Searcharxiv⌕ Search

arXiv subjects

Hassan Sajjad

Publications and source records attributed to Hassan Sajjad.

At least 19 recordsLinked to original sources

Understanding AI from the Inside: The State of Mechanistic Interpretability for AI Safety

Neural networks are likely to be among the most consequential artefacts in human history. As these systems become more capable and autonomous, their potential to cause societal harm also grows, making AI safety one of the defining challenges of our time. Yet we continue to build increasingly powerful systems without fully understanding how they work, how they will behave in unfamiliar situations, or how to reliably control them. Mechanistic interpretability offers a path towards understanding these systems by studying their internal representations and computations, with the aim of evaluating, predicting, and controlling behaviour. We survey mechanistic interpretability for AI safety in order to answer the question: Does understanding how models work actually make them safer? We place mechanistic interpretability in the broader context of interpretability research and compare its mechanistic evidence with behavioural evaluation. We introduce the foundations of mechanistic interpretability, including features, polysemanticity, transformer computation, and circuits, and survey methods for feature extraction and analysis, circuit discovery and causal verification, explaining intermediate representations in natural language, learning dynamics, and training-data attribution. Beyond surveying existing methods, we propose a unified framework for evaluating mechanistic interpretability claims and characterise six research orientations by their goals, assumptions, and trade-offs. Finally, we examine how mechanistic understanding can support AI safety, including knowledge editing, steering, and monitoring, and assess its limitations, failure modes, open problems, and relationship to behavioural safety measures. We also release a repository of code notebooks with worked examples that reproduce key findings from surveyed methods, allowing readers to build intuition through hands-on experimentation.

cs.CL↗

TRACE: Temporal Risk Assessment and Concept Editing for Safer Language Generation

Large language models (LLMs) can produce harmful content while generating a response. A useful safeguard should not only detect this risk during generation but also intervene while leaving safe responses largely unchanged and provide an auditable record of the internal patterns supporting its decisions. We propose TRACE (Temporal Risk Assessment and Concept Editing), a framework that tracks harmful-response risk during generation, selectively edits token activations, and leaves an auditable trace of its detection and steering decisions. TRACE learns a codebook of recurring activation patterns, with each codebook vector representing a latent concept. At inference time, it assigns each token activation to the nearest codebook vector to obtain the corresponding concept. A gated recurrent unit (GRU) processes the activations and their assigned concept features to predict completed-response harmfulness and a monotone cumulative streaming risk. During steering, activations assigned to harmful concepts are steered along directions defined by nearby benign targets. An external LLM labels the learned concepts post hoc, making detection sequences and harmful-to-benign edit records interpretable. Experiments across three generators and the WildGuard and S-Eval benchmarks show that TRACE improves streaming F1 over Kelp in five of six comparisons while remaining competitive on completed-response detection. TRACE's selective steering reduces Harmbench attack success with little change in refusal on originally safe responses. The post-hoc audit reveals recurring semantic structure in the learned concepts and distinct traces associated with harmful detection, safe generation, and activation edits.

cs.AI↗

Instance-Level Difficulty: A Missing Perspective in Machine Unlearning

Current research on deep machine unlearning primarily focuses on improving or evaluating the overall effectiveness of unlearning methods while overlooking the varying difficulty of unlearning individual training samples. As a result, the broader feasibility of machine unlearning remains under-explored. This paper studies the cruxes that make machine unlearning difficult through a thorough instance-level unlearning performance analysis over various unlearning algorithms and datasets. In particular, we summarize four factors that make unlearning a data point difficult, and we empirically show that these factors are independent of a specific unlearning algorithm but only relevant to the target model and its training data. Given these findings, we argue that machine unlearning research should pay attention to the instance-level difficulty of unlearning.

cs.LG↗

Removing Information Content Does Not Certify Tamper Resistance in Open-Weight Models

