{"total":21,"items":[{"citing_arxiv_id":"2606.00959","ref_index":49,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Towards Understanding Modality Interaction in Multimodal Language Models via Partial Information Decomposition","primary_cat":"cs.AI","submitted_at":"2026-05-31T02:29:36+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":5.0,"formal_verification":"none","one_line_summary":"PID applied to MLLMs identifies task-specific modality interaction profiles that generalize across models, extend to tri-modal cases, and yield initial performance gains via reweighting.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.07604","ref_index":30,"ref_count":2,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Contribution Weights: A Geometrical Analysis of Self-Attention Transformers","primary_cat":"cs.LG","submitted_at":"2026-05-29T09:40:38+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"Contribution Weights combine attention, value magnitude, and directional alignment to measure token influence more faithfully than attention alone, and show attention sinks actively suppress information via a convex sink-rate to output-norm relationship.","context_count":1,"top_context_role":"background","top_context_polarity":"background","context_text":"The previous empirical results are telling us that both ∥ ˆwi∥ and ⟨ ˆwi,˜v0⟩ are approximately constant. Since ˆwi is the expected value under pi,· this roughly follows from other observations on the ˜vj, namely that their norms are of the same magnitude and that their dot product with˜v0 is constant (and negative). Note that the approximate equality ||oi|| ≃ p α2||˜v0||2 + 2α(1−α)γ+ (1−α) 2κ2 (30) 17 Contribution Weights 2 5 8 11 14 17 20 23 26 29 Layer 0.5 0.6 0.7 0.8 0.9 1.0 R2 (∥w∥vs α) Mean(median) = 0.822 R2 for ∥w∥vs Sink Rate 2 5 8 11 14 17 20 23 26 29 Layer −0.2 0.0 0.2 0.4 0.6 0.8 1.0 R2 (⟨˜v1,w⟩vs α) Mean(median) = 0.729 R2 for Dot Product vs Sink Rate Mean Median Quartiles (25th-75th) 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Sink Rate [ α (l,h)"},{"citing_arxiv_id":"2605.19848","ref_index":11,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"CLIF: Concept-Level Influence Functions for Transparent Bottleneck Models","primary_cat":"cs.CL","submitted_at":"2026-05-19T13:42:38+00:00","verdict":null,"verdict_confidence":null,"novelty_score":null,"formal_verification":null,"one_line_summary":null,"context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.14255","ref_index":9,"ref_count":2,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Architecture-Aware Explanation Auditing for Industrial Visual Inspection","primary_cat":"cs.LG","submitted_at":"2026-05-14T01:48:00+00:00","verdict":null,"verdict_confidence":null,"novelty_score":null,"formal_verification":null,"one_line_summary":null,"context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.13434","ref_index":21,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Rescaled Asynchronous SGD: Optimal Distributed Optimization under Data and System Heterogeneity","primary_cat":"cs.LG","submitted_at":"2026-05-13T12:27:22+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"Rescaled ASGD recovers convergence to the true global objective by rescaling worker stepsizes proportional to computation times, matching the known time lower bound in the leading term under non-convex smoothness and bounded heterogeneity.","context_count":1,"top_context_role":"background","top_context_polarity":"background","context_text":"converge faster, albeit to the wrong objective function ˆF. Under Assumption 3.1, if simultaneous updates are processed in ascending order of worker indices, the iteration-specific delay δk for every update delivered by worker i remains constant across all iterationskwithi k =i. With a slight abuse of notation, the worker-specific delay is then δi = X j̸=i τi τj =τ i n τH −1.(21) Substituting (21) into (20), we find Γi =γ τmaxτH n 1 τ 2 i . With this, the cycle stepsize and sum of squared stepsizes become α=γ τmaxτH n nX i=1 τ −2 i , A=γ 2 τmaxτ 2 H n2 nX i=1 τ −3 i . 29 Following the same steps as in Sections B.3 and B.4, we can now derive the cycle and wall-clock time complexities. Under the same assumptions as in Theorem B."