Understanding which parameters are influential in Large Language Models (LLMs) is central to improving their efficiency, reliability, and interpretability. We introduce Weight-Adjusted Gradients (WAG), a simple yet effective approach for estimating parameter importance that explicitly captures the interaction between model weights and first-order gradient information and identifies parameters that disproportionately influence model behavior, such as those responsible for collapse phenomena in LLMs. Across a range of models and settings, we show that WAG surfaces a tiny but critical subset of parameters whose modification leads to dramatic degradation in performance, a failure mode that existing importance metrics overlook. These findings reveal a previously underexplored interplay between weights and gradients, suggesting that parameter importance cannot be fully understood through either signal alone. The surprising effectiveness of WAG points to fundamental structural properties of trained networks and motivates new open questions about the role of zeroth-order and first-order information in deep learning. We demonstrate the practical utility of WAG across multiple applications, including expert allocation in mixture-of-expert architectures, parameter-specific unlearning, mixed-precision quantization, and layer selection for knowledge editing. Our results position WAG as a unified approach for analyzing, debugging, and controlling LLMs, and opens new directions for principled model-level interpretation.
Results support IAIML as a compact, interaction-aware framework appropriate for settings where bounded explanation size and controlled treatment of feature interactions are design requirements.
Causal self-attention models such as SASRec are widely used in sequential recommendation, where learned attention weights are often assumed to provide crucial importance weighting over past interactions. Yet it is unclear when predictive performance truly depends on such non-uniform weighting. We study a controlled SASRec variant that replaces learned attention weights with uniform aggregation and is trained under an otherwise identical block structure and training recipe. Across fourteen benchmark datasets, this modification often yields performance comparable to the original model, with clear dataset-dependent exceptions. To explain this heterogeneity, we introduce a stage-wise norm-based decomposition that quantifies self-preserving vs. cross-position mixing within attention blocks. Across datasets, we find distinct regimes: low mixing yields robustness to uniformization; higher mixing tends to coincide with sensitivity, while some datasets exhibit substantial mixing without dependence on learned weighting. Our results provide a practical diagnostic for identifying when attention weighting is functionally utilized in sequential recommendation. The code is available at: https://github.com/keito0329/revisiting-sasrec.
K.Ozaki K.Ozaki, Keigo Sakurai, Ren Togo et al.· Annual International ACM SIG...· 0 citations
Variable importance may describe either intrinsic predictive information in a population or extrinsic importance for a fitted prediction rule. Quantifying the uncertainty in variable importance estimates is critical for interpretation. Methods for estimating intrinsic variable importance (we will refer to these as VIMP) and the minipatch leave-one-covariate-out procedure (MPLOCO) target intrinsic and extrinsic importance, respectively, and provide methods for computing standard errors. These two approaches have a shared structure, comparing prediction performance with and without features, but the relationship between them has not been formally characterized. We establish conditions under which the two perspectives align. Under squared-error loss, if the fitted full and reduced learners converge to their oracle counterparts sufficiently fast, then MPLOCO is asymptotically equivalent to VIMP. We provide further conditions extending this result to general loss functions and formalize grouped MPLOCO for potentially overlapping feature groups. Through simulations, we show that VIMP and MPLOCO agree most closely when the fitted learner is well aligned with the data-generating mechanism. In a high-dimensional grouped simulation, both procedures identified the signal-containing groups. In an analysis of HIV-1 VRC01 neutralization sensitivity, both methods placed the same three biologically relevant feature groups among their highest-ranked groups. These results clarify when intrinsic and extrinsic importance can be interpreted similarly and when they provide complementary information.
A central goal of mechanistic interpretability is to understand how neural networks work and what each individual component does. Dominant circuit-finding approaches focus on a specific behavior and reverse-engineer the role of components on the associated sub-distribution. However, past work has shown that components can have different functions that are active on different subsets of the input distribution. In this work we ask whether a single weight can be understood globally across the full training distribution by characterizing when it matters (the inputs on which ablating it changes the model's predictions). We introduce an automated LLM pipeline that writes a short, human-readable description of when a weight matters and verifies it on held-out text, crediting a weight only if its description generalizes. Across two sparse and two dense transformers, the fraction of weights that are interpretable (in this sense) is higher in sparse transformers than in dense ones, a gap that widens once unreliable descriptions are discarded. Our results show that a meaningful fraction of a sparse transformer model's weights can be interpreted: 12 to 31% of weights have a single short description that identifies what the weight is used for.
Recent work identified Super Weights, individual parameters whose removal degrades model performance by orders of magnitude. We show that this degradation due to pruning Super Weights does not universally apply to all LLMs. Furthermore, if these parameters are so important, Super Weight-aware training should be effective. We show the opposite. Training Super Weights in isolation (100 to 8,192 parameters) drops accuracy to random-guessing levels on both OLMo-1B and OLMo-7B, and expanding to local neighborhoods of up to 36K parameters provides no improvement. The failure is specific to Super Weight coordinates: training an equal number of randomly chosen positions in the same down_proj layers instead improves over the baseline, so the collapse comes from targeting Super Weights, not from sparsity itself. Vanilla LoRA, updating every position in attention weight matrices through low-rank structure, succeeds with only 0.16% of parameters, and applying the same low-rank update to down_proj succeeds as well. A 10-seed ablation confirms that constraining LoRA updates at positions corresponding to Super Weight coordinates yields statistically indistinguishable results. These findings establish that parameter importance does not imply parameter trainability in isolation, and that effective fine-tuning relies on structured decompositions over entire layers rather than targeting individually important weights.
S. Subramanian, Adewale Akinfaderin, Akarsha Sehwag· 0 citations
Benchmark scores are reported as properties of a model, yet the inference framework used to produce them, such as HuggingFace, vLLM, or Ollama, are considered non-influential and their names and versions are almost never disclosed. In this work we investigate how much this choice can influence the model output. In a fully-crossed study (three instruction-tuned models x five inference frameworks x six benchmarks x four generation modes) we investigate how different tools (wrappers/backend) influence benchmark scores and how their score changes is influenced by generation hyper-parameters. We find backend to be a non-negligible factor where even under greedy, sampling-noise-free decoding, changing the backend can significantly alter models performance and this effect is structural and strongly model-dependent. Decomposing the variance according to generation mode reveal that considerable portion of the variability (roughly 39\%) a practitioner sees out-of-the-box can stem from the backend, while the remaining stems from sampling noise and each framework's default generation parameters, both of which are avoidable by disclosing and matching the generation configuration. These divergences are more pronounced on factual than on social-bias benchmarks. Overall, benchmark numbers are not backend-agnostic therefore, we recommend disclosing the backend, its version, and the full generation configuration, also using deterministic decoding for cross-backend comparison.
Shahed Masoudian, Passant Shafaei, Monorama Swain et al.· 0 citations