Does removing harmful information make open-weight models resistant to fine-tuning attacks? We show that mutual information at release alone cannot universally certify slow recovery. Function-preserving reparameterizations leave information unchanged while altering gradient-descent geometry, so an invariant certificate is bounded by the fastest reachable parameterization. We apply this principle to weight--data mutual information under training-data filtering and label--representation mutual information under capability removal. Training order can change recovery time at fixed weight--data information, while exact representation-level independence can preserve the entire parameter Jacobian. An explicit construction has both information quantities equal to zero and recovers in one gradient step. Controlled experiments illustrate order-dependent recovery and parameterization-dependent attack speed. These results identify the missing requirement for certification: constraints on attack dynamics beyond mutual information at release.

cs.LG↗

Cross-Layer Discrete Concept Discovery for Interpreting Language Models

Interpreting language models remains challenging due to the existence of residual stream, which linearly mixes and duplicates features across adjacent layers, causing single-layer analyses to miss this cross-layer structure. Cross-layer sparse autoencoders (SAEs) address layer mixing but operate in continuous space, where concepts split across many neurons without clear boundaries. We introduce Cross-Layer Vector Quantized-Variational Autoencoder (CLVQ-VAE), a novel framework which maps representations from a lower layer to a higher layer through a discrete vector-quantization bottleneck, collapsing duplicated residual-stream features into compact, interpretable concept vectors. Our approach combines top-k temperature-based sampling with exponential moving average (EMA) codebook updates, providing controlled exploration of the discrete latent space while maintaining codebook diversity. Across both encoder- and decoder-based models on ERASER-Movie, Jigsaw, and AGNews, CLVQ-VAE outperforms clustering, single-layer vector quantized-variational autoencoder (VQ-VAE), and sparse autoencoder (SAE) baselines across three evaluation axes: removing identified concepts drops downstream probe accuracy by up to 93%, LLM judges rank our concepts first in 66.7% of comparisons, and human annotators recover model predictions from our visualizations with 78% accuracy versus 54% for clustering.

cs.LG↗

Multilingual in Name Only? Cultural and Linguistic Weaknesses of LLMs in Urdu

Multilingual large language models (LLMs) are increasingly used for open-ended text generation, yet their behaviour in low-resource languages remains poorly understood. In this work, we question how correct and reliable is the generation of multilingual LLMs when used for the task of story generation. We consider Urdu language as a representative low-resource language. We generate Urdu-Stories, a corpus of 93 stories generated using three contemporary LLMs (GPT-5.1, Qwen-3-Max, DeepSeek-3.1). We manually annotate the errors present in them under a nine-label linguistic, semantic, and cultural taxonomy. Our notable findings suggest that LLMs often make basic errors of grammar and semantics. The stories lack coherence, have unnatural repetition and show pervasive cultural shallowness. We further show using few-shot prompting that the cultural and context errors largely remain unresolved. Our findings highlight the limitations of current LLMs as a reliable source of content generation and information retrieval for low-resource languages.

cs.CL↗

GAPS: Dimension-Level Gates for Conditional Activation Steering

Activation steering suppresses undesired behaviors in language models by adding a steering vector to the hidden state during generation. Recent conditional methods such as CAST and DSAS improve the behavior-capability trade-off by deciding when to intervene, but once active, they apply the full dense vector to all hidden dimensions, regardless of whether a neuron carries concept information or already lies in the desired regime. We introduce dimension-level conditioning as a complementary axis of selectivity that also decides which neurons to intervene on. Our method, GAPS (Gated Activation steering via Posterior and Separability), combines two training-free gates: a static separability gate that restricts steering to neurons with statistically reliable concept information (via AUROC), and a dynamic posterior gate that steers a neuron only when its current activation is better explained by the undesired concept under a Gaussian model. The gates add O(D) overhead per token, and they plug into existing conditional methods. On toxicity mitigation (RealToxicityPrompts) and concept removal (OneSeC) with Gemma-3 (4B) and Qwen-3 (1.7B), GAPS consistently matches or improves the Pareto front of its token-level counterparts; under a fixed capability budget, DSAS+GAPS reduces Gemma-3's toxicity rate from 6.52% to 0.48%, versus 3.52% for DSAS alone. Ablations attribute most of the gain to the posterior gate.

