Searcharxiv⌕ Search

arXiv subjects

Chaipat Tirapongprasert

Publications and source records attributed to Chaipat Tirapongprasert.

5 recordsLinked to original sources

Tracking Representations in LLM Reasoning Traces for Iterated Games

Large language model (LLM) agents achieve state-of-the-art performance across domains through chain-of-thought (CoT) reasoning. Despite their ubiquitous use, little is understood about the relationship between the externalized trace and the model's actual computation. We address this by investigating how LLM agents make decisions in iterative games with incomplete information. Using a series of mechanistic interpretability techniques, we find that the model forms and iteratively revises an internal representation of the game throughout the CoT. Once the model prematurely commits to an incorrect belief, it rarely recovers. This failure is manifested in both the literal trace and the layer activations. Ultimately, as multi-agent systems are more prevalent, our work opens the door to a more focused diagnosis of their strategic reasoning errors. For illustrative purposes, we offer an interactive visualizer of our methodology on selected sample traces at https://interp-self.github.io/gametrace/

cs.CL↗

Agentic Systems for Calibrating Cosmological Simulations

Calibrating a cosmological simulation is an expensive, high-dimensional parameter optimization task. Previous works employ Bayesian optimization for automation, which treats the simulator as a black box and precludes the incorporation of human domain knowledge or workflow into the selection policy. We propose AgenticTuner, an agentic calibration pipeline with a present-day large language model serving as the optimizer who proposes the next parameter trial. We show that these agents can achieve competitive calibration performance at a fraction of the CPU cost when they are given informal, human-readable priors (such as codebase, documentation, and heuristic guides that seasoned scientists may possess). We hope this work serves as an early step toward AI scientists that can carry out complex, long-horizon scientific tasks.

astro-ph.IM↗

Don't Cut Corners: How Training Outside the Prior Makes Simulation-Based Inference More Robust

Large astrophysical simulation campaigns often generate training data by sampling parameters across a Uniform prior box. Due to the proposal's sharp edge, neural posterior estimators struggle to learn accurate approximations near the boundaries. We propose Tailed-Uniform, a family of hybrid proposal distributions for sampling training simulations for robust simulation-based inference. By padding the original hard-truncated training box with decaying tails, Tailed-Uniform-trained networks yield more accurate posteriors near and beyond the edges. We demonstrate these improvements on a family of tail shapes, including a widened Uniform box as a control. Our results suggest that additional simulations near the prior boundary better constrain the networks as it approaches the edge of the training box, even for Uniform assumed priors. We show these advantages on a toy problem and cosmological parameter inference from the matter power spectrum. These benefits increase in high dimensions, where boundaries dominate parameter space volume.

astro-ph.IM↗

DegenDetector: Symbolic Recovery of Parameter Degeneracies in Bayesian Posteriors

We introduce DegenDetector, a framework for identifying and characterizing parameter degeneracies in posterior distributions as closed-form symbolic equations. By combining mutual information screening with alternating symbolic regression, we facilitate automated and interpretable identification of degenerate relationships without domain-specific input. While standard tools such as corner plots can indicate that correlations exist, they do not reveal the underlying functional form. DegenDetector fills this gap by expressing multi-parameter degeneracies as closed-form equations, providing interpretable structure that scales to high-order parameter spaces.

astro-ph.IM↗

Learning at the Edge: Tailed-Uniform Sampling for Robust Simulation-Based Inference

We introduce the Tailed-Uniform proposal distribution for generating training simulations in simulation-based inference. Instead of sampling parameters uniformly within bounded regions, we extend the distribution beyond prior boundaries with smooth Gaussian tails. This eliminates sharp transitions that cause neural posterior estimators to fail when the posterior distribution intersects or extends beyond the prior bounds. We show these benefits on a toy problem and cosmological parameter inference from the matter power spectrum. Such an advantage grows in high dimensions, where boundaries dominate parameter space volume. All code is publicly available on Github at https://github.com/chaipattira/tailed-uniform-sbi.

astro-ph.IM↗