs may differ from reality? This question lies at the heart of understanding the Theory of Mind (ToM) capabilities of LMs. We analyze LMs' ability to reason about characters’ beliefs using causal mediation and abstraction. We construct a dataset, CausalToM, consisting of simple stories where two characters independently change the state of two objects, potentially unaware of each other's actions. Our investigation uncovered a pervasive algorithmic pattern that we call a lookback mechanism, which enables the LM to recall important information when it becomes necessary. The LM binds each character-object-state triple together by co-locating their reference information, represented as Ordering IDs (OIs), in low-rank subspaces of the state token's residual stream. When asked about a character's beliefs regarding the state of an object, the binding lookback retrieves the correct state OI and then the answer lookback retrieves the corresponding state token. When we introduce text specifying that one character is (not) visible to the other, we find that the LM first generates a visibility ID encoding the relation between the observing and the observed character OIs. In a visibility lookback, this ID is used to retrieve information about the observed character and update the observing character's beliefs. Our work provides insights into belief tracking mechanisms, taking a step toward reverse-engineering ToM reasoning in LMs.

Social Aspects · Accountability, Transparency and Interpretability

Arnab Sen Sharma, Giordano Rogers, Natalie Shapira, David Bau

We investigate the mechanisms underlying a range of list-processing tasks in LLMs, and we find that they have learned to encode a compact, causal representation of a general filtering operation that mirrors the generic ``filter'' function of functional programming. Using causal mediation analysis on a diverse set of list-processing tasks, we find that a small number of attention heads, which we dub *filter heads*, encode a compact representation of the filtering predicate in their query states at certain tokens. We demonstrate that this predicate representation is general and portable: it can be extracted and reapplied to execute the same filtering operation on different collections, presented in different formats, languages, or even in tasks. However, we also identify situations where LMs can exploit a different strategy for filtering: eagerly evaluating if an item satisfies the predicate and storing this intermediate result as a flag directly in the item representations. Our results reveal that transformer LMs can develop human-interpretable implementations of abstract computational operations that generalize in ways that are surprisingly similar to strategies used in traditional functional programming patterns.

Social Aspects · Accountability, Transparency and Interpretability

Thibaud Gloaguen, Robin Staab, Nikola Jovanović, Martin Vechev

We introduce the first watermark tailored for diffusion language models (DLMs), an emergent LLM paradigm able to generate tokens in arbitrary order, in contrast to standard autoregressive language models (ARLMs) which generate tokens sequentially. While there has been much work in ARLM watermarking, a key challenge when attempting to apply these schemes directly to the DLM setting is that they rely on previously generated tokens, which are not always available with DLM generation. In this work we address this challenge by: (i) applying the watermark in expectation over the context even when some context tokens are yet to be determined, and (ii) promoting tokens which increase the watermark strength when used as context for other tokens. This is accomplished while keeping the watermark detector unchanged. Our experimental evaluation demonstrates that the DLM watermark leads to a >99\% true positive rate with minimal quality impact and achieves similar robustness to existing ARLM watermarks, enabling for the first time reliable DLM watermarking.

Social Aspects · Accountability, Transparency and Interpretability

Thibaud Gloaguen, Robin Staab, Nikola Jovanović, Martin Vechev

Most LLM fingerprinting methods teach the model to respond to a few fixed queries with predefined atypical responses (keys). This memorization often does not survive common deployment steps such as finetuning or quantization, and such keys can be easily detected and filtered from LLM responses, ultimately breaking the fingerprint. To overcome these limitations we introduce *LLM fingerprinting via semantically conditioned watermarks*, replacing fixed query sets with a broad semantic domain, and replacing brittle atypical keys with a statistical watermarking signal diffused throughout each response. After teaching the model to watermark its responses only to prompts from a predetermined domain e.g., French language, the model owner can use queries from that domain to reliably detect the fingerprint and verify ownership. As we confirm in our thorough experimental evaluation, our fingerprint is both stealthy and robust to all common deployment scenarios.

