Retrieval pipelines, vector search, chunking and reranking: how models are grounded in a corpus instead of their weights.
arXiv:2609.18898v1 Announce Type: new Abstract: Training-free weighted aggregation is widely used to lift 2D semantic features onto 3D Gaussians for open-vocabulary scene understanding, yet its theor...
The paper introduces WFM, a Wiki Foundation Model designed to support complex agentic reasoning by combining dense document contexts with markdown-based topological linkages. It formalizes a Wiki Graph schema that preserves explicit topologies while embedding continuous semantics, and employs a query‑conditioned attentive aggregation for efficient message passing. The authors also propose an NCCL‑based protocol to reduce distributed system overhead, achieving a 10.5× training speedup and strong performance on long‑term memory and multi‑hop reasoning benchmarks.
The paper investigates when compressed vector representations can provide exact linear or affine readouts for a finite lexicon’s truth conditions, establishing a necessary and sufficient row‑space condition. It shows that the augmented truth matrix’s rank determines the minimal dimension needed for exact linear (rank r) and affine (rank r − 1) readouts, and that exact readouts preserve Boolean connectives. Experiments on GloVe and word2vec embeddings reveal that while many predicates are linearly separable, none achieves exact affine recovery from pretrained embeddings, yet supervised transductive training can attain exact affine recovery at dimensions meeting the theoretical bound, preserving most of the original variance. "whyItMatters":"The results provide a precise mathematical criterion for when vector embeddings can faithfully encode logical truth conditions, informing both theoretical understanding and practical training of language models."
EdgeReMIND is a linear memorization model designed for temporal multi-relational link prediction on the Temporal Graph Benchmark 2.0 (TGB 2.0). It achieves the highest reported test mean reciprocal rank (MRR) on six of eight TGB 2.0 datasets and is the only relation‑aware method that runs on all datasets, overcoming the scalability ceiling that limits existing embedding methods. The model uses learned per‑relation weights over data‑calibrated features, making it a practical state‑of‑the‑art baseline for large‑scale temporal graphs.
Subspace-Decomposed JEPAs (SD-JEPA) split the latent space of Joint-Embedding Predictive Architectures into two orthogonal subspaces: a low-dimensional progression subspace trained with a cosine-margin triplet loss and a high-dimensional content subspace regularised by SIGReg. The authors prove that the anti-collapse forces act on disjoint coordinates, allowing additive composition rather than competition. SD-JEPA outperforms the LeWM baseline on most control benchmarks and the strongest non-LeWM JEPA baseline on Push‑T, with a subspace-ablation confirming the split as essential. The 1‑D angular progression coordinate serves as a scene-aware compass, advancing with task progress, regressing on backtracking, and relocalising under perturbations to separate surprise from meaning.
arXiv:2609.17749v1 Announce Type: new Abstract: A lot of expectations have been put for years on integrating domain expert knowledge in image classification models. Several approaches have been explo...
The paper re‑examines the impact of removing the encoder from Action Chunking Transformers (ACT), a model used for robot manipulation learning. Contrary to the original claim that encoder removal drops success rates from 35% to 2%, the authors find no such dramatic decline in their re‑runs, though minor variations remain uncertain. They attribute the discrepancy to factors like training length and checkpoint selection, and note that the encoder’s latent variable offers little reconstruction benefit on the tested benchmark, while its removal speeds up training.
VoiceTrace introduces a new benchmark, VoiceTrace-Bench, for hybrid speech retrieval that combines a textual query specifying "what" to retrieve with a reference speech specifying "who" to retrieve. The authors propose a two‑stage framework: VoiceTrace‑Emb, which learns unified audio‑text embeddings for efficient large‑scale retrieval, and VoiceTrace‑Reranker, which fine‑grains relevance by jointly examining query‑candidate pairs. Experiments show VoiceTrace outperforms existing methods on both traditional semantic speech retrieval benchmarks and the new hybrid setting.
The paper investigates which historical examples are most useful for time‑series forecasting by defining predictive relevance as the expected future utility conditioned on inference‑time information. It introduces a two‑stage approach: a normalized‑pattern retriever generates a coarse candidate set, and a lightweight MLP reranks these candidates using future‑supervised relevance while keeping inference strictly past‑only. Experiments on six benchmarks show that this reranker improves pattern retrieval and outperforms a matched‑protocol baseline, revealing that historical relevance is structured, domain‑dependent, and not governed by a single universal retrieval rule.
The paper introduces Multiple Embedding Replay Selection (MERS), a graph‑based method that combines supervised and self‑supervised embeddings to improve sample selection for replay buffers in continual learning. MERS replaces traditional buffer selection modules and demonstrates consistent performance gains over state‑of‑the‑art strategies, especially in low‑memory settings. Experiments on CIFAR‑100 and TinyImageNet show that MERS outperforms single‑embedding baselines without adding model parameters or increasing replay volume, making it a practical, drop‑in enhancement for replay‑based continual learning.
