arXiv:2604.03260v1 Announce Type: cross Abstract: We introduce Focus, a method that learns which token pairs matter rather than approximating all of them. Learnable centroids assign tokens to groups; distant attention is restricted to same-group pairs while local attention operates at full resolution. Because all model weights stay frozen, Focus is purely additive: centroid-only training (as few as 148K parameters) improves domain perplexity with zero degradation on downstream benchmarks--from 124M to 70B parameters, across five attention architectures. No existing efficient attention method achieves this in the retrofit setting. At 124M, Focus surpasses full attention (30.3 vs 31.4 PPL); trained from scratch at 7B scale (2B tokens), Focus again beats full attention (13.82 vs 13.89 PPL). At inference, restricting each token to its top-k highest-scoring groups discretizes the soft routing into a hard sparsity pattern, yielding 2x speedup while beating the pretrained baseline (41.3 vs 42.8 PPL); decomposing this pattern into two standard FlashAttention calls reaches 8.6x wall-clock speedup at 1M tokens with no custom kernels. Unlike LoRA, centroid routing preserves alignment: instruction-tuned models retain TruthfulQA scores after adaptation, while LoRA degrades at every learning rate and rank. Sinkhorn normalization enforces balanced groups as a hard constraint, and the resulting groups discover interpretable linguistic categories without supervision.

arXiv:2604.04516v1 Announce Type: cross Abstract: Adapting LLMs to new domains causes forgetting because standard methods (full fine-tuning, LoRA) inject new directions into the weight space. We propose GAIN, which re-emphasizes existing features through multiplicative modulation W_new = S * W. The learned diagonal matrix S is applied to the attention output projection and optionally the FFN. The principle mirrors gain modulation in neuroscience, where neurons adapt to context by scaling response strength while preserving selectivity. We evaluate GAIN on five models from four families (774M to 70B), adapting sequentially across eight domains. GAIN-FFN matches LoRA's in-domain adaptation, but their effects on previously trained domains are opposite: GAIN-FFN improves them by 7-13% (validation PPL), while LoRA degrades them by 18-36%. Downstream accuracy confirms the pattern: for example, after seven sequential adaptations on Qwen2.5, GAIN-FFN degrades BoolQ by only 0.8% while LoRA damages it by 14.9%. GAIN adds 46K-230K parameters per model and can be absorbed into the pretrained weights for zero inference cost.

arXiv:2604.04843v1 Announce Type: cross Abstract: Human-object-scene interactions (HOSI) generation has broad applications in embodied AI, simulation, and animation. Unlike human-object interaction (HOI) and human-scene interaction (HSI), HOSI generation requires reasoning over dynamic object-scene changes, yet suffers from limited annotated data. To address these issues, we propose a coarse-to-fine instruction-conditioned interaction generation framework that is explicitly aligned with the iterative denoising process of a consistency model. In particular, we adopt a dynamic perception strategy that leverages trajectories from the preceding refinement to update scene context and condition subsequent refinement at each denoising step of consistency model, yielding consistent interactions. To further reduce physical artifacts, we introduce a bump-aware guidance that mitigates collisions and penetrations during sampling without requiring fine-grained scene geometry, enabling real-time generation. To overcome data scarcity, we design a hybrid training startegy that synthesizes pseudo-HOSI samples by injecting voxelized scene occupancy into HOI datasets and jointly trains with high-fidelity HSI data, allowing interaction learning while preserving realistic scene awareness. Extensive experiments demonstrate that our method achieves state-of-the-art performance in both HOSI and HOI generation, and strong generalization to unseen scenes. Project page: https://yudezou.github.io/InfBaGel-page/

arXiv:2506.17585v3 Announce Type: replace Abstract: Trustworthy language models should provide both correct and verifiable answers. However, citations generated directly by standalone LLMs are often unreliable. As a result, current systems insert citations by querying an external retriever at inference time, introducing latency, infrastructure dependence, and vulnerability to retrieval noise. We explore whether LLMs can be made to reliably attribute to the documents seen during continual pretraining without test-time retrieval, by revising the training process. To study this, we construct CitePretrainBench, a benchmark that mixes real-world corpora (Wikipedia, Common Crawl, arXiv) with novel documents and probes both short-form (single-fact) and long-form (multi-fact) citation tasks. Our approach follows a two-stage process: (1) continual pretraining to index factual knowledge by binding it to persistent document identifiers; and (2) instruction tuning to elicit citation behavior. We introduce Active Indexing for the first stage, which creates generalizable, source-anchored bindings by augmenting training with synthetic data that (i) restate each fact in diverse, compositional forms and (ii) enforce bidirectional training (source-to-fact and fact-to-source). This equips the model to both generate content from a cited source and attribute its own answers, improving robustness to paraphrase and composition. Experiments with Qwen-2.5-7B&3B show that Active Indexing consistently outperforms a Passive Indexing baseline, which simply appends an identifier to each document, achieving citation precision gains of up to 30.2% across all tasks and models. Our ablation studies reveal that performance continues to improve as we scale the amount of augmented data, showing a clear upward trend even at 16x the original token count. Finally, we show that internal citations complement external ones by making the model more robust to retrieval noise.