Note The authors improve LM-generated descriptions for automatic interpretability by accounting for the effect of each feature on the LM output, either via vocabulary projections or direct probability comparison. Results show that such approaches are less expensive than common input-based autointerp methods, and can be integrated with those to improve the quality of feature descriptions.

Note Thread: https://bsky.app/profile/megamor2.bsky.social/post/3ldlwiz42j22c

Note Authors train a model to answer open-ended questions about model activations in natural language. The approach can be used for reading properties of LM activations, or steering the generation by directly optimizing the latents to closely match a natural language prompt. Thread: https://x.com/aypan_17/status/1867626789847413119

Note Chain of Continuous Thought (CoConuT) uses terminal states as starting point for subsequent generation steps, effectively preserving more information rather than committing to a single output token. Answers are then generated in natural language. The method is shown to outperform CoT and its variants on mathematical reasoning and related tasks. Thread: https://x.com/Ber18791531/status/1866561188664087017

Note Thread: https://bsky.app/profile/michaelwhanna.bsky.social/post/3ldnyqfcrzc2f

Note Thread: https://x.com/jkminder/status/1856671617029398952 Code: https://github.com/kdu4108/context-vs-prior-finetuning

Note Thread: https://bsky.app/profile/bkhmsi.bsky.social/post/3ldo5jnpgm22m

Note Authors apply ablation techniques to popular VLM models, showing that token position corresponding to input image elements preserve most of information from the corresponding locations. Logit lens is also used to show that visual tokens tend naturally towards meaningful word embeddings despite no explicit conditioning. Late layers of the LLM are found responsible for the extraction of visual information from visual tokens.

Note Authors show that LMs encode form complexity linearly and meaning complexity nonlinearly. Form complexity correlates near-perfectly to linear embedding dimension (PCA) of representations since the start of training. Results suggest LMs encode form and meaning differently: form is an inductive bias represented in high-dimensional linear subspaces, while meaning is learned over time in low-dimensional nonlinear manifolds. Thread: https://x.com/sparse_emcheng/status/1842835883726078311

Note Currently SAEs are evaluated via unsupervised metrics like L0. This work proposes using a constrained setting like board games to fix a predefined set of known features (i.e. board states & their properties), and use SAE performances in recovering those as a quality metric. p-annealing is proposed to transition from L0 to L1 during training, improving SAE quality. Thread: https://x.com/a_karvonen/status/1819399813441663042

Note This work introduces Strict IIT (SIIT) to intervene in low-level NN nodes while preserving high-level task coherence. Authors train several LMs with SIIT to follow circuits of interest, producing models that can be used as benchmarks for component attribution methods while remaining realistic, as opposed to Tracr-compiled RASP models. Experiments on various methods show that EAP-IG produces results on par with ACDC. Thread: https://x.com/IvanArcus/status/1815412941677850949

Note This paper examines how language model circuits and components develop throughout pre-training, studying models from 70M -> 2.8B params across 300B training tokens. Task performance and functional components emerge at similar token counts across model scales, with the overall algorithmic structure of circuits remaining roughly stable across training. This suggests a mechanistic understanding of early checkpoints and smaller models could improve MI research efficiency.

Note This work hypothesizes the existence of an implicit vocabulary in autoregressive LMs mapping multi-tokens to word-like embeddings in early layers. A token information erasure phenomenon is discovered for multi-word expressions and used as a signal to retrieve possible implicit vocabulary items for Llama 2 and Llama 3.

Note This work studies optimal strategies for multi-property activation steering of LLMs to combine language conditioning with stylistic properties like safety and formality. Steering intensity is shown to be property-dependent, with a trade-off between steering accuracy and output fluency. Dynamic Activation Composition is proposed to adaptively modulate the steering intensity from the expected shift in the LLM predictive distribution, ensuring good steering while minimizing fluency degradation.

Note This work focuses on neuron-level mechanisms for confidence calibration in LLMs, identifying entropy and token frequency neurons. Entropy neurons have high weight norms but minimal direct impact on logits, effectively modulating output distribution entropy. Token frequency neurons boost logits proportionally to their frequency. Experiments demonstrate the interaction between entropy neurons and other mechanisms, such as induction heads.

