{"total":33,"items":[{"citing_arxiv_id":"2605.23780","ref_index":32,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Beyond Binary Edits Robust Multimodal Knowledge Editing with Adversarial Subspace Alignment","primary_cat":"cs.AI","submitted_at":"2026-05-22T15:46:10+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":7.0,"formal_verification":"none","one_line_summary":"Introduces Latent Adversarial Robustification and Rank-Constrained Subspace Learning to enable robust generalization in multimodal knowledge editing through adversarial subspace alignment.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.18879","ref_index":14,"ref_count":2,"confidence":0.98,"is_internal_anchor":true,"paper_title":"ZeroUnlearn: Few-Shot Knowledge Unlearning in Large Language Models","primary_cat":"cs.LG","submitted_at":"2026-05-16T03:10:36+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":"2605.16686","ref_index":10,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Scalable Knowledge Editing for Mixture-of-Experts LLMs via Tensor-Structured Updates","primary_cat":"cs.LG","submitted_at":"2026-05-15T22:46:58+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"A MEMIT-style knowledge editing framework for MoE LLMs that formulates per-expert updates via tensor structure and applies Woodbury identity for low-rank inversions, achieving up to 6x speedup with comparable editing quality.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.16600","ref_index":9,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Where Pretraining writes and Alignment reads: the asymmetry of Transformer weight space","primary_cat":"cs.LG","submitted_at":"2026-05-15T20:00:59+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":7.0,"formal_verification":"none","one_line_summary":"Pretraining and alignment induce asymmetric geometric traces in transformer weights because alignment updates concentrate in read pathways due to activation covariance while write pathways inherit less structure from alignment losses.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.12809","ref_index":80,"ref_count":1,"confidence":0.9,"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":"2605.12357","ref_index":11,"ref_count":1,"confidence":0.9,"is_internal_anchor":true,"paper_title":"$\\delta$-mem: Efficient Online Memory for Large Language Models","primary_cat":"cs.AI","submitted_at":"2026-05-12T16:31:44+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"δ-mem augments frozen LLMs with an 8x8 online memory state updated by delta-rule learning to generate low-rank attention corrections, delivering 1.10x average gains over the backbone and larger improvements on memory-heavy tasks.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.12207","ref_index":63,"ref_count":1,"confidence":0.9,"is_internal_anchor":true,"paper_title":"Not How Many, But Which: Parameter Placement in Low-Rank Adaptation","primary_cat":"cs.LG","submitted_at":"2026-05-12T14:46:00+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"Gradient-informed placement of LoRA parameters recovers full performance under GRPO while random placement does not, due to differences in gradient rank and stability across training regimes.","context_count":1,"top_context_role":"background","top_context_polarity":"background","context_text":"[61] Neel Nanda, Lawrence Chan, Tom Lieberum, Jess Smith, and Jacob Steinhardt. Progress measures for grokking via mechanistic interpretability. arXiv preprint arXiv:2301.05217, 2023. [62] Mathematical Association of America (MAA). Amc 2023: American mathematics competitions 2023 dataset. https://huggingface.co/datasets/math-ai/amc23, 2023. Accessed: 2024-05-05. 13 [63] Rui Pan, Xiang Liu, Shizhe Diao, Renjie Pi, Jipeng Zhang, Chi Han, and Tong Zhang. Lisa: Layerwise importance sampling for memory-efficient large language model fine-tuning. arXiv preprint arXiv:2403.17919, 2024. [64] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward"},{"citing_arxiv_id":"2605.10198","ref_index":34,"ref_count":1,"confidence":0.9,"is_internal_anchor":true,"paper_title":"Empty SPACE: Cross-Attention Sparsity for Concept Erasure in Diffusion Models","primary_cat":"cs.LG","submitted_at":"2026-05-11T08:46:29+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":5.0,"formal_verification":"none","one_line_summary":"SPACE induces sparsity in cross-attention parameters via closed-form iterative updates to erase target concepts more effectively than dense baselines in large diffusion models.","context_count":1,"top_context_role":"method","top_context_polarity":"use_method","context_text":"based methods [10, 11, 12, 13, 14, 30, 31, 32] typically align the model distribution conditioned on a target concept with its generation conditioned on an anchor concept. The main drawback of these methods is their high computational cost, as they rely on backpropagation to fine-tune the model. As a more computationally- and memory-efficient alternative, methods that update the model weights in closed-form have been presented. UCE [15], which builds upon [33] and [34], introduces a unified framework for concept editing by performing a closed-form update of the cross-attention parameters of T2I models. RECE [16] is an iterative method that extends UCE by proposing a robust unlearning framework that generates prompts capable of re-activating erased concepts and modifies the T2I model based on such prompts. SPEED"},{"citing_arxiv_id":"2605.09314","ref_index":29,"ref_count":1,"confidence":0.9,"is_internal_anchor":true,"paper_title":"How LLMs Are Persuaded: A Few Attention Heads, Rerouted","primary_cat":"cs.AI","submitted_at":"2026-05-10T04:15:24+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":7.0,"formal_verification":"none","one_line_summary":"Persuasion in LLMs works by redirecting a small set of attention heads to copy the target option token instead of reasoning over evidence, via a rank-one routing feature that can be directly edited or removed.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.09195","ref_index":27,"ref_count":1,"confidence":0.9,"is_internal_anchor":true,"paper_title":"The Geometry of Forgetting: Temporal Knowledge Drift as an Independent Axis in LLM Representations","primary_cat":"cs.AI","submitted_at":"2026-05-09T22:27:31+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"Temporal knowledge drift is encoded as a geometrically orthogonal direction in LLM residual streams, independent of correctness and uncertainty.","context_count":1,"top_context_role":"background","top_context_polarity":"background","context_text":"Temporal cutoffs and hallucination detection.Benchmarks for time-stamped question answer- ing TimeQA [ 8], TempLAMA [ 12], StreamingQA [ 22], and FreshLLMs [ 37] measure whether models produce stale answers but treat drift as an output-level failure; we instead ask whether drift is encoded internally regardless of whether it surfaces. Knowledge editing (ROME [ 26], MEMIT [27], SERAC [29]) overwrites stale facts post-hoc and is complementary to our diagnostic. The hallucination-detection literature has converged on output- or token-level signals semantic en- tropy [21], SEP [20], CCS [6], SAPLMA [3] but these conflate uncertainty, falsehood, and refusal. §6 shows that none isolatestemporaldrift: a model can be confidently, fluently wrong because its"},{"citing_arxiv_id":"2605.04972","ref_index":11,"ref_count":1,"confidence":0.9,"is_internal_anchor":true,"paper_title":"Why Expert Alignment Is Hard: Evidence from Subjective Evaluation","primary_cat":"cs.CL","submitted_at":"2026-05-06T14:28:09+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":5.0,"formal_verification":"none","one_line_summary":"Expert alignment in subjective LLM evaluations is difficult because expert judgments are heterogeneous, partly tacit, dimension-dependent, and temporally unstable.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.02234","ref_index":16,"ref_count":1,"confidence":0.9,"is_internal_anchor":true,"paper_title":"Bucketing the Good Apples: A Method for Diagnosing and Improving Causal Abstraction","primary_cat":"cs.AI","submitted_at":"2026-05-04T05:09:21+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":7.0,"formal_verification":"none","one_line_summary":"A four-step recipe partitions the input space using interchange intervention behavior to diagnose where causal abstractions hold and to guide improvements, demonstrated by recovering a full hypothesis from scratch in a toy logic task.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.02083","ref_index":9,"ref_count":1,"confidence":0.9,"is_internal_anchor":true,"paper_title":"EditPropBench: Measuring Factual Edit Propagation in Scientific Manuscripts","primary_cat":"cs.CL","submitted_at":"2026-05-03T22:46:47+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":7.0,"formal_verification":"none","one_line_summary":"EditPropBench evaluates LLM editors on propagating factual edits to dependent claims in synthetic scientific manuscripts, showing that even the strongest systems miss roughly 30% of required updates on hard cases.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.08143","ref_index":20,"ref_count":2,"confidence":0.98,"is_internal_anchor":true,"paper_title":"HoReN: Normalized Hopfield Retrieval for Large-Scale Sequential Model Editing","primary_cat":"cs.LG","submitted_at":"2026-05-02T15:51:31+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":5.0,"formal_verification":"none","one_line_summary":"HoReN is a parameter-preserving editor that wraps an MLP with a Hopfield codebook memory and scales to 50K sequential edits on ZsRE while maintaining performance above 0.93.","context_count":1,"top_context_role":"background","top_context_polarity":"background","context_text":"Advances in neural information processing systems, 35:17359-17372, 2022. [19] Kevin Meng, Arnab Sen Sharma, Alex Andonian, Yonatan Belinkov, and David Bau. Mass- editing memory in a transformer.arXiv preprint arXiv:2210.07229, 2022. [20] Eric Mitchell, Charles Lin, Antoine Bosselut, Chelsea Finn, and Christopher D Manning. Fast model editing at scale.arXiv preprint arXiv:2110.11309, 2021. [21] Eric Mitchell, Charles Lin, Antoine Bosselut, Christopher D Manning, and Chelsea Finn. Memory-based model editing at scale. InInternational Conference on Machine Learning, pages 15817-15831. PMLR, 2022. [22] Raymond Ng, Thanh Ngan Nguyen, Yuli Huang, Ngee Chia, Weiqi Leong, et al. Sea-lion: Southeast asian languages in one network. InIJCNLP-AACL, 2025."