cs.CL↗

SCOPE: Selective Conformal Optimized Pairwise LLM Judging

Large language models (LLMs) are increasingly used as scalable judges in pairwise evaluation, but they remain prone to miscalibration and biases. We propose \textsc{Scope} (Selective Conformal Optimized Pairwise Evaluation), a framework that calibrates an acceptance threshold so that, under exchangeability, the error rate among non-abstained judgments is at most a user-specified level $α$. To supply \textsc{Scope} with a bias-neutral uncertainty signal, we introduce Bidirectional Preference Entropy (BPE), which queries the judge under both response positions and converts the order-averaged preference probability into an entropy-based score. Across various pairwise judging benchmarks, BPE outperforms standard confidence proxies in calibration and discrimination, while \textsc{Scope} consistently satisfies the target risk bound (empirical FDR $\approx 0.097$--$0.099$ at $α=0.10$) and retains substantial coverage. Compared to vanilla baselines, \textsc{Scope} accepts up to $2.4\times$ more judgments under the same risk constraint, demonstrating that BPE enables reliable and high-coverage LLM-based evaluation.

cs.CL↗

Judge, Retrieve, or Abstain: Uncertainty-Guarded LLM Judging with Provable Risk Guarantees

Using LLMs as judges has become standard practice for evaluating model outputs at scale. This is particularly common for subjective, open-ended tasks such as assessing helpfulness or alignment, where no single reference answer exists. However, objective tasks introduce a distinct reliability challenge for reference-free LLM judging. In the absence of a reference answer, the judge evaluates factual correctness either through its parametric knowledge or through tool augmentation. Although the former enables efficient evaluation, the judge may hallucinate or lack sufficient evidence for its verdict. Conversely, tool augmentation can provide additional evidence but introduces extra computational cost and requires an appropriate mechanism to determine when and how that evidence should be used reliably. More importantly, neither approach alone provides formal control over the risk of accepted verdicts or guarantees their reliability at a specified level. We propose a risk-controlled framework that calibrates uncertainty thresholds on a held-out set so that the false discovery rate among accepted verdicts remains below a user-specified level~$α$ with high probability, using finite-sample Clopper--Pearson intervals. When the parametric mode is not sufficiently confident, the instance is routed to a retrieval-augmented mode, where the judge gathers web evidence and re-evaluates the instance under a second calibrated threshold. The finite-sample guarantee carries over to this two-threshold routing without additional assumptions. Across open-domain QA benchmarks and judges of varying scales, the framework maintains the target error rate while achieving substantially higher coverage than single-mode baselines.

cs.CL↗

Multi-Granular Node Pruning for Causal Circuit Discovery

Circuit discovery aims to identify minimal subnetworks that are responsible for specific behaviors in large language models (LLMs). Existing approaches primarily rely on iterative edge pruning, which is computationally expensive and limited to coarse-grained units such as attention heads or MLP blocks, overlooking finer structures like individual neurons. We propose a node-level pruning framework for circuit discovery that addresses both scalability and granularity limitations. Our method introduces learnable masks across multiple levels of granularity, from entire blocks to individual neurons, within a unified optimization objective. Granularity-specific sparsity penalties guide the pruning process, allowing a comprehensive compression in a single fine-tuning run. Empirically, our approach identifies circuits that are smaller in nodes than those discovered by prior methods; moreover, we demonstrate that many neurons deemed important by coarse methods are actually irrelevant, while still maintaining task performance. Furthermore, our method has a significantly lower memory footprint, 5-10x, as it does not require keeping intermediate activations in the memory to work.

cs.AI↗

LangFIR: Discovering Sparse Language-Specific Features from Monolingual Data for Language Steering

