{"total":15,"items":[{"citing_arxiv_id":"2607.07937","ref_index":34,"ref_count":1,"confidence":0.88,"is_internal_anchor":false,"paper_title":"When Debiasing Backfires: Counterintuitive Side Effects of Preprocessing-Based Stereotype Mitigation","primary_cat":"cs.CL","submitted_at":"2026-07-08T21:34:49+00:00","verdict":"CONDITIONAL","verdict_confidence":"HIGH","novelty_score":6.0,"formal_verification":"none","one_line_summary":"Debiasing language-model training data for a target group frequently increases stereotyping or counter-stereotyping for non-target groups across categories, models, and scales.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.29139","ref_index":9,"ref_count":1,"confidence":0.88,"is_internal_anchor":false,"paper_title":"How Token Influence Decays with Distance: A Green-Function View of Trained Language Models","primary_cat":"cs.LG","submitted_at":"2026-06-28T01:00:22+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":5.0,"formal_verification":"none","one_line_summary":"Empirical Jacobian analysis reveals that token influence in trained language models decays as a power law with distance (exponent ~0.8), a learned property not present in random models.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.20561","ref_index":124,"ref_count":1,"confidence":0.88,"is_internal_anchor":false,"paper_title":"TimeProVe: Propose, then Verify for Efficient Long Video Temporal Reasoning in Activities of Daily Living","primary_cat":"cs.CV","submitted_at":"2026-06-18T17:59:48+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"TimeProVe proposes a propose-then-verify framework using lightweight action-based candidate evidence generation followed by targeted VLM verification for efficient long video temporal reasoning, achieving 7.3% improvement on OTB with 75% fewer VLM calls.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.13115","ref_index":45,"ref_count":1,"confidence":0.88,"is_internal_anchor":false,"paper_title":"G-Long: Graph-Enhanced Memory Management for Efficient Long-Term Dialogue Agents","primary_cat":"cs.CL","submitted_at":"2026-06-11T09:42:13+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":5.0,"formal_verification":"none","one_line_summary":"G-Long uses graph-enhanced triplet memory and attention-aware scoring from a T5 summarizer to achieve up to 9.8% better response quality on MSC and 40.8% better retrieval recall on LME with lower overhead.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.04514","ref_index":3,"ref_count":1,"confidence":0.88,"is_internal_anchor":false,"paper_title":"SAILRec: Steering LLM Attention to Dual-Side Semantically Aligned Collaborative Embeddings for Recommendation","primary_cat":"cs.IR","submitted_at":"2026-06-03T06:46:32+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":4.0,"formal_verification":"none","one_line_summary":"SAILRec uses dual-side semantic alignment and hierarchical attention steering to improve how LLMs incorporate collaborative embeddings for recommendations, outperforming baselines on MovieLens-1M and Amazon-Book datasets.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.27561","ref_index":10,"ref_count":1,"confidence":0.88,"is_internal_anchor":false,"paper_title":"Clinical Validation of the Melanoscope AI Mobile Dermoscopy Clinical Decision Support System","primary_cat":"cs.CV","submitted_at":"2026-05-26T18:29:53+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":3.0,"formal_verification":"none","one_line_summary":"Prospective single-center validation of a cascade deep learning dermoscopy CDSS found no false negatives for five malignant lesions and 88.3% specificity, with quantitative IoU assessment of attention maps.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.23393","ref_index":16,"ref_count":1,"confidence":0.88,"is_internal_anchor":false,"paper_title":"Every Component is a Lookup: Token Attribution and Composition from a Single Decomposition","primary_cat":"cs.LG","submitted_at":"2026-05-22T09:03:01+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"Unpack decomposes transformer credit via a unified backward recursion on the φ(S)U template, recovering known IOI circuits with mode labels and showing consistent duplicate-name suppression across Pythia scales from a single forward pass.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.20158","ref_index":1,"ref_count":1,"confidence":0.88,"is_internal_anchor":false,"paper_title":"Rethinking Visual Attribution for Chest X-ray Reasoning in Large Vision Language Models","primary_cat":"cs.CV","submitted_at":"2026-05-19T17:46:40+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"Existing visual attribution methods often fail to identify the visual evidence used by LVLMs in chest X-ray reasoning, while MedFocus using unbalanced optimal transport and targeted interventions substantially outperforms them across multiple models and settings.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.12952","ref_index":1,"ref_count":1,"confidence":0.88,"is_internal_anchor":false,"paper_title":"Debunking Grad-ECLIP: A Comprehensive Study on Its Incorrectness and Fundamental Principles for Model Interpretation","primary_cat":"cs.CV","submitted_at":"2026-05-13T03:35:23+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":4.0,"formal_verification":"none","one_line_summary":"Grad-ECLIP is an equivalent but flawed variant of attention-based interpretation, with two principles proposed to ensure model explanations reflect the original model.