},{"citing_arxiv_id":"2604.27401","ref_index":18,"ref_count":1,"confidence":0.9,"is_internal_anchor":true,"paper_title":"Perturbation Probing: A Two-Pass-per-Prompt Diagnostic for FFN Behavioral Circuits in Aligned LLMs","primary_cat":"cs.CL","submitted_at":"2026-04-30T04:13:33+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"Perturbation probing identifies tiny sets of FFN neurons that control refusal templates and language routing in LLMs, enabling precise ablations and directional interventions that alter behavior on benchmarks while preserving safety.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2604.26686","ref_index":48,"ref_count":1,"confidence":0.9,"is_internal_anchor":true,"paper_title":"When Model Editing Meets Service Evolution: A Knowledge-Update Perspective for Service Recommendation","primary_cat":"cs.SE","submitted_at":"2026-04-29T13:51:56+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"EVOREC integrates locate-then-edit model editing with FA-constrained decoding to improve LLM-based service recommendation under evolution, reporting 25.9% average relative gain in Recall@5 over baselines and 22.3% over fine-tuning in dynamic scenarios.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2604.23877","ref_index":3,"ref_count":1,"confidence":0.9,"is_internal_anchor":true,"paper_title":"Knowledge Vector of Logical Reasoning in Large Language Models","primary_cat":"cs.CL","submitted_at":"2026-04-26T20:37:26+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"Distinct linear knowledge vectors for deductive, inductive, and abductive reasoning in LLMs can be refined via complementary subspace constraints to improve performance through mutual knowledge sharing.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2604.23750","ref_index":26,"ref_count":2,"confidence":0.9,"is_internal_anchor":true,"paper_title":"The Override Gap: A Magnitude Account of Knowledge Conflict Failure in Hypernetwork-Based Instant LLM Adaptation","primary_cat":"cs.LG","submitted_at":"2026-04-26T14:59:14+00:00","verdict":"CONDITIONAL","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"Knowledge conflicts in hypernetwork LLM adaptation stem from constant adapter margins losing to frequency-dependent pretrained margins; selective layer boosting and conflict-aware triggering raise deep-conflict accuracy to 71-72.5% on Gemma-2B and Mistral-7B.","context_count":1,"top_context_role":"background","top_context_polarity":"background","context_text":"communication during adapter generation. Text-to-LoRA [4] extends the paradigm to task adapta- tion from natural language descriptions. In the vision domain, Vision-as-LoRA [34] and LoRA.rar [31] generate adapters from images and style combinations respectively. All these methods are evaluated primarily on recall tasks. Knowledge Editing.ROME [25] and MEMIT [26] directly modify MLP weights to change spe- cific facts stored in language models. MEND [27] and SERAC [28] instead learn auxiliary networks that predict weight edits or route around them, scaling to larger models. Transformer-Patcher [16] adds a single neuron per edit in the last feed-forward layer, preserving behavior on unrelated in- puts. These methods treat knowledge as isolated key-value pairs, an assumption that leads to ripple"},{"citing_arxiv_id":"2604.19089","ref_index":30,"ref_count":1,"confidence":0.9,"is_internal_anchor":true,"paper_title":"Towards Scalable Lifelong Knowledge Editing with Selective Knowledge Suppression","primary_cat":"cs.AI","submitted_at":"2026-04-21T05:02:29+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":5.0,"formal_verification":"none","one_line_summary":"LightEdit enables scalable lifelong knowledge editing in LLMs via selective knowledge retrieval and probability suppression during decoding, outperforming prior methods on ZSRE, Counterfact, and RIPE while reducing training costs.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2604.15166","ref_index":30,"ref_count":2,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Class Unlearning via Depth-Aware Removal of Forget-Specific Directions","primary_cat":"cs.CV","submitted_at":"2026-04-16T15:46:02+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"DAMP performs one-shot class unlearning by depth-aware projection removal of forget-specific directions, producing forgetting behavior closer to retraining from scratch than prior methods on image classification tasks.