The paper introduces Vocabulary‑Guided Gait Recognition (Gait‑World), a new paradigm that leverages human vocabularies and Vision‑Language Models (VLMs) to better understand and recognize gait. It proposes the α‑Gait model, which uses a Vocabulary Relation Mapper and a Gait Fine‑grained Detector to align vocabulary prompts with gait features, addressing the challenge of heterogeneous embedding spaces. Experiments on multiple gait datasets demonstrate the potential of incorporating VLM‑derived vocabulary information into gait recognition.
The study investigates whether contextual embeddings can detect meaning changes in scientific terminology beyond traditional frequency counts. Using Astrophysics and NLP corpora from 2010 to 2024, the authors extract candidate terms with KeyBERT, filter for significant frequency rises, and then evaluate semantic drift via multiple embedding‑based metrics. Results show that frequency methods slightly outperform embedding metrics in aligning with expert judgments, yet embedding‑only detections (e.g., "primordial black holes") reveal critical conceptual shifts missed by frequency alone, suggesting complementary value.
The paper investigates how to efficiently repair stale key-value (KV) caches in retrieval‑augmented generation systems after document edits. It proposes a budgeted in‑place recomputation approach and evaluates training‑free position‑selection policies on a factual RAG benchmark. Across three model families, a contiguous edit‑local window consistently recovers most of the post‑edit answer quality while being 13–21 times faster than a full re‑prefill, though its effectiveness diminishes when answer‑bearing text moves downstream.
The study investigates which components of a neural network contribute to rapid generalization (grokking) and how stable that improvement remains during further training. By transferring internal attention and MLP weights along with token embeddings and readout, the authors achieve a 5.46‑percentage‑point boost in early accuracy and a 558‑step reduction in confirmation latency, while also demonstrating that freezing transferred representations largely prevents post‑grokking relapse. The work delineates clear component‑level differences between acceleration and stability, and identifies architectural limits where omitting donor embeddings leads to significant performance loss.
Q-BIOLAT is a framework that converts pretrained protein-language-model embeddings into compact binary codes and trains a quadratic unconstrained binary optimization (QUBO) surrogate with unary and pairwise latent interactions for protein fitness optimization. The study demonstrates that binary encodings with similar predictive accuracy can produce different Hamming neighborhoods, affecting local optima and search trajectories, and shows that PCA followed by per‑coordinate median thresholding yields a more balanced binary space than AE/VAE baselines. Experimental evaluation on GFP and AAV fitness landscapes from ProteinGym confirms that simulated annealing, genetic algorithms, and greedy hill climbing can retrieve high‑percentile variants, with decoded candidates reported via surrogate‑predicted scores.
Attention-based rerankers use a null-query calibration step to reduce positional and structural bias, but this step can inadvertently remove useful signal when prompts contain detailed instructions, constraints, personas, or demonstrations. The study demonstrates that such calibration harms performance on instruction-heavy tasks and introduces an interpolated null calibration that selectively limits instruction influence on the null baseline. This modification restores or surpasses the effectiveness of standard rerankers on instruction-heavy tasks while maintaining benefits when the null pass remains relevance-agnostic, and shows that in-context demonstrations can improve reranking with minimal calibration interference.
The paper introduces the Tri‑Metric Router, a deterministic, training‑free policy that chooses among Raw, Neural, and Lexical pipelines for retrieval‑augmented generation on commodity GPUs. It uses three CPU‑side signals—spatial complexity, syntactic density, and type‑token ratio—to balance VRAM headroom and latency, calibrated on LongBench qasper. The method eliminates out‑of‑memory failures and improves alignment and F1 scores compared to always‑on lexical compression without extra VRAM or training costs.
NeuroECG is a deep learning framework that repurposes a pretrained ECG foundation model to predict neurological outcomes after cardiac arrest without using electroencephalography (EEG). The model fine‑tunes the backbone with a gradual unfreezing strategy on single‑channel bedside ECG, aggregates multiple ECG segments via quantile pooling and PCA, and achieves an AUROC of 0.7333 using ECG alone. When combined with static clinical covariates, NeuroECG improves performance to an AUROC of 0.8077 and an AUPRC of 0.8970, demonstrating that bedside ECG can serve as a low‑cost, auxiliary prognostic tool in an EEG‑free setting.
The paper introduces LGM, a neuro‑symbolic framework that disentangles long‑term memory by mapping historical interactions into a continuous latent graph. Instead of static memory graphs, LGM uses a sparse autoencoder to create query‑aware latent nodes and edges, then applies a graph encoder conditioned on the query to perform non‑linear message passing. Experiments on long‑term personalization benchmarks show that LGM outperforms existing methods in capturing both explicit and implicit user preferences and generating personalized responses.
The paper compares Knowledge-Graph Based Augmentation (Graph-RAG) with Retrieval-Augmented Generation (RAG) for answering culturally specific questions. Using the LatamQA dataset, Graph-RAG, built automatically from Wikipedia via KGGen, matches RAG performance and reduces the base LLM’s error by 72% with a standard KG and 78% with a benchmark-aware variant. The approach also transfers zero‑shot to Portuguese, showing multilingual applicability.