Large language models (LLMs) show strong multilingual capabilities, yet reliably controlling the language of their outputs remains difficult. Representation-level steering addresses this by adding language-specific vectors to model activations at inference time, but identifying language-specific directions in the residual stream often relies on multilingual or parallel data that can be expensive to obtain. Sparse autoencoders (SAEs) decompose residual activations into interpretable, sparse feature directions and offer a natural basis for this search, yet existing SAE-based approaches face the same data constraint. We introduce LangFIR (Language Feature Identification via Random-token Filtering), a method that discovers language-specific SAE features using only a small amount of monolingual data and random-token sequences. Many SAE features consistently activated by target-language inputs do not encode language identity. Random-token sequences surface these language-agnostic features, allowing LangFIR to filter them out and isolate a sparse set of language-specific features. We show that these features are extremely sparse, highly selective for their target language, and causally important: directional ablation increases cross-entropy loss only for the corresponding language. Using these features to construct steering vectors for the multilingual generation control task, LangFIR achieves the best average accuracy x BLEU among steering methods across three models (Gemma 3 1B, Gemma 3 4B, and Llama 3.1 8B), three datasets, and twelve target languages, outperforming the strongest monolingual baseline by up to 4.7x and surpassing methods that use parallel data. Our results suggest that language identity in multilingual LLMs is localized in a sparse set of feature directions discoverable with monolingual data. Code is available at https://github.com/JaMussCraft/LangFIR

cs.CL↗

Distribution-Specific Curvature Control with Finite-Sample Guarantees for Open-Weight Safety

A short fine-tuning run can undo the safety guards of an open-weight model---retraining a refusal-trained assistant to aid weapons development or produce hate speech. Preventing such harmful fine-tuning while retaining benign adaptability remains difficult: the only prior method with an explicit curvature certificate, spectral deformation, inflates curvature globally and thereby obstructs benign adaptation along with harmful adaptation. We propose HarmAlign, which applies function-preserving spectral deformation along a estimated contrastive activation subspace. We derive finite-sample bounds for the estimated subspace energy and the resulting local harmful-distribution curvature lower bound. A stability--progress dichotomy for constant-step gradient descent turns the certified curvature into conditional convergence-rate control. Empirically, within a fixed-architecture, finite-budget first-order threat model, HarmAlign blocks direct fine-tuning and three data- or objective-adaptive attacks across a hazardous-knowledge relearning setting and a harmful-assistance fine-tuning setting, while the protected benign tasks remain trainable. The block persists across the tested first-order optimizer variants over every attack checkpoint, and under out-of-distribution harmful fine-tuning, and it extends to important cases in our threat model: accidental safety degradation and emergent misalignment.

cs.LG↗

SAGE: A Search-AuGmented Evaluation of Large Language Models on Free-Form QA

As Large Language Models (LLMs) become increasingly used for question-answering (QA), relying on static, pre-annotated references for evaluation poses significant challenges in cost, scalability, and completeness. Meanwhile, using LLMs themselves as evaluators without external grounding remains unreliable for objective tasks, as they systematically over-accept incorrect answers, fabricate supporting rationales, and degrade sharply on questions that fall outside their training data. We propose Search-AuGmented Evaluation (SAGE), a framework to assess LLM outputs without fixed ground-truth answers. Unlike conventional metrics that compare to static references or depend solely on LLM-as-a-judge knowledge, SAGE acts as an agent that actively retrieves and synthesizes external evidence. It iteratively generates web queries, collects information, summarizes findings, and refines subsequent searches through reflection. By reducing dependence on static reference-driven evaluation protocols, SAGE offers a scalable and adaptive alternative for evaluating the factuality of LLMs. Experimental results on multiple free-form QA benchmarks show that SAGE achieves substantial to perfect agreement with human evaluations.

cs.CL↗

Vector Quantized Latent Concepts: A Scalable Alternative to Clustering-Based Concept Discovery

