Tag
The paper introduces DIVE, a diversity-driven framework that enables frozen LLMs to self-improve by evolving persistent natural-language skills from task experience and verifier feedback, without parameter updates. It outperforms existing methods on math and logical reasoning tasks and transfers across model scales.
This paper introduces SHAPER, a self-evolving framework for embodied agents that keeps model parameters frozen and improves performance by evolving reusable skills and context-code harnesses through target-environment rollouts. Evaluated on VLABench and ESI-Bench, it proposes a practical alternative to fine-tuning when training is expensive or unavailable.
Introduces Latent-IM, a framework for recovering interaction management from frozen speech LLMs using activation-based selection and steering for conversational moves. It improves end-to-end move accuracy by 12.5 points over the unsteered backbone.
This paper investigates why text-based optimization (TextGrad) fails for language agents, showing that while frozen agents can follow good policies, they cannot reliably learn and select policies from their own trajectories.
HALO introduces a hybrid adaptive latent refinement method for frozen language models that selectively applies second-stage refinement to a subset of tokens, achieving better performance than fixed refinement steps while using less compute.
This research introduces a technique to loop frozen, off-the-shelf transformer checkpoints at inference time by using damped Runge-Kutta substeps, treating transformer layers as Euler steps in a residual ODE. This allows extra latent compute without fine-tuning, architecture changes, or new weights, showing gains on knowledge tasks like MMLU-Pro, GPQA, and ARC.
Proposes Life-Harness, a method that improves frozen LLM agents by adapting the runtime interface instead of model weights, achieving an average 88.5% relative improvement across 126 settings and 18 backbones.
Residual Coupling (RC) connects frozen language models in parallel using lightweight learned linear bridges, enabling horizontal scaling without weight modification. It reduces perplexity by up to 80.7% compared to MoE and improves accuracy on TruthfulQA by 9.1 percentage points.