Note This work proposes to use LLM internals to create RAG citations, ensuring citations reflect actual context usage during generation rather than surface-level similarity. The proposed method MIRAGE (Model Internals-based RAG Explanations) detects context-sensitive answer tokens and pairs them with retrieved documents contributing to their prediction via saliency methods. The method outperforms 11B NLI validators without model/data requirements, and is more robust and faithful than self-citation.

Note This work aims to test whether model knowledge about an entity can be estimated only from its internal computation. In particular, model internals are used to estimate the models' confidence in QA about the entity, and the factuality of responses. Experiments show that KEEN, a simple probe trained over internal subject representations, succeeds at both tasks. Moreover, KEEN aligns with the model's hedging behavior and faithfully reflects changes in the model's knowledge after fine-tuning.

Note Interpretability and analysis (IA) of LMs is often criticized for lacking actionable insights towards advancing NLP. This study analyzes a citation graph of 185K+ papers from ACL/EMNLP, and surveys 138 NLP researchers about their views on IA. IA work is well-cited, with many works building on IA findings. Authors call IA researchers to center humans in their work, think about the big picture, provide actionable insights and work towards standardized, robust methods.

Note LLMs struggle to reliably recall items from a list as the length of the list increases, depending on the items' positions in it. This work shows that models' prompt sensitivity is caused by failures in information communication across model layers using "communication channels", i.e. subspaces of the residual stream. A new SVD-based procedure is proposed to identify these channels from weight matrices.

Note This work focuses on structural ICL, i.e., extrapolation to unseen tokens within familiar structures. The authors find that structural ICL is transient (i.e., it decays over training) in both natural and synthetic settings. Active forgetting can maintain structural ICL in the synthetic setting. Temporary forgetting is introduced, enabling a dual process strategy where models use in-weight solutions for frequent tokens, and structural ICL solutions for rare ones.

Note CoT reasoning leads to inconsistencies between models' middle and final layers, suggesting some uncertainty. Authors propose to use the internal consistency (IC) of model predictions produced from intermediate layers as a measure of the model's confidence in the reasoning process. IC is found to effectively distinguishes between correct and incorrect reasoning paths, and up-weighting reasoning paths with high internal consistency leads to significant improvements in reasoning performance.

Note This work shows that representation processing in Transformer-based LMs undergoes a phase transition characterized by an expansion of intrinsic dimensionality, cross-model information sharing, and a switch to abstract information processing. This dynamic is reduced in the presence of random text, and absent in untrained model. The depth at which such dynamic appears is found to empirically correlate with model quality.

Note While some features in Transformer LMs are represented linearly in a single dimension, authors find evidence of multidimensional features that can be extracted using techniques such as SAEs. In particular, authors find evidence for circular representations of time-related concepts such as days, months and years in the representations of multiple LMs.

Note This work shows near-linear behavior across various LLM architecture, showing how pre-training tends to increase nonlinear dynamics while fine-tuning can make them more pronounced. Authors speculate that, as embeddings become more uniform across neighboring layers, models may compensate for the reduced variability by amplifying non-linear processing in the residual stream.

Note This work summarizes recent trends in interpretability research for Transformer-based LMs, presenting model components, popular methods, and findings using a joint notation. Approaches are categorized based on their usage for either behavior localization or information decoding, and a section is reserved to present popular tools for conducting interpretability research.

Note This study introduces "clamping," a method inspired by optogenetics, for training-time causal interventions to study mechanism formation. Authors apply clamping to study the emergence of induction heads (IH), finding IHs contribs are additive and redundant, with competition emerging due to optimization pressures, and many-to-many dependent on previous token heads in lower layers. Three critical induction subcircuits are identified, and their formation is connected to data-dependent properties.

Note The LLM transparency toolkit is an open-source toolkit and visual interface to efficiently identify component circuits in LMs responsible for their predictions using Information Flow Routes. The tool can highlight the importance of granular components, and vocabulary projections are provided to examine intermediate predictions of the residual stream, and tokens promoted by specific component updates.

