arXiv subjects

Shangming Chen

Publications and source records attributed to Shangming Chen.

3 recordsLinked to original sources

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↗

arXiv2023

Dynamical Analysis of an Allelopathic Phytoplankton Model with Fear Effect

This paper is the first to propose an allelopathic phytoplankton competition ODE model influenced by a fear effect based on natural biological phenomena. It is shown that the interplay of this fear effect and the allelopathic term cause rich dynamics in the proposed competition model, such as global stability, transcritical bifurcation, pitchfork bifurcation, and saddle-node bifurcation. We also consider the spatially explicit version of the model and prove analogous results. Numerical simulations verify the feasibility of the theoretical analysis. The results demonstrate that the primary cause of the extinction of non-toxic species is the fear of toxic species compared to toxins. Allelopathy only affects the density of non-toxic species. The discussion provides guidance for the conservation of species and the maintenance of biodiversity.

math.DS↗

arXiv2023

Dynamical Analysis of a Lotka-Volterra Competition Model with both Allee and Fear Effect

Population ecology theory is replete with density dependent processes. However trait-mediated or behavioral indirect interactions can both reinforce or oppose density-dependent effects. This paper presents the first two species competitive ODE and PDE systems where an Allee effect, which is a density dependent process and the fear effect, which is non-consumptive and behavioral are both present. The stability of the equilibria is discussed analytically using the qualitative theory of ordinary differential equations. It is found that the Allee effect and the fear effect change the extinction dynamics of the system and the number of positive equilibrium points, but they do not affect the stability of the positive equilibria. We also observe some special dynamics that induce bifurcations in the system by varying the Allee or fear parameter. Interestingly we find that the Allee effect working in conjunction with the fear effect, can bring about several qualitative changes to the dynamical behavior of the system with only the fear effect in place, in regimes of small fear. That is, for small amounts of the fear parameter, it can change a competitive exclusion type situation to a strong competition type situation. It can also change a weak competition type situation to a bi-stability type situation. However for large fear regimes the Allee effect reinforces the dynamics driven by the fear effect. The analysis of the corresponding spatially explicit model is also presented. To this end the comparison principle for parabolic PDE is used. The conclusions of this paper have strong implications for conservation biology, biological control as well as the preservation of biodiversity.

q-bio.PE↗