Listening to the Wise Few: Query-Key Alignment Unlocks Latent Correct Answers in Large Language Models
Read the original on arXiv Computation and Language →The Flow has not summarised this story yet — read it at arXiv Computation and Language.
The Flow has not summarised this story yet — read it at arXiv Computation and Language.
arXiv:2602.22453v4 Announce Type: replace Abstract: Retrieval heads, a subset of attention heads in Transformers, were studied in English, showing its crucial role in retrieving information from the...
The paper introduces First-Order In-Context Shapley, an inference‑only method that evaluates data utility for large language model alignment by computing the first‑order term of an in‑context cooperative game over semantic k‑NN neighborhoods in O(k) forward passes. It projects log‑likelihood shifts onto a directed influence graph, yielding three signals—Exemplifying Advantage, Exemplified Advantage, and One‑of‑a‑Kind Score—to identify high‑utility demonstrations and rare prompts. The approach reduces compute by 19× compared to Monte‑Carlo Shapley, filters 1–5% of records, and improves error detection and reward accuracy in DPO training and benchmark auditing. whyItMatters":"By enabling scalable, inference‑only data auditing, the method addresses label noise and safety inversions that hinder post‑training alignment of large language models, leading to measurable improvements in model reliability and reward accuracy."
arXiv:2609.39929v1 Announce Type: cross Abstract: Long-context failures of RoPE-based language models can arise from RoPE's intrinsic tradeoff between maintaining stable token preferences and disting...
arXiv:2608. 15445v1 Announce Type: new Abstract: When a reward is correct on every training example yet consistent with more than one goal, a model can acquire an unintended one, a failure known as goal misgeneralization.
The paper uncovers a systematic mismatch in modern text embeddings: while sentiment can be extracted with high accuracy from frozen encoders, the embeddings’ cosine similarity metric barely reflects polarity differences. Across eleven models from 2021–2026, a linear probe achieves up to 91.3% accuracy, yet cosine similarity margins are only 0.003–0.064, leading to poor retrieval performance (e.g., Qwen3-Embedding-4B’s 64.3 MAP). Introducing a lightweight supervised‑contrastive head (Sentiment‑Injected Sentence Embeddings) dramatically improves probe accuracy, retrieval MAP, and polarity margins, but at the cost of downstream semantic similarity. "whyItMatters":"The findings reveal that widely used retrieval and clustering systems may ignore valuable sentiment information encoded in embeddings, impacting their effectiveness."
arXiv:2607. 01002v1 Announce Type: cross Abstract: In long-context use, large language models frequently synthesize answers from the meaning of a relevant context span rather than literally copy-pasting them.