Note This work applies contrastive activation addition, the tuned lens and probing for eliciting latent knowledge in quirky models to Mamba and RWKV LMs, finding these Transformer-specific methods can be applied with slight adaptation to these architectures, obtaining similar results.

Note This work examines the influence of context versus memorized knowledge in LMs through the lens of the shift caused by contexts at various degrees of informativeness to the models' predictive distribution. Authors propose information-theoretic metrics to measure the persuasiveness of a context and the susceptibility of an entity to be influenced by contextual information. Analysis reveals important differences due to model size, query formulation and context assertiveness/negation.

Note This work applies the ROME method to Mamba, finding weights playing the role of MLPs in encoding factual relations across several Mamba layers, and can be patched to perform model editing. A new SSM-specific technique is also introduced to emulate attention knockout (value zeroing) revealing information flows similar to the ones in Transformers when processing factual statements.

Note This work introduces Representation fine-tuning (ReFT), a framework using learned inference-time interventions as efficient yet effective alternatives to PEFT weight adaptation. LoReFT, a ReFT variant intervening linearly on a representation subspaces, is evaluated against several PEFT approaches showing SOTA performances across popular benchmark with 10-50x speedup. The HF-compatible pyreft library is introduced to simplify ReFT usage.

Note This work aims to evaluate whether language models exhibit implicit planning during generation. In a synthetic setting and employing a myopic variant of gradient descent ignoring off-diagonal information, authors find that LMs can implicitly plan for future predictions. However, the same behavior is observed to a much lesser extent for natural language, where computation for current predictions are also functional to future results.

Note This work proposes using features and errors from sparse autoencoders trained to reconstruct LM activations as interpretable units for circuit discovery. The authors then introduce SHIFT, a technique for editing model behavior by ablating interpretable elements from sparse feature circuits. This method is applied alongside unsupervised circuit discovery at scale by means of clustering, showing highly interpretable feature circuits interacting in behaviors like predicting sequence increments.

Note This work introduces an efficient approach for circuit discovery, EAP-IG, using integrated gradients to perform edge attribution patching. Circuits found by EAP-IG and other methods are evaluated in terms of faithfulness, i.e. consistency of pre- and post-patching behavior, finding EAP-IG outperforms EAP across all tested tasks. The overlap between circuits found by activation patching and EAP is a faithfulness indicator only when full or no overlap is present, but not for partial overlap cases

Note This work proposes an efficient approach for circuit discovery. Information flow routes require a single forward pass, and are derived from decomposing component updates into the Transformer residual stream. Experiments on LLaMA 7B show how the contrastive formulation of activation patching (which can be avoided with Information flow routes) can lead to misleading results depending to selected templates.

Note Authors identify two failure modes for the attribution patching (AtP) method for estimating component importance in LMs, leading to false negatives due to attention saturation or cancellation of direct and indirect effects. An improved version named AtP* is proposed to improve the method’s robustness in such settings. A diagnostic procedure is also proposed to bound the error caused by gradient approximation.

Note This work investigates the effect of fine-tuning on circuit-level mechanisms in LLMs, focusing on the entity tracking task on LLaMA 7B variants. Authors find that circuits from the base model persist in fine-tuned models, and their individual components preserve their functionalities. Cross-Model Activation Patching (CMAP) reveals that gains in performance can be attributed to improvements in circuit components, rather than overall functional rearrangement.

Note This work proposes Simple Detectors Aggregation (STARE), an aggregation procedure to leverage hallucination detectors’ complementary strengths in the context of machine translation. Authors experiment with two popular hallucination detection benchmarks (LFAN-HALL and HalOmi), showing that an aggregation of detectors using only model internals can outperform ad-hoc trained hallucination detectors.

Note This work extends Logit Lens vocabulary projections of FFNs in Transformers to gradients to study the knowledge editing performed by backward passes. Authors prove that a gradient matrix can be cast as a low-rank linear combination of its forward and backward passes’ inputs, and identify an imprint-and-shift mechanism driving knowledge updating in FFNs. Finally, an efficient editing method driven by the linearization above is evaluated, showing strong performances in simple editing settings.