},{"citing_arxiv_id":"2605.18849","ref_index":15,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"INSIGHTS: Demonstration-Based Summaries of Time Series Predictors","primary_cat":"cs.LG","submitted_at":"2026-05-13T08:17:35+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":5.0,"formal_verification":"none","one_line_summary":"INSIGHTS creates manageable global summaries of time series model behavior by balancing sample importance and diversity with domain-specific utility functions, validated via experiments and user studies.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.12809","ref_index":39,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Correcting Influence: Unboxing LLM Outputs with Orthogonal Latent Spaces","primary_cat":"cs.LG","submitted_at":"2026-05-12T23:01:29+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"A latent mediation framework with sparse autoencoders enables non-additive token-level influence attribution in LLMs by learning orthogonal features and back-propagating attributions.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.06032","ref_index":297,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"Does Synthetic Data Help? Empirical Evidence from Deep Learning Time Series Forecasters","primary_cat":"cs.LG","submitted_at":"2026-05-07T11:22:45+00:00","verdict":"ACCEPT","verdict_confidence":"MODERATE","novelty_score":7.0,"formal_verification":"none","one_line_summary":"Synthetic data augmentation helps channel-mixing time series models but degrades channel-independent ones, with reliable gains only from seasonal-trend generators and gradual schedules in low-resource settings.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.05668","ref_index":12,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"Large Vision-Language Models Get Lost in Attention","primary_cat":"cs.AI","submitted_at":"2026-05-07T04:45:52+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"In LVLMs, attention can be replaced by random Gaussian weights with little or no performance loss, indicating that current models get lost in attention rather than efficiently using visual context.","context_count":1,"top_context_role":"background","top_context_polarity":"background","context_text":"The default settings are highlighted in bold red. General Reasoning Model Layers VQAv2VizWiz OKVQAMMVetA-OKVQAMMStarSEED SciQA RWQA MMMU GQA LLaVA-1.5-7B[2, 7] 46.30 42.40 49.88 41.73 47.53 31.83 42.22 38.69 28.21 28.12 46.45 [2, 13] 36.81 34.31 39.26 28.78 32.47 26.37 30.86 36.20 30.73 23.80 35.21 [6, 11] 58.33 51.96 61.23 44.60 56.24 28.27 43.21 43.89 27.29 32.93 47.19 [12, 17] 68.52 68.87 73.58 46.04 67.06 32.54 57.04 52.71 35.09 35.34 63.57 [14, 25] 71.30 66.42 77.78 49.64 70.35 34.68 60.99 58.37 38.30 36.0671.88 [18, 23] 71.30 72.06 81.73 53.24 73.41 38.24 63.46 56.79 36.24 35.10 70.17 [18, 29] 72.9265.93 79.51 47.48 70.35 34.44 58.77 57.69 36.24 34.62 70.42 [22, 31] 67.82 68.14 77.53 47.48 67.76 31.35 59.51 57.24 36."},{"citing_arxiv_id":"2605.02707","ref_index":39,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"SAIL: Structure-Aware Interpretable Learning for Anatomy-Aligned Post-hoc Explanations in OCT","primary_cat":"cs.CV","submitted_at":"2026-05-04T15:13:47+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":5.0,"formal_verification":"none","one_line_summary":"SAIL integrates anatomical priors at the representation level with semantic features via fusion to produce more anatomically aligned attribution maps in OCT without altering existing explainability techniques.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2604.18487","ref_index":17,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"Adversarial Humanities Benchmark: Results on Stylistic Robustness in Frontier Model Safety","primary_cat":"cs.CL","submitted_at":"2026-04-20T16:37:27+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"Stylistic rewrites of harmful prompts raise attack success rates from 3.84% to 36.8-65% across 31 frontier models, indicating weak generalization in safety refusals.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2604.13258","ref_index":17,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"Hessian-Enhanced Token Attribution (HETA): Interpreting Autoregressive LLMs","primary_cat":"cs.CL","submitted_at":"2026-04-14T19:43:43+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":5.0,"formal_verification":"none","one_line_summary":"HETA is a new attribution framework for decoder-only LLMs that combines semantic transition vectors, Hessian-based sensitivity scores, and KL divergence to produce more faithful and human-aligned token attributions than prior methods.","context_count":1,"top_context_role":"background","top_context_polarity":"background","context_text":"Finally, fMT [i]≤1 absorbs the gate in the second summand of equation 15. Aggregate error bounds.Let ∥ · ∥ 1 denote the sum over tokens. Summing equation 16 over i= 1, . . . , Tand using P i S(T) i ≤ ∥H T ∥1 andP i Ii ≤C I (finite by bounded logits) gives TX i=1 Attri −gAttri ≤ϵ M(W) β∥H T ∥1+γ C I \u0001 +β T c d τk + γ µ \u0010 T εorig + TX i=1 ε(i) mask \u0011 .