Social Aspects · Accountability, Transparency and Interpretability

Piotr Komorowski, Elena Golimblevskaia, Reduan Achtibat, Thomas Wiegand, Sebastian Lapuschkin, Wojciech Samek

The capacity of Large Language Models (LLMs) to follow complex instructions and generate factually accurate text is critical for their real-world application. However, standard decoding methods often fail to robustly satisfy these requirements, while existing control techniques frequently degrade general output quality. In this work, we introduce Attribution-Guided Decoding (AGD), an interpretability-based decoding strategy. Instead of directly manipulating model activations, AGD considers a set of high-probability output token candidates and selects the one that exhibits the highest attribution to a user-defined Region of Interest (ROI). This ROI can be flexibly defined over different parts of the model's input or internal components, allowing AGD to steer generation towards various desirable behaviors. We demonstrate AGD's efficacy across three challenging domains. For instruction following, we show that AGD significantly boosts adherence (e.g., improving the overall success rate on Llama 3.1 from 66.0\% to 79.1\%). For knowledge-intensive tasks, we show that guiding generation towards usage of internal knowledge components or contextual sources can reduce hallucinations and improve factual accuracy in both closed-book and open-book settings. Furthermore, we propose an adaptive, entropy-based variant of AGD that mitigates quality degradation and reduces computational overhead by applying guidance only when the model is uncertain. Our work presents a versatile, more interpretable, and effective method for enhancing the reliability of modern LLMs.

Reinforcement Learning · Multi-agent

HyunJi Nam, Yanming Wan, Mickel Liu, Peter Ahnn, Jianxun Lian, Natasha Jaques

As everyday use cases of large language model (LLM) AI assistants have expanded, it is becoming increasingly important to personalize responses to align to different users' preferences and goals. While reinforcement learning from human feedback (RLHF) is effective at improving LLMs to be generally more helpful and fluent, it does not account for variability across users, as it models the entire user population with a single reward model, meaning it assumes that everyone's preferences are the same. We present a novel framework, **P**reference **L**earning **U**sing **S**ummarization (**PLUS**), that uses reinforcement learning (RL) to learn to produce text-based summaries of each user's preferences, characteristics, and past conversations. These summaries condition the reward model, enabling it to make personalized predictions about the types of responses valued by each user. Both the user-summarization model and reward model are trained simultaneously, creating an online co-adaptation loop. We show that in contrast to the standard Bradley–Terry model, summaries produced by PLUS capture diverse aspects of user preferences, achieving a 11–77\% improvement in reward model accuracy. Key strengths of PLUS are: (1) robust performance with new users and conversation topics, achieving a 25\% improvement over the best personalized reward model technique used for RLHF; (2) zero-shot personalization with state-of-the-art proprietary models like GPT-4 (e.g., PLUS-summary-conditioned responses achieved a 72\% win rate compared to 28\% for default GPT-4o); (3) learning from flexible user contexts beyond preference labels, and (4) interpretable representation of users, enabling greater transparency and user control in pluralistic LLM alignment.

Yuhan Li 0001, Xinni Zhang, Linhao Luo, Heng Chang, Yuxiang Ren, Irwin King, Jia Li 0009

Explainable recommendation has demonstrated significant advantages in informing users about the logic behind recommendations, thereby increasing system transparency, effectiveness, and trustworthiness. To provide personalized and interpretable explanations, existing works often combine the generation capabilities of large language models (LLMs) with collaborative filtering (CF) information. CF information extracted from the user-item interaction graph captures the user behaviors and preferences, which is crucial for providing informative explanations. However, due to the complexity of graph structure, effectively extracting the CF information from graphs still remains a challenge. Moreover, existing methods often struggle with the integration of extracted CF information with LLMs due to its implicit representation and the modality gap between graph structures and natural language explanations. To address these challenges, we propose G-Refer, a framework using Graph Retrieval-augmented large language models (LLMs) for explainable recommendation. Specifically, we first employ a hybrid graph retrieval mechanism to retrieve explicit CF signals from both structural and semantic perspectives. The retrieved CF information is explicitly formulated as human-understandable text by the proposed graph translation and accounts for the explanations generated by LLMs. To bridge the modality gap, we introduce knowledge pruning and retrieval-augmented fine-tuning to enhance the ability of LLMs to process and utilize the retrieved CF information to generate explanations. Extensive experiments show that G-Refer achieves superior performance compared with existing methods in both explainability and stability. Codes and data are available at https://github.com/Yuhan1i/G-Refer.

