PACT: Can Enterprise AI Assistants Be Trusted Under Pressure?
研究通过引入PACT框架,评估企业级AI助手在压力下遵守规则的能力,揭示了不同模型在合规性上的差异及风险。
研究通过引入PACT框架,评估企业级AI助手在压力下遵守规则的能力,揭示了不同模型在合规性上的差异及风险。
研究通过系统性调整学习率、批次大小等参数,对比LoRA和全微调方法,在不同模型和数据集上优化监督微调效果。
This study investigates whether large language models can continuously learn and accumulate new factual knowledge through weight updates without catastrophically forgetting previously acquired information. Focusing on the Qwen3 model, the authors sequentially inject fictional facts via weight editing and systematically evaluate the impact of different training formulations on knowledge retention and usability using held-out question answering, KL divergence analysis, and context-based recovery tests. The findings reveal that training with diverse phrasings boosts knowledge retention to 46%, compared to merely 1% with a single statement. Moreover, ostensibly “forgotten” facts remain implicitly encoded in the weights and can be recovered to 77–80% accuracy with appropriate prompting. These results suggest that reliable knowledge transfer depends more on contextual activation than on static weight embedding, indicating that forgetting primarily stems from query redirection masking rather than actual knowledge erasure.
This work addresses the severe memory bottleneck caused by KV caching in long-context inference, necessitating compression methods that simultaneously achieve light weight, high expressivity, and cross-trajectory reusability. The authors propose Still, a lightweight Perceiver module embedded at each layer of a frozen base model, which enables synthetic KV cache compression in a single forward pass without per-context optimization. Trained solely on top of frozen foundation models, Still is the first method to jointly satisfy all three desiderata—lightweight design, strong representational capacity, and reusability—enabling efficient inference and flexible summarization even under extreme compression ratios (8–200×). Evaluated on Qwen and Gemma with 128k context lengths, Still significantly outperforms existing approaches, achieving gains of 8–22 points over the strongest baseline on the RULER benchmark and setting new state-of-the-art results on LongBench summarization tasks.
This work addresses the significant redundancy in causal Transformer batch inference, where shared prefixes—such as system prompts or few-shot examples—are redundantly processed through repeated MLP activations. To mitigate this, the authors propose a stateless, prefix-Trie-based compressed inference method that constructs a Trie structure over shared input segments, enabling their unified representation in a single forward pass and restoring outputs only at attention boundaries. This approach uniquely integrates prefix deduplication with position-aware MLPs, LayerNorm, and embedding layers, achieving intra-batch redundancy elimination without caching or maintaining state. Compatible with standard Transformer architectures, the method yields end-to-end speedups of 1.44–1.59× on MS MARCO v1.1 using Qwen3 models (0.6B–8B parameters), with synthetic data demonstrating up to 5× acceleration.
研究通过引入PACT框架,评估企业级AI助手在压力下遵守规则的能力,揭示了不同模型在合规性上的差异及风险。
研究通过系统性调整学习率、批次大小等参数,对比LoRA和全微调方法,在不同模型和数据集上优化监督微调效果。
This study investigates whether large language models can continuously learn and accumulate new factual knowledge through weight updates without catastrophically forgetting previously acquired information. Focusing on the Qwen3 model, the authors sequentially inject fictional facts via weight editing and systematically evaluate the impact of different training formulations on knowledge retention and usability using held-out question answering, KL divergence analysis, and context-based recovery tests. The findings reveal that training with diverse phrasings boosts knowledge retention to 46%, compared to merely 1% with a single statement. Moreover, ostensibly “forgotten” facts remain implicitly encoded in the weights and can be recovered to 77–80% accuracy with appropriate prompting. These results suggest that reliable knowledge transfer depends more on contextual activation than on static weight embedding, indicating that forgetting primarily stems from query redirection masking rather than actual knowledge erasure.
This work addresses the severe memory bottleneck caused by KV caching in long-context inference, necessitating compression methods that simultaneously achieve light weight, high expressivity, and cross-trajectory reusability. The authors propose Still, a lightweight Perceiver module embedded at each layer of a frozen base model, which enables synthetic KV cache compression in a single forward pass without per-context optimization. Trained solely on top of frozen foundation models, Still is the first method to jointly satisfy all three desiderata—lightweight design, strong representational capacity, and reusability—enabling efficient inference and flexible summarization even under extreme compression ratios (8–200×). Evaluated on Qwen and Gemma with 128k context lengths, Still significantly outperforms existing approaches, achieving gains of 8–22 points over the strongest baseline on the RULER benchmark and setting new state-of-the-art results on LongBench summarization tasks.
This work addresses the significant redundancy in causal Transformer batch inference, where shared prefixes—such as system prompts or few-shot examples—are redundantly processed through repeated MLP activations. To mitigate this, the authors propose a stateless, prefix-Trie-based compressed inference method that constructs a Trie structure over shared input segments, enabling their unified representation in a single forward pass and restoring outputs only at attention boundaries. This approach uniquely integrates prefix deduplication with position-aware MLPs, LayerNorm, and embedding layers, achieving intra-batch redundancy elimination without caching or maintaining state. Compatible with standard Transformer architectures, the method yields end-to-end speedups of 1.44–1.59× on MS MARCO v1.1 using Qwen3 models (0.6B–8B parameters), with synthetic data demonstrating up to 5× acceleration.