(17) When the window covers most causal paths (ϵM(W)→0 as W↑ ) and the Hessian spectrum is rapidly decaying (τk →0 as k↑ ), the approximation error vanishes. If, additionally, truncation minimally perturbs next-token distributions (small εorig and ε(i) mask), the KL component is stable by equation 14. Discussion of constants. cd = √ d is the block-size factor connecting entrywise ℓ1 to Frobenius"},{"citing_arxiv_id":"2604.08885","ref_index":7,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"Uncertainty-Aware Transformers: Conformal Prediction for Language Models","primary_cat":"cs.LG","submitted_at":"2026-04-10T02:48:50+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":5.0,"formal_verification":"none","one_line_summary":"CONFIDE applies conformal prediction to transformer embeddings for valid prediction sets, improving accuracy up to 4.09% and efficiency over baselines on models like BERT-tiny.","context_count":1,"top_context_role":"background","top_context_polarity":"background","context_text":"This ensures each class attains the nominal coverage, not just the average over classes. •Credibility and confidence: Credibility= max yj∈Y p(yj),Confidence= 1−max yj̸=y∗ p(yj),(6) wherey∗= arg maxyjp(yj). Credibility reflects how well the top label conforms; confidence measures its separation from the rest. •Efficiency and correct efficiency: The average set size E [ |Γε(x)| ] (7) measuresefficiency(lower is better). Because one can appear \"efficient\" by shrinking sets while omitting the truth, we also report a normalized,correct compactness score, called correct efficiency, restricted to correctly covering sets: cEff = 1−E [ |Γε(x)| |Y| ⏐⏐⏐Y∈Γε(x) ] ∈[0,1](higher is better). In all results we report, coverage, efficiency, and correct efficiency together (Huang et al."},{"citing_arxiv_id":"2604.03463","ref_index":20,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"Super Agents and Confounders: Influence of surrounding agents on vehicle trajectory prediction","primary_cat":"cs.LG","submitted_at":"2026-04-03T21:15:06+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"Surrounding agents frequently degrade trajectory prediction accuracy in interactive driving scenes, and integrating a Conditional Information Bottleneck improves results by ignoring non-beneficial contextual signals.","context_count":1,"top_context_role":"background","top_context_polarity":"support","context_text":"often cancel each other out. The limitations of standard interpretability tools, such as Transformer attention weights or gradient-based saliency maps, further justify the need for more robust attribution methods. While these mechanisms are often used to identify influential features, research indicates that they are frequently insufficient for explanation. Jain et al. [20] demonstrate that attention weights often do not correlate with feature impor- tance metrics. Furthermore, Adebayo et al. [21] highlight that many saliency methods fail basic sanity checks, acting as simple edge detectors that remain independent of both model parameters and the data-generating process. These observations motivate our Shapley-based analysis"},{"citing_arxiv_id":"2602.24176","ref_index":182,"ref_count":4,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Beyond Explainable AI (XAI): An Overdue Paradigm Shift and Post-XAI Research Directions","primary_cat":"cs.CY","submitted_at":"2026-02-27T16:58:27+00:00","verdict":null,"verdict_confidence":null,"novelty_score":null,"formal_verification":null,"one_line_summary":null,"context_count":2,"top_context_role":"background","top_context_polarity":"support","context_text":"reading books, and not through data-driven computation. Human knowledge validation occurs through structured scientific community processes-peer review, expert verification, and system- atic validation protocols. Dewey conceptualizes communication as a fundamental mechanism for knowledge formation, serving to transform individual experiences into collectively shared, public knowledge[182]. Moreover, this