Mutable Inference: Architecting LLM Accelerators for Inference-Time Learning
Yimin Gao, Liangtao Dai, Melika Morsali, Junting Huo, Mircea Stan
AI 速读 基于摘要
提出Learning While Serving(LWS)加速器架构,让推理时学习(ITL)的权重更新与延迟敏感的预测前向计算共享硬件,从而在保持预测吞吐的同时隐藏学习开销。
cs.AR
查看原文摘要
Long-context LLMs must retain growing histories, but full attention makes each new token increasingly expensive. Inference-time learning (ITL) offers another approach. The model keeps training on the context during inference, retaining what it learns in its weights. Recent work shows that ITL can match full-context attention quality while attending only to a recent fixed-size window. This bounds attention costs but adds Learning (backward computation and weight updates) alongside Prediction (the forward pass) costs. Inference accelerators assume fixed weights, while training accelerators target complete training steps. This work identifies two opportunities. Attention's non-matrix operations leave enough idle matrix capacity for Learning, while selected layers can tolerate delayed weight updates with little quality loss. We exploit both with Learning While Serving (LWS), an inference accelerator architecture that shares compute hardware between latency-critical Prediction and deadline-flexible Learning. Hardware-managed per-layer weight versioning allows Learning to publish updates while in-flight computations read older versions. Resumable reduction blocks let Learning kernels span fragmented idle windows and yield to Prediction, while dedicated softmax-gradient units relieve vector contention. On modeled datacenter and edge NPUs, LWS hides Learning within Prediction, sustaining Prediction-only throughput while preserving 99.3% of fully fresh Learning's quality benefit for TTT-E2E 3B at 128K tokens. Across nine models, LWS achieves 1.6x the throughput of serial execution and 1.3x that of the best prior co-execution architecture. RTL synthesis shows that LWS adds 1.5% to NPU core area without another MAC array. By efficiently turning idle inference capacity into learning capacity, LWS establishes inference-time learning as a first-class workload for future LLM accelerators.
Zhimin Gao, Meibo Hu, Jiamian Wang, Zhiqiang Tao, Pichao Wang
原文摘要
Reasoning models are trained to produce correct answers, yet selective deployment requires ranking which answers to accept or defer, and outcome-reward training does not directly optimize the post-rationale distribution used for that decision. We identify two mismatches: expected correctness is mode-seeking rather than distribution-eliciting, and scoring an average across rationales rewards disagreement unavailable to a single-rationale decision. We introduce Post-Rationale Distribution Optimization (PRDO), which applies a proper score to the answer distribution after each sampled rationale, without target rationales or separate confidence labels. Across four 0.6--3.8-billion-parameter models on MMLU-Pro, PRDO improves Brier score, error ranking and secondary accuracy relative to budget-matched, temperature-scaled GRPO. At 20\% realized error, oracle coverage rises by \npj{Q3a_vs_grpo.qwen3-0.6b.mp.oracle_cov@0.2.ppabs}--\npj{Q3a_vs_grpo.phi-4-mini.mp.oracle_cov@0.2.ppabs} percentage points, and ranking gains transfer to unseen tasks. Matched controls identify proper, per-rationale scoring as the active design choices, while verification tasks show the boundary: better selective ranking can coexist with lower full-coverage accuracy. Code is available at https://github.com/ZimmyGao/openjev-rlcd
Shangxin Yang, Zongwei Wang, Jing Long, Junliang Yu, Min Gao
原文摘要
Semantic ID-based recommendation represents each item as a sequence of semantic tokens obtained through residual quantization. This paradigm generates the next item's tokens autoregressively from coarse to fine, entailing each prediction attending to historical tokens at all semantic levels. However, our inference-time analysis reveals that historical tokens at the target-token's semantic level tend to be more informative, whereas other levels can introduce interference. Motivated by these findings, we propose LESSRec that separately encodes the historical token sequence at each semantic level and restricts attention over historical tokens to the sequence at the target token's semantic level during prediction. To complement these level-specific histories, LESSRec explicitly provides an item-level semantic sequence that represents each historical item's overall semantics in chronological order. The decoder accesses the level-aligned token sequence and the item-level semantic sequence through two separate cross-attention branches and fuses their contexts. We further introduce LESSRec+, which uses separate encoders for different semantic-level histories with the same total parameter count to reduce encoder computation. Extensive experiments across three public datasets and three tokenizers demonstrate the recommendation effectiveness of LESSRec and LESSRec+, with efficiency analyses highlighting the computational benefits of LESSRec+.
Chi Wang, Min Gao, Yifei Zhang, Bincheng Gu, Zongwei Wang, Wentao Li
原文摘要
Fake news undermines public trust and the integrity of online information ecosystems. Temporally disordered fake news is especially elusive because it rearranges factual events into a false chronology. Individual events may remain factually valid even when their temporal relations are misrepresented. When similar events involving the same entity recur, retrieval may return semantically plausible evidence from the wrong period. Consequently, independent event verification may rely on locally coherent but temporally mismatched evidence and incorrectly validate the article. We call this the \textit{evidence coherence trap}. Because existing benchmarks rarely isolate this setting, we construct ChronoFake, a source-disjoint benchmark of 30,706 articles derived from temporally anchored real events. It distinguishes temporal manipulation from valid narration, including flashbacks. To address this, we propose DelSA, a \underline{D}ocument S\underline{el}ection \underline{S}elf-Updating \underline{A}gent that iteratively refines evidence selection during training and applies a frozen rulebook for one-pass evidence selection at inference. DelSA uses an LLM to characterize candidate evidence, while a verifier-trained reranker balances local relevance against timeline consistency. Finally, the classifier compares the article-implied chronology extracted by the LLM with the evidence-derived chronology to reveal global temporal conflicts. DelSA outperforms existing methods on ChronoFake, while remaining competitive on PolitiFact and GossipCop.