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2604.14717","ref_index":28,"ref_count":2,"confidence":0.9,"is_internal_anchor":true,"paper_title":"Layered Mutability: Continuity and Governance in Persistent Self-Modifying Agents","primary_cat":"cs.AI","submitted_at":"2026-04-16T07:27:35+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":5.0,"formal_verification":"none","one_line_summary":"Persistent self-modifying AI agents exhibit compositional drift from mismatches across five mutability layers, with governance difficulty rising under rapid mutation, strong coupling, weak reversibility, and low observability, as indicated by a 0.68 identity hysteresis ratio in a preliminary ratchet","context_count":1,"top_context_role":"background","top_context_polarity":"background","context_text":"Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. Overcoming catastrophic for- getting in neural networks.Proceedings of the National Academy of Sciences, 114(13):3521-3526, 2017. 16 [27] Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. Locating and editing factual associations in GPT.Advances in Neural Information Processing Systems, 35, 2022. [28] Kevin Meng, Arnab Sen Sharma, Alex Andonian, Yonatan Belinkov, and David Bau. Mass- editing memory in a transformer.arXiv preprint arXiv:2210.07229, 2023. [29] S. Schneider et al. Time, identity and consciousness in language model agents.arXiv preprint arXiv:2603.09043, 2026. [30] Yijia Yan, Wenshuo Yao, Jiacheng Huang, Rui Wang, Yuxuan Wang, and Tat-Seng Chua."},{"citing_arxiv_id":"2604.08284","ref_index":10,"ref_count":1,"confidence":0.9,"is_internal_anchor":true,"paper_title":"Distributed Multi-Layer Editing for Rule-Level Knowledge in Large Language Models","primary_cat":"cs.CL","submitted_at":"2026-04-09T14:22:19+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"Rule knowledge in LLMs is localized by form across layers; a distributed multi-layer editing method improves instance portability by 13.91 and rule understanding by 50.19 percentage points over baselines on multiple models.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2603.22241","ref_index":38,"ref_count":1,"confidence":0.9,"is_internal_anchor":true,"paper_title":"MemDLM: Memory-Enhanced DLM Training","primary_cat":"cs.CL","submitted_at":"2026-03-23T17:39:56+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":7.0,"formal_verification":"none","one_line_summary":"MemDLM embeds a simulated denoising trajectory into DLM training via bi-level optimization, creating a parametric memory that improves convergence and long-context performance even when the memory is dropped at test time.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2603.14259","ref_index":19,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"GenRecEdit: Adapting Model Editing for Generative Recommendation with Cold-Start Items","primary_cat":"cs.IR","submitted_at":"2026-03-15T07:31:28+00:00","verdict":"CONDITIONAL","verdict_confidence":"MODERATE","novelty_score":7.0,"formal_verification":"none","one_line_summary":"GenRecEdit injects cold-start items into generative recommendation models via context-aware token editing and interference-reducing triggers, boosting cold-start accuracy while using only 9.5% of retraining time.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2601.21577","ref_index":13,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Collaborative Parameter Learning: Mitigating Forgetting via Parameter-Level Gradient Analysis","primary_cat":"cs.LG","submitted_at":"2026-01-29T11:42:30+00:00","verdict":"CONDITIONAL","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"Collaborative Parameter Learning freezes 50-75% of parameters whose updates cause forgetting and updates only the 25-50% that mitigate it, allowing LLMs to learn 20-48% more new questions with negligible forgetting and lower compute cost.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2601.19208","ref_index":13,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"How Do Transformers Learn to Associate Tokens: Gradient Leading Terms Bring Mechanistic Interpretability","primary_cat":"cs.CL","submitted_at":"2026-01-27T05:22:34+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":7.0,"formal_verification":"none","one_line_summary":"Transformer weights at early training stages are closed-form compositions of bigram, token-interchangeability, and context mappings that directly reflect text-corpus statistics and explain the emergence of semantic associations.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2601.14053","ref_index":90,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"LLMOrbit: A Circular Taxonomy of Large Language Models -From Scaling Walls to Agentic AI Systems","primary_cat":"cs.LG","submitted_at":"2026-01-20T15:06:19+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":3.0,"formal_verification":"none","one_line_summary":"A survey taxonomy of LLMs identifies three scaling crises and six efficiency paradigms while tracing the shift from generation to tool-using agents.","context_count":1,"top_context_role":"background","top_context_polarity":"support","context_text":"onlyhalfthe attention head dimensions, leaving the rest unchanged: Partial-RoPE(q) = [RoPE(q1:d/2);q d/2+1:d](58) Why:At inference on sequences