arXiv subjects

Jan Chojnacki

Publications and source records attributed to Jan Chojnacki.

8 recordsLinked to original sources

arXiv2025

Interpretable Risk Mitigation in LLM Agent Systems

Autonomous agents powered by large language models (LLMs) enable novel use cases in domains where responsible action is increasingly important. Yet the inherent unpredictability of LLMs raises safety concerns about agent reliability. In this work, we explore agent behaviour in a toy, game-theoretic environment based on a variation of the Iterated Prisoner's Dilemma. We introduce a strategy-modification method-independent of both the game and the prompt-by steering the residual stream with interpretable features extracted from a sparse autoencoder latent space. Steering with the good-faith negotiation feature lowers the average defection probability by 28 percentage points. We also identify feasible steering ranges for several open-source LLM agents. Finally, we hypothesise that game-theoretic evaluation of LLM agents, combined with representation-steering alignment, can generalise to real-world applications on end-user devices and embodied platforms.

cs.AI↗

arXiv2024

Non-Linear Inference Time Intervention: Improving LLM Truthfulness

In this work, we explore LLM's internal representation space to identify attention heads that contain the most truthful and accurate information. We further developed the Inference Time Intervention (ITI) framework, which lets bias LLM without the need for fine-tuning. The improvement manifests in introducing a non-linear multi-token probing and multi-token intervention: Non-Linear ITI (NL-ITI), which significantly enhances performance on evaluation benchmarks. NL-ITI is tested on diverse multiple-choice datasets, including TruthfulQA, on which we report over 16% relative MC1 (accuracy of model pointing to the correct answer) improvement with respect to the baseline ITI results. Moreover, we achieved a 10% relative improvement over the recently released Truth Forest (TrFf) method that also focused on ITI improvement.

cs.CL↗

arXiv2023

Stability of the de-Sitter spacetime. The anisotropic case

We elaborate on the cosmological implications of the recently established non-perturbative $O(D,D$)-symmetric approach. Following the $α'$-complete Friedmann equations, we provide a qualitative description of the dynamics of anisotropy perturbations. We find that the spacetime stability depends on the sign of the generalized dilaton time derivative. We consider isotropic limits and find possible mechanism for pre-big bang cosmology, governed by the value of cosmological constant.

hep-th↗

arXiv2022

Non-linear stability of $\alpha'$-corrected Friedmann equations

We study the non-linear stability of fixed-point solutions to the $\alpha'$-exact equations from O$(d,d)$ invariant cosmology, with and without matter perturbations. Previous non-linear analysis in the literature is revisited, and its compatibility with known linear perturbation results is shown. Some formal aspects of cosmological perturbations in duality invariant cosmology are discussed, and we show the existence of time-reparameterization invariant variables for perturbations.

hep-th↗

arXiv2022

Stability of the non-perturbative $O(D,D)$ de-Sitter spacetime. The isotropic case

The existence of de-Sitter solutions and their stability within string theory became, in recent years, one of the central research questions within the string theory community. The so-called de-Sitter conjecture states that the de-Sitter vacua are unstable within string theory. Here, we study the de-Sitter solutions in string theory and their stability in the time direction using the $O(d,d)$ symmetry approach allowing for the calculation of the generic form of the string theory effective action. We study the undiscussed case of the constant generalized dilaton as well as the general form of dilaton couplings. Our findings suggest that de-Sitter solutions are unstable for the constant generalized dilaton both with and without matter. Furthermore, we found that in order for the theory to be weakly coupled the general form of dilaton coupling is in the form of a cosmological constant.

hep-th↗

arXiv2021

Finite Action Principle and Horava-Lifshitz Gravity: early universe, black holes and wormholes

The destructive interference of the neighbouring field configurations with infinite classical action in the gravitational path integral approach serves as a dynamical mechanism resolving the black hole singularity problem. It also provides an isotropic and homogeneous early universe without the need of inflation. In this work, we elaborate on the finite action in the framework of Horava-Lifshitz gravity -- a ghost-free QFT. Assuming the mixmaster chaotic solutions in the projectable H-L theory, we show that the beginning of the universe is homogeneous and isotropic. Furthermore, we show that the H-L gravity action selects only the regular black-hole spacetimes. We also comment on possibility of traversable wormholes in theories with higher curvature invariants.

gr-qc↗

arXiv2021

Finite Action Principle and wormholes

In this work, we elaborate on the finite action for wormholes in higher derivative theories. Both non-traversable and traversable wormholes in theories with higher curvature invariants posses finite action.

gr-qc↗

arXiv2021

Is asymptotically safe inflation eternal?

Recently, based on swampland considerations in string theory, the (no) eternal inflation principle has been put forward. The natural question arises whether similar conditions hold in other approaches to quantum gravity. In this article, the asymptotic safety hypothesis is considered in the context of eternal inflation. As exemplary inflationary models the SU(N) Yang-Mills in the Veneziano limit and various RG-improvements of the gravitational action are studied. The existence of UV fixed point generically flattens the potential and our findings suggest no tension between eternal inflation and asymptotic safety, both in the matter and gravitational sector in contradistinction to string theory. Moreover, the eternal inflation cannot take place in the range of applicability of effective field quantum gravity theory. We employ the analytical relations for eternal inflation to some of the models with single minima, such as Starobinsky inflation, alpha-attractors, or the RG-improved models and verify them with the massive numerical simulations. The validity of these constraints is also discussed for a multi-minima model.

gr-qc↗