Mátyás Vincze, Laura Ferrarotti, Leonardo Lucio Custode, Bruno Lepri, Giovanni Iacca

Continuous control tasks often involve high-dimensional, dynamic, and non-linear environments. State-of-the-art performance in these tasks is achieved through complex closed-box policies that are effective, but suffer from an inherent opacity. Interpretable policies, while generally underperforming compared to their closed-box counterparts, advantageously facilitate transparent decision-making within automated systems. Hence, their usage is often essential for diagnosing and mitigating errors, supporting ethical and legal accountability, and fostering trust among stakeholders. In this paper, we propose SMoSE, a novel method to train sparsely activated interpretable controllers, based on a top-1 Mixture-of-Experts architecture. SMoSE combines a set of interpretable decision-makers, trained to be experts in different basic skills, and an interpretable router that assigns tasks among the experts. The training is carried out via state-of-the-art Reinforcement Learning algorithms, exploiting load-balancing techniques to ensure fair expert usage. We then distill decision trees from the weights of the router, significantly improving the ease of interpretation. We evaluate SMoSE on six benchmark environments from MuJoCo: our method outperforms recent interpretable baselines and narrows the gap with non-interpretable state-of-the-art algorithms.

Xiusi Chen, Shanyong Wang, Cheng Qian, Hongru Wang, Peixuan Han, Heng Ji

In high-stakes domains such as healthcare and finance, effective decision-making demands not just accurate outcomes but transparent and explainable reasoning. However, current language models often lack the structured deliberation needed for such tasks, instead generating decisions and justifications in a disconnected, post-hoc manner. To address this, we propose DecisionFlow, a novel decision modeling framework that guides models to reason over structured representations of actions, attributes, and constraints. Rather than predicting answers directly from prompts, DecisionFlow builds a semantically grounded decision space and infers a latent utility function to evaluate trade-offs in a transparent, utility-driven manner. This process produces decisions tightly coupled with interpretable rationales reflecting the model’s reasoning. Empirical results on two high-stakes benchmarks show that DecisionFlow not only achieves up to 30% accuracy gains over strong prompting baselines but also enhances alignment in outcomes. Our work is a critical step toward integrating symbolic reasoning with LLMs, enabling more accountable, explainable, and reliable LLM decision support systems. Code and data are at https://github.com/xiusic/DecisionFlow.

Social Aspects · Accountability, Transparency and Interpretability

Linjiang Zhou, Chao Ma, Zepeng Wang, Libing Wu, XIAOCHUAN SHI

Gradient smoothing is an efficient approach to reducing noise in gradient-based model explanation methods. SmoothGrad adds Gaussian noise to mitigate much of this noise. However, the crucial hyperparameter in this method, the variance $\sigma$ of the Gaussian noise, is often set manually or determined using a heuristic approach. This results in the smoothed gradients containing extra noise introduced by the smoothing process. In this paper, we aim to analyze the noise and its connection to the out-of-range sampling in the smoothing process of SmoothGrad. Based on this insight, we propose AdaptGrad, an adaptive gradient smoothing method that controls out-of-range sampling to minimize noise. Comprehensive experiments, both qualitative and quantitative, demonstrate that AdaptGrad could effectively reduce almost all the noise in vanilla gradients compared to baseline methods. AdaptGrad is simple and universal, making it a practical solution to enhance gradient-based interpretability methods to achieve clearer visualization.