Note The authors introduce a revisited psycholinguistic benchmark to evaluate the effectiveness and reliability of intervention-based mechanistic interpretability methods across several linguistic tasks. Across several model sizes, Distributed alignment search (DAS) and probing are found to be the most reliable approaches, and are used to investigate the emergence of features linked to linguistically plausible predictions in the initial phases of model training.

Note This work introduces InfICL, a demonstration selection method using influence functions to identify salient training examples to use as demonstrations at inference time. InfICL is tested alongside other examples selection baselines for prompting medium-sized LLMs for COLA and RTE, showing improvements over other methods especially when a smaller number of in-context examples is used.

Note This paper introduces SpectralDeTuning, an method to recover original pre-trained weights of a model from a set of LoRA fine-tunes with merged weights. Authors introduce the LoWRA Bench dataset to measure progress in this task, and show that the method performs well for both language and vision models. The current limitations of the approach are 1) assuming the attacker knowledge of the rank used in LoRAs and 2) need for a good amount of LoRAs to reconstruct the original pre-training effectively

Note Authors propose SyntaxSHAP, a variant of the model-agnostic SHAP approach enforcing tree-based coalition based on the syntax of the explained sequence, while preserving most properties of SHAP explanations. The approach is found to be more faithful and semantically meaningful than other model-agnostic methods when explaining the predictions of LMs such as GPT-2 and Mistral, especially in edge cases such as negation.

Note Authors propose a fine-tuning procedure with natural language explanation to clarify intermediate reasoning steps. Several LMs are fine-tuned on ListOps dataset, containing synthetically-generated instructions on sequences of numbers. Authors find that explanations improve model performances across all tested model sizes and explanations lengths. Smaller language models benefit the most from explanations, especially when long-form.

Note This works introduces a model editing approach using individual “canonical” examples to showcase desired/unwanted behavior. The approach is tested on regular LMs and Backpack LMs, which are more controllable thanks to disentangled sense vector representations. For the latters, authors propose sense fine-tuning, i.e. updating few sense vectors with canonical examples to apply desired changes in an efficient and effective way, outperforming other model editing approaches and full/LoRa fine-tuning.

Note This work proposes extending the LRP feature attribution framework to handling Transformers-specific layers. Authors show that AttnLRP is significantly more faithful than other popular attribution methods, has minimal time requirements for execution and can be employed to identify model components associated to specific concepts in generated text.

Note This work discusses the dichotomy between faithfulness and plausibility in LLMs’ self-explanations (SEs) employing natural language (CoT, counterfactual reasoning, and token importance), which tend to be plausible but unfaithful to models' reasoning process. Authors call for a community effort to 1) develop reliable metrics to characterize the faithfulness of explanations and 2) pioneering novel strategies to generate more faithful SEs.

Note While most internals-based hallucination detection methods use surface-level information, this work proposes EigenScore, an internal measure of responses’ self-consistency using the eigenvalues of sampled responses' covariance matrix in intermediate model layers to quantify answers’ diversity in the dense embedding space. EigenScore outperforms logit-level methods for hallucination detection on QA tasks, especially with feature clipping to control overconfident generations.

Note In this opinion piece, authors contend that the new capabilities of LLMs can transform the scope of interpretability, moving from low-level explanations such as saliency maps and circuit analysis to natural language explanations. This goal is hindered by LM’s natural tendency to hallucinate, their large size and their inherent opaqueness. Authors highlight in particular dataset explanations for knowledge discovery, explanations’ reliability and interactive explanations as key areas moving ahead.

Note Authors propose Recursive Attribution Generation (ReAGent), a perturbation-based feature attribution approach specifically conceived for generative LMs. The method employs a lightweight encoder LM to replace sampled input spans with valid alternatives and measure the effect of the perturbation on the drop in next token probability predictions. ReAGent is shown to consistentlyoutperform other established approaches across several models and generation tasks in terms of faithfulness.