longer than train- ing's maximum length, full RoPE can produce \"unseen\" rotation angles, potentially degrading quality. Partial RoPE provides an \"unrotated escape hatch\" that carries position-agnostic information, improving length extrapo- lation [90]. No Positional Embeddings (NoPE):SmolLM3 (3B) and Kimi Linear take the radical approach of omitting po- sitional encodings entirely.How does the model know token order?The causal attention mask provides im- plicit positional information: tokentcan only attend to tokens≤t, creating directional flow. Theoretically, this \"masked self-attention as implicit position encoding\" is"},{"citing_arxiv_id":"2508.00570","ref_index":26,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"SPRINT: Scalable and Predictive Intent Refinement for LLM-Enhanced Session-based Recommendation","primary_cat":"cs.IR","submitted_at":"2025-08-01T12:11:10+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":5.0,"formal_verification":"none","one_line_summary":"SPRINT refines LLM-generated intents for session-based recommendation via a global intent pool, performance validation, selective LLM invocation during training, and a lightweight intent predictor for scalable inference without LLM calls.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2507.03724","ref_index":69,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"MemOS: A Memory OS for AI System","primary_cat":"cs.CL","submitted_at":"2025-07-04T17:21:46+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":5.0,"formal_verification":"none","one_line_summary":"MemOS introduces a unified memory management framework for LLMs using MemCubes to handle and evolve different memory types for improved controllability and evolvability.","context_count":1,"top_context_role":"background","top_context_polarity":"background","context_text":"LightRAG [11], NodeRAG [ 47, 48], HyperGraphRAG [ 49], HippoRAG [ 50, 51], PGRAG[ 52], Zep [ 53], A-MEM [ 54], Mem0[55] Implicit Parametric Knowledge BERT [56], RLHF [57], CTRL [58], SLayer [59] Modular Parameter Adapta- tion LoRA [60], PRAG [ 61], DyPRAG [ 62], SERAC [ 63], CaliNet [64], DPM [65], GRACE [66] Parametric Memory Editing ROME [ 67], MEMIT [ 68], AlphaEdit [ 69], AnyEdit [ 70], EasyEdit [71], AdaPLE [72], MEMAT [73] fine-tunes the model to decouple output into separate memory and reasoning components, fully leveraging memory for inference. SLayer [59] identifies memory-relevant layers in the model and locally fine-tunes them to enhance specific knowledge representation. It is worth noting that relying solely on memorization of"},{"citing_arxiv_id":"2506.04042","ref_index":8,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Causal Path Alignment: Anchoring the Optimization Trajectory for Controllable In-Parameter Knowledge Editing","primary_cat":"cs.CL","submitted_at":"2025-06-04T15:06:46+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":5.0,"formal_verification":"none","one_line_summary":"Causal Path Alignment anchors optimization trajectories in in-parameter knowledge editing to follow relation-aware causal paths, reducing subject-dominant memory interference in LLMs.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2310.12508","ref_index":143,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"SalUn: Empowering Machine Unlearning via Gradient-based Weight Saliency in Both Image Classification and Generation","primary_cat":"cs.LG","submitted_at":"2023-10-19T06:17:17+00:00","verdict":"CONDITIONAL","verdict_confidence":"MODERATE","novelty_score":6.0,"formal_verification":"none","one_line_summary":"SalUn uses gradient-based weight saliency to achieve effective machine unlearning of data, classes, or concepts in image classification and generation, narrowing the gap to exact retraining.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2303.08112","ref_index":63,"ref_count":1,"confidence":0.9,"is_internal_anchor":true,"paper_title":"Eliciting Latent Predictions from Transformers with the Tuned Lens","primary_cat":"cs.LG","submitted_at":"2023-03-14T17:47:09+00:00","verdict":"ACCEPT","verdict_confidence":"MODERATE","novelty_score":7.0,"formal_verification":"none","one_line_summary":"Training per-layer affine probes on frozen transformers yields more reliable latent predictions than the logit lens and enables detection of malicious inputs from prediction trajectories.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2202.05262","ref_index":26,"ref_count":1,"confidence":0.98,"is_internal_anchor":true,"paper_title":"Locating and Editing Factual Associations in GPT","primary_cat":"cs.CL","submitted_at":"2022-02-10T18:59:54+00:00","verdict":"ACCEPT","verdict_confidence":"LOW","novelty_score":8.0,"formal_verification":"none","one_line_summary":"Factual associations in autoregressive transformers are localized to mid-layer feed-forward modules and can be edited via rank-one model editing while preserving both specificity and generalization on counterfactual tests.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null}],"limit":50,"offset":0}