third assumption faces a critical challenge in an AI-saturated infor- mation ecosystem. As AI-generated content proliferates across the internet, the epistemic value of computationally-derived knowledge-including XAI explanations themselves-becomes increasingly questionable. Research on model collapse demonstrates that genuine human interactions with systems"},{"citing_arxiv_id":"2602.16608","ref_index":12,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Explainable AI: Context-Aware Layer-Wise Integrated Gradients for Explaining Transformer Models","primary_cat":"cs.CL","submitted_at":"2026-02-18T17:03:10+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"CA-LIG is a unified hierarchical attribution method that computes layer-wise Integrated Gradients fused with class-specific attention gradients to generate signed, context-sensitive explanations for transformer models.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2604.06172","ref_index":1,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"EviSnap: Faithful Evidence-Cited Explanations for Cold-Start Cross-Domain Recommendation","primary_cat":"cs.IR","submitted_at":"2026-01-09T18:21:43+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"EviSnap creates cross-domain recommendations whose scores decompose exactly into evidence-cited concept contributions via offline LLM facet extraction, clustering, and linear transfer.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2508.04427","ref_index":100,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Decoding the Multimodal Maze: A Systematic Review on the Adoption of Explainability in Multimodal Attention-based Models","primary_cat":"cs.LG","submitted_at":"2025-08-06T13:14:20+00:00","verdict":null,"verdict_confidence":null,"novelty_score":null,"formal_verification":null,"one_line_summary":null,"context_count":1,"top_context_role":"background","top_context_polarity":"background","context_text":"Model-specific or the decompositional approach generates explanations from the in- ternal structure, parameters, and feature representation of the prediction model. As this review focuses on attention-based methods, most studies applying model-specific XAI methods leverage the attention weights. The use of attention weights as explanations, al- though debated [100], allows the generation of intuitive methods of observing the internal mechanics of a model [22]. Apart from attention weights, there are several other methods found in the literature for generating model-specific explanations. These methods are described following the classes used by Fantozziet al.[28]. 23 Attention-based methods.The attention scores are a matrix of cross-token probabilities."},{"citing_arxiv_id":"2503.16771","ref_index":26,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Enabling Global, Human-Centered Explanations for LLMs:From Tokens to Interpretable Code and Test Generation","primary_cat":"cs.SE","submitted_at":"2025-03-21T01:00:45+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"CodeQ aggregates token rationales into code categories to enable global interpretability of LLMs, claiming over 50% entropy reduction and revealing model preference for syntactic cues plus human misalignment in a 37-person study.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"1907.00570","ref_index":9,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Do Transformer Attention Heads Provide Transparency in Abstractive Summarization?","primary_cat":"cs.CL","submitted_at":"2019-07-01T06:46:43+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":5.0,"formal_verification":"none","one_line_summary":"Analysis of transformer attention heads in abstractive summarization shows specialization in some heads and proposes a method to measure model reliance on learned attention distributions.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"1906.10924","ref_index":9,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Interpretable Question Answering on Knowledge Bases and Text","primary_cat":"cs.CL","submitted_at":"2019-06-26T09:10:33+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":5.0,"formal_verification":"none","one_line_summary":"Compares LIME, input perturbation and attention for explaining QA on KB+text; proposes automatic evaluation paradigm and finds input perturbation superior in both automatic and human studies.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null}],"limit":50,"offset":0}