Note This work introduces a new methodology for human verification of reasoning chains and adopts it to annotate a dataset of chain-of-thought reasoning chains produced by 3 LMs. The annotated dataset, REVEAL, can be used to benchmark automatic verifiers of reasoning in LMs. In their analysis, the authors find that LM-produced CoTs generally contain faulty steps often leading to wrong automatic verification.

Note This work proposes Gradient-Based Red Teaming (GBRT), a red teaming method for automatically generating diverse prompts inducing an LM to output unsafe responses. In practice, prompts are learned by scoring LM responses with a safety-trained probing classifier, and back-propagating through frozen classifier and LM to update the prompt. GBRT prompts are shown to be more likely to generate unsafe responses and evade safety-tuning measures than those produced by RL-based methods.

Note Audits conducted on AI systems can identify potential risks and ensure their compliance to safety requirements. Authors categorise audits based on the access to model-related resources and highlight how levels of transparency on audited AI system enable broader and more effective auditing procedures. Technical, physical, and legal safeguards for performing audits are also introduced to ensure minimal security risks for audited companies.

Note This work evaluates the human confidence in LLM responses to multiple-choice MMLU questions based on explanations the LLM provides together with selected answers. The authors experiment with altering the model prompt to reflect the actual prediction confidence in models’ explanations, showing improved calibration for users’ assessment of LLM’s reliability and a better ability to discriminate between correct and incorrect answers.

Note This work methodically evaluates of in-context learning on formal languages across several model architectures, showing how Transformers work best in this setting. These results are attributed to the presence of “n-gram heads” able to retrieve the token following a context already seen in the current context window and copy it. These insights are used to design static attention layers mimicking the behavior of n-gram head, leading to lower perplexity despite the lower computational cost.

Note This survey summarizes recent works in interpretability research, focusing mainly on pre-trained Transformer-based LMs. The authors categorize current approaches as either local or global and discuss popular applications of LM interpretability, such as model editing, enhancing model performance, and controlling LM generation.

Note Authors introduce LLMCheckup, a conversational interface connecting an LLM to several interpretability tools (feature attribution methods, similarity, counterfactual/rationale generation) allowing users to inquire about LLM predictions using natural language. The interface consolidates several interpretability methods in a unified chat interface, simplifying future investigations into natural language explanations.

Note This work investigates the universality of individual neurons across GPT2 models trained from different initial random seeds, starting from the assumption that such neurons are likely to exhibit interpretable patterns. 1-5% of neurons consistently activate for the same inputs, and can be grouped into families exhibiting similar functional roles, e.g. modulating prediction entropy, deactivating attention heads, and promoting/suppressing elements of the vocabulary in the prediction.

Note This study uses self-consistency checks to measure the faithfulness of LLM explanations: if an LLM says a set of words is important for making a prediction, then it should not be able to make the same prediction without these words. Results demonstrate that LLM self-explanations faithfulness of self-explanations cannot be reliably trusted, as they prove to be very task and model dependent, with bigger model generally producing more faithful explanations.

Note Authors introduce a new taxonomy for fine-grained annotation of hallucinations in LM generations and propose Factuality Verification with Augmented Knowledge (FAVA), a retrieval-augmented LM fine-tuned to detect and edit hallucinations in LM outputs, outperforming ChatGPT and LLama2 Chat on both detection and editing tasks.

Note Patchscopes is a generalized framework for verbalizing information contained in LM representations. This is achieved via a mid-forward patching operation inserting the information into an ad-hoc prompt aimed at eliciting model knowledge. Patchscope instances for vocabulary projection, feature extraction and entity resolution in model representation are show to outperform popular interpretability approaches, often resulting in more robust and expressive information.

Note This work raises concerns that gains in factual knowledge after model editing can result in a significant degradation of the general abilities of LLMs. Authors evaluate 4 popular editing methods on 2 LLMs across eight representative tasks, showing model editing does substantially hurt model general abilities.

Note This work systematically examines the impact of methodological details in activation patching, a popular technique with causal guarantees to quantify the importance of model components in driving model predictions. Authors' provide several recommendations concerning the type of patching (noise vs. counterfactual), the metric to use (probability vs. logit vs. KL), the number of layers to patch and which tokens to corrupt.