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.12168","ref_index":43,"ref_count":1,"confidence":0.88,"is_internal_anchor":false,"paper_title":"On What We Can Learn from Low-Resolution Data","primary_cat":"cs.LG","submitted_at":"2026-05-12T14:16:05+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"Low-resolution data improves high-resolution model performance when high-resolution samples are limited, via KL-divergence bounds and experiments on vision transformers and CNNs.","context_count":1,"top_context_role":"background","top_context_polarity":"background","context_text":"\u0012 P(θ| X l) P(θ| X h) \u0013 = log   1 Zl exp \u0010 −PN i=1 ℓ(θ,x i)−ℓ(θ,x l) \u0011 1 Zh exp \u0010 −PN i=1 ℓ(θ,x i)−ℓ(θ,x h) \u0011   (40) = log \u0012 Zh Zl \u0013 + (ℓ(θ,x h)−ℓ(θ,x l)).(41) Introducing∆(θ) =ℓ(θ,x h)−ℓ(θ,x l)for notational convenience. A rewrite ofZ h yields, Zh = Z exp − NX i=1 ℓ(θ,x i)−ℓ(θ,x l) ! exp (ℓ(θ,x l)−ℓ(θ,x h))dθ(42) = Z Zlp(θ| X l) exp (−∆(θ))dθ(43) =Z lEθ|Xl[exp (−∆(θ))].(44) Substitute back into eq. (41): log \u0012 Zh Zl \u0013 + ∆(θ) = log Eθ|Xl[exp(−∆(θ))] \u0001 + ∆(θ),(45) and subsequently into eq. (39), the difference in KL divergence admits the following closed-form expression: KLh −KL l = Z p(θ| X) log Eθ|Xl[exp(−∆(θ))] \u0001 + ∆(θ) \u0001 dθ(46) = log Eθ|Xl[exp(−∆(θ))] \u0001 +E θ|X [∆(θ)].(47) The use of the cumulant generating function used in Proposition 4 can similarly be applied here:"},{"citing_arxiv_id":"2605.07772","ref_index":65,"ref_count":1,"confidence":0.88,"is_internal_anchor":false,"paper_title":"Training-Induced Escape from Token Clustering in a Mean-Field Formulation of Transformers","primary_cat":"cs.LG","submitted_at":"2026-05-08T14:12:13+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":5.0,"formal_verification":"none","one_line_summary":"Training a mean-field Transformer under L2 regularization induces an escape from attention-driven token clustering in later layers after initial clustering.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.06212","ref_index":1,"ref_count":1,"confidence":0.88,"is_internal_anchor":false,"paper_title":"Playing the network backward: A Game Theoretic Attribution Framework","primary_cat":"cs.LG","submitted_at":"2026-05-07T13:15:07+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"Backward attribution is reframed as integrals over trajectories in a two-player game on the network, unifying gradients and alpha-beta-LRP while enabling new adaptations that outperform prior methods on ViT-B/16 localization metrics.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2604.09368","ref_index":1,"ref_count":1,"confidence":0.88,"is_internal_anchor":false,"paper_title":"Through Their Eyes: Fixation-aligned Tuning for Personalized User Emulation","primary_cat":"cs.MM","submitted_at":"2026-04-10T14:38:45+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"Personalized soft prompts steer VLM attention to match user-specific gaze patterns, yielding better attention alignment and click prediction in recommendation simulations.","context_count":1,"top_context_role":"method","top_context_polarity":"use_method","context_text":"coefficients to construct personalized soft prompts that steer the VLM's attention toward each user's characteristic fixation pattern. To our knowledge, our work is the first to apply personal- ized gaze-guided alignment to VLMs in the recommendation setting.We evaluate the framework across two architecturally distinct VLM backbones [ 3, 28] and three interpretability-based probing operators [1, 2, 23] to mitigate model-specific confounds. Our experiments demonstrate that personalized fixation alignment effectively shifts the VLM's visual attention distribution toward each user's characteristic gaze pattern, and that this perceptual alignment contributes to improved consistency between the simu- lator's predicted decisions and the user's actual behavior, yielding"},{"citing_arxiv_id":"2604.13073","ref_index":9,"ref_count":1,"confidence":0.88,"is_internal_anchor":false,"paper_title":"OmniTrace: A Unified Framework for Generation-Time Attribution in Omni-Modal LLMs","primary_cat":"cs.CL","submitted_at":"2026-03-20T17:25:00+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":7.0,"formal_verification":"none","one_line_summary":"OmniTrace converts token-level signals into span-level cross-modal attributions for open-ended generation in omni-modal LLMs via generation-time tracing.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2603.13652","ref_index":1,"ref_count":1,"confidence":0.88,"is_internal_anchor":false,"paper_title":"Causal Attribution via Activation Patching","primary_cat":"cs.CV","submitted_at":"2026-03-13T23:25:49+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"CAAP produces patch attributions in ViTs by direct activation patching on intermediate layers to measure causal contribution to the target class score.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null}],"limit":50,"offset":0}