Shu Yuan | arXiv Science

arXiv2026

MAIGO: Mitigating Lost-in-Conversation with History-Cleaned On-Policy Self-Distillation

Large language models often solve tasks from a fully specified prompt but degrade when the same requirements unfold over multiple turns, known as the lost-in-conversation (LiC) gap. We trace part of this degradation to self-contamination: intermediate assistant replies enter later context and carry early deviations forward. Motivated by this mechanism, we propose MAIGO, an on-policy self-distillation method that reduces this contamination using history-cleaned references from the model's own policy. For middle turns, MAIGO removes prior assistant replies while preserving the user-visible sharded prefix; for answer turns, it distills from paired full-view references conditioned on the completed user-side dialogue. A reliability weight downweights middle-turn samples that disagree with the clean reference. MAIGO requires no verifier rewards, state labels, or inference-time scaffolding. Under the LiC paired-view protocol with deterministic verifiers, MAIGO improves Qwen2.5-7B-Instruct SHARDED accuracy from 52.8 to 66.1 and the SHARDED/FULL ratio from 66.5% to 84.1%, while keeping FULL accuracy within 2.3 points. These results show that self-contamination is a trainable component of the LiC gap.

cs.CL↗

arXiv2019

Spectro-polarimetric Observations at the NVST: I. Instrumental Polarization Calibration and Primary Measurements

This paper is devoted to the primary spectro-polarimetric observation performed at the New Vacuum Solar Telescope of China since 2017, and our aim is to precisely evaluate the real polarimetric accuracy and sensitivity of this polarimetry by using full Stokes spectro-polarimetric observations of the photospheric line Fe I 532.4 nm. In the work, we briefly describe the salient characteristic of the NVST as a polarimeter in technology and then characterize its instrumental polarization based on the operation in 2017 and 2019. It is verified that the calibration method making use of the instrumental polarization calibration unit (ICU) is stable and credible. The calibration accuracy can reach up to 3$\times 10^{-3}$ . Based on the scientific observation of the NOAA 12645 on April 5th, 2017, we estimate that the residual cross-talk from Stokes $I$ to Stokes $Q$, $U$ and $V$, after the instrumental polarization calibration, is about 4$\times10^{-3}$ on average, which is consistent with the calibration accuracy and close to the photon noise. The polarimetric sensitivity (i.e., the detection limit) for polarized light is of the order of $10^{-3}$ with an integration time over 20 seconds. Slow modulation rate is indeed an issue for the present system. The present NVST polarimeter is expected to be integrated with an high-order adaptive optics system and a field scanner to realize 2D magnetic field vector measurements in the following instrumentation update.

astro-ph.SR↗