Four new research papers published on October 17, 2026, present distinct approaches to improving efficiency and reasoning capabilities in large language models.
According to arxiv.org, TOPS (Token Optimal Preservation Sets) addresses computational overhead in multimodal large language models by pruning visual tokens. The researchers formulated token pruning “from a first-principles perspective” and identified “three fundamental principles for effective token selection: Task Relevance, Information Coverage, and Semantic Diversity.” TOPS is described as “training-free and model-agnostic.”
A separate study on arxiv.org examined information density in chain-of-thought reasoning, measuring density using “LP variance” and “Caveman compression ratio.” According to the paper, experiments across six model scales found that “lower LP variance predicts correctness in supervised fine-tuned models, but this relationship weakens with scale and reverses in reinforcement-learning distilled models.” The researchers introduced “Entropy-Gated Density Routing (EGDR), a self-supervised method using token-level Shannon entropy,” which improved accuracy by 10.0 percentage points.
CoTS (Competitive Token Selection), detailed on arxiv.org, proposes “a training-free framework that reframes latent intervention from binary mode selection into continuous allocation at the token level,” using probability competition between top candidates to balance discrete and soft representations.
Finally, Shared Residual Distillation (SRD), according to arxiv.org, enables models to “learn from the model’s own unverified peer trajectories” without external supervision. Testing on Qwen3 models showed substantial improvements, raising “mean AIME24-26 avg@12 from 11.02 to 16.20 on Qwen3-1.7B” and to 54.44 on Qwen3-8B.