Large language models (LLMs) encode rich semantic information in their hidden states, yet it remains difficult to understand what information these internal representations capture. Latent concepts extracted from hidden states offer a promising direction for interpreting LLMs, but existing clustering-based methods face a trade-off: hierarchical clustering produces coherent concepts but is limited to small datasets due to its quadratic memory cost, while K-Means scales efficiently but may yield less semantically coherent concepts. We propose Vector Quantized Latent Concept (VQLC), a discrete concept learning framework that learns a codebook of latent concepts on frozen hidden states. Across 12 dataset-model settings, VQLC stays close to K-Means in computational cost, scales better than hierarchical clustering, and remains competitive in faithfulness, with the clearest gains on decoder-only models. LLMs-based evaluation, qualitative analysis, and a Sparse Autoencoder (SAE) comparison demonstrate that the learned concepts are interpretable and task-relevant.

cs.LG↗

On the Persistent Effects of Lexicality in Large Language Models

Representations extracted from large language models (LLMs) play an important role in many downstream applications. However, the structure of these representations is often influenced by lexical overlap rather than semantic content. Our understanding of the relationship between this lexical influence and semantic content, and its implications for downstream tasks, remains limited. In this work, we investigate representations to quantify the effect of lexical overlap relative to semantic content. We consider several adversarial semantic stress tests and further connect our findings to the information theory perspective. We find that lexical influence extends across the depth of models, consistently across architectures, training regimes, and objective functions, including the models trained for semantic similarity. Moreover, we observe a mid-depth region in which both lexical and semantic signals degrade simultaneously, indicating a transitional regime where representations are poor for both surface form and meaning. We further demonstrate the effect of lexical influence on downstream uses of LLMs using summarization and model editing as a case study.

cs.CL↗

Neurons Speak in Ranges: Breaking Free from Discrete Neuronal Attribution

Pervasive polysemanticity in large language models (LLMs) undermines discrete neuron-concept attribution, posing a significant challenge for model interpretation and control. We systematically analyze both encoder and decoder based LLMs across diverse datasets, and observe that even highly salient neurons for specific semantic concepts consistently exhibit polysemantic behavior. Importantly, we uncover a consistent pattern: concept-conditioned activation magnitudes of neurons form distinct, often Gaussian-like distributions with minimal overlap. Building on this observation, we hypothesize that interpreting and intervening on concept-specific activation ranges can enable more precise interpretability and targeted manipulation in LLMs. To this end, we introduce NeuronLens, a novel range-based interpretation and manipulation framework, that localizes concept attribution to activation ranges within a neuron. Extensive empirical evaluations show that range-based interventions enable effective manipulation of target concepts while causing substantially less collateral degradation to auxiliary concepts and overall model performance compared to neuron-level masking.

cs.LG↗

LLMs Underperform Graph-Based Parsers on Supervised Relation Extraction for Complex Graphs

Relation extraction represents a fundamental component in the process of creating knowledge graphs, among other applications. Large language models (LLMs) have been adopted as a promising tool for relation extraction, both in supervised and in-context learning settings. However, in this work we show that their performance still lags behind much smaller architectures when the linguistic graph underlying a text has great complexity. To demonstrate this, we evaluate four LLMs against a graph-based parser on six relation extraction datasets with sentence graphs of varying sizes and complexities. Our results show that the graph-based parser increasingly outperforms the LLMs, as the number of relations in the input documents increases. This makes the much lighter graph-based parser a superior choice in the presence of complex linguistic graphs.

cs.CL↗

Limits of Convergence-Rate Control for Open-Weight Safety

Open-weight foundation models can be fine-tuned for harmful purposes after release, yet no existing training resistance methods provide theoretical guarantees. Treating these interventions as convergence-rate control problems allows us to connect optimization speed to the spectral structure of model weights. We leverage this insight to develop a novel understanding of convergence rate control through spectral reparameterization and derive an algorithm, SpecDef, that can both provably and empirically slow first- and second-order optimization in non-adversarial settings. In adversarial settings, we establish a fundamental limit on a broad class of convergence rate control methods including our own: an attacker with sufficient knowledge can restore fast convergence at a linear increase in model size. In order to overcome this limitation, future works will need to investigate methods that are not equivalent to controlling convergence rate.

math.OC↗