Researchers introduced AdaptiveSpec, a training-free per-step speculative decoding method that adapts both draft verification and tree shape from internal decoding signals, improving throughput over E…
A new arXiv study (2609.02893v1) finds that projecting Llama-3.1-8B-Instruct activations onto a small subset of principal components from the training distribution enables cross-domain deception detec…
Researchers propose CARE (Contrastive Anchor-based Rubric Evolution), a method that uses anchor responses from frontier models to dynamically evolve rubrics during LLM post-training, addressing reward…
A new arXiv preprint (2608.28911v1) introduces SemKV, a semantic mixed-precision KV cache quantization method that achieves a measured 6.0x storage reduction for long-context LLM inference with no sta…
Researchers introduced SHIFT-LLM, a training-free framework that corrects distribution shifts in depth-pruned large language models by inserting Linear Residual Adapters at each pruning site, recoveri…
Researchers introduced computation-conditioned credit transport (CCT) and its concrete algorithm CompPO, which uses the Transformer policy's internal attention to parameterize credit assignment in lar…
A new arXiv study (2608.20362v1) finds that exact-match verifiers in reinforcement learning with verifiable rewards (RLVR) introduce language-dependent false-negative reward noise in multilingual sett…
Researchers introduced Attention-Aware Transform Coding (AATC), a KV cache compression method that reduces memory use by approximately 5.8x while maintaining near-lossless accuracy on models like Llam…
A BlueDot Impact Technical AI Safety project demonstrated that large language models (LLMs) can be trained to perform steganography, embedding hidden information in their outputs to evade monitoring. …
A study labeling 7,440 answers from four open-weight LLMs found that more than half of the hallucination labels were incorrect, and correcting them reordered the results. The author, Priyanshi Jain, r…
A study by an anonymous researcher, conducted as part of Neel Nanda's MATS 10.0 stream, found that linear 'trust' vectors extracted from the residual streams of Llama-3.2-3B-Instruct and Llama-3.1-8B-…
A new study from arXiv finds that linear readouts trained on the final-token hidden state of large language models can decode whether diagnostic evidence supports, challenges, or fails to address a ca…
A study presented at the 2026 Annual Meeting of the Special Interest Group on Discourse and Dialogue (SIGDIAL) finds that large language models used as role-playing agents exhibit a trade-off between …
A new study from researchers at Forethought Foundation finds that training language models to be risk-averse on low-stakes gambles (prizes up to $100) can generalize to astronomically high stakes (pri…
A new benchmark evaluates KV-cache optimization techniques—quantization, pruning, and merging—for long-context LLM serving, finding that compression ratio alone poorly predicts end-to-end performance.…
Researchers propose SPSD, an edge-based pipeline that compresses user prompts using a small language model before sending them to a cloud LLM, reducing input tokens by an average of 99.9 per call whil…