arXiv subjects

Poulami Ghosh

Publications and source records attributed to Poulami Ghosh.

8 recordsLinked to original sources

arXiv2026

Language Models are not Equally Robust to Non-Canonical Tokenization across Languages

Despite the existence of exponentially many valid tokenizations for a given string, language models operate on a single canonical sequence deterministically produced by the tokenizer, leaving the broader tokenization space largely uncharacterized. In this paper, we investigate this overlooked space by studying the behavior of language models under non-canonical tokenizations across diverse languages. For English, prior work shows that models are largely invariant to alternative tokenizations that represent the same underlying string. We ask whether this invariance generalizes to other languages beyond English. We conduct a multilingual study across 27 languages spanning diverse scripts and evaluate LLM behavior under alternative tokenizations across six downstream tasks. We find that tokenization invariance does not generalize: model behavior varies substantially across languages with instruction-tuned models exhibiting an average relative performance drop of 23.7% for Llama-3.1-8B, 11.4% for Qwen3-8B, and 9.9% for Gemma-3-12B. The variation of tokenization invariance is systematic across languages. Languages that exhibit higher token fragmentation show significantly greater sensitivity to non-canonical tokenizations. Our study of tokenization robustness serves as a diagnostic of how tightly a model is coupled to its tokenizer. These results demonstrate that tokenization robustness is not a universal property of language models, but depends strongly on the language and its interaction with the tokenizer. We also show that LoRA fine-tuning with multi-tokenization training data provides an effective mitigation for tokenization sensitivity. Fine-tuning on English alone improves tokenization robustness across languages, while systematically sampling diverse non-canonical tokenizations achieves the strongest overall performance.

cs.CL↗

arXiv2026

Long-lived relaxation channel and exciton-phonon coupling in Ta2NiSe5 via non-degenerate pump-probe spectroscopy

An excitonic insulator represents a quantum phase in which spontaneous condensation of excitons leads to novel many-body phenomena. Ta$_2$NSi$_5$ (TNSe), a layered narrow-gap semiconductor, has emerged as a model platform to probe these correlated excitonic phases and their underlying dynamics below 327 K. In this work, we investigate the nonequilibrium dynamics of TNSe using temperature-dependent, non-degenerate optical pump-probe spectroscopy with a 3.14 eV pump and a 1.57 eV probe, extending the accessible pump-probe delay window up to 500 ps. In addition to the well-established sub-picosecond relaxation channel ($\sim$ 0.7- 0.9 ps) associated with carrier cooling and recombination, accompanied by exciton reformation, we uncover a much slower recovery process with a decay time of $\sim$ 280-600~ps, significantly longer than previously reported. We attribute this unusually prolonged recovery to enhanced scattering between excitons and nonequilibrium phonons, which delays the re-establishment of equilibrium excitonic correlations. On top of this bi-exponential background, we observe two coherent phonon modes at 1.0 and 2.9 THz with distinctly different coupling behaviors. The 1.0 THz mode exhibits an order-parameter-like temperature dependence, consistent with strong coupling to the excitonic condensate in TNSe. In contrast, the 2.9 THz mode does not exhibit any discernible coupling to the excitonic order parameter, and appears to arise from anharmonic lattice dynamics associated with the structural phase transition. Together, these results elucidate the hierarchy of relaxation pathways in TNSe and highlight the importance of extending the temporal detection window in pump-probe measurements to fully capture long-lived exciton-phonon dynamics.

cond-mat.str-el↗

arXiv2025

Terahertz Time-Domain Spectroscopy and Density Functional Theory Analysis of Low-Frequency Vibrational Modes of a Benzoxazolium-Coumarin Donor-pi-Acceptor Chromophore

To elucidate low-frequency vibrational modes, we investigate a benzoxazolium--coumarin (BCO+) donor-pi-acceptor derivative using transmission terahertz time-domain spectroscopy (THz-TDS). The retrieved complex refractive index reveals distinct modes at 0.62, 0.85, 1.30, 1.81, and 2.07 THz. Gas-phase density functional theory (DFT) agrees well with these features and enables assignment of specific intramolecular motions. Together, THz-TDS and DFT identify the characteristic low-frequency modes of BCO+ and suggest their connection to intramolecular charge transfer-relevant nuclear motions, highlighting that THz-TDS can serve as a sensitive probe of vibrational signatures in donor-pi-acceptor systems.

cond-mat.mtrl-sci↗

arXiv2025

How Good Are LLMs at Processing Tool Outputs?

Most realistic task automation problems require large language models (LLMs) to call tools, which often return complex JSON responses. These responses must be further processed to derive the information necessary for task completion. The ability of LLMs to do so is under-studied. In this paper, we study the tool response processing task and LLMs' abilities to process structured (JSON) responses. We created a dataset for this task, and evaluated 15 open and closed weight models using multiple prompting approaches. Our results show that JSON processing remains a difficult task even for frontier models across multiple prompting strategies. The optimal response processing strategy depends on both the nature and size of the tool outputs, as well as the complexity of the required reasoning. Variations in processing approaches can lead to performance differences ranging from 3\% to 50\%.

cs.LG↗

arXiv2025

Ultrafast Dynamics of Spin-Orbit Entangled Excitons Coupled to Magnetic Ordering in van der Waals Antiferromagnet NiPS3

Spin-orbit entangled excitons (SOEE) in two-dimensional (2D) antiferromagnets provide direct access to explore unconventional many body interactions in correlated electron systems. In this work, we carry out a detailed investigation using non-degenerate isotropic and anisotropic pump-probe reflection spectroscopy to probe the ultrafast dynamics of SOEE and their coupling to spin fluctuations in NiPS3. Transient reflectivity data reveals acoustic phonon oscillations at ~ 27 GHz, along with two distinct relaxation timescales: fast (1-9 ps) and slower components (1-4 ns) associated with SOEE coherence and spin reordering, respectively. Both timescales exhibit pronounced temperature dependence near the exciton dissociation (TED = 120 K) and Neel (TN = 155 K) temperatures. The SOEE coherence shortens from ~ 8-9 ps at T < TED to ~ 3 ps at T > TED with a finite tail persisting beyond TN. The spin reordering time grows near 120 K, and shows critical slowing down around TN. Pump fluence studies further corroborate their spin origin. Our findings uncover the direct interplay between the excitonic and spin degrees of freedom across ultrafast and longer timescales, offering new opportunities to probe and engineer emergent many-body interactions in 2D antiferromagnets.

cond-mat.other↗

arXiv2025

Ultrafast dynamics of carriers, coherent acoustic phonons and strain pulses in BiSbTe1.5Se1.5 topological insulator thin films

We Investigate the ultrafast carrier, coherent acoustic phonons (CAPs), and acoustic strain pulse dynamics in topological insulator BiSbTe1.5Se1.5 (BSTS) thin films of varying thickness using degenerate pump-probe reflection spectroscopy. Here, Sapphire has been chosen as the main substrate due to its maximum acoustic reflectivity at the BSTS-sapphire interface compared to BSTS-GaAs, BSTS-Si, and BSTS-MgO interfaces. For the films with thickness more than twice the penetration depth, the transient reflectivity data predominantly exhibits travelling acoustic strain pulses (TASP) on the top of single-exponential electronic decay (~ 2 ps). In contrast, films with thickness less than penetration depth are dominated by CAPs and a bi exponential electronic background with decay times of ~ 2 ps and ~ 260-380 ps. The observed TASP dynamics are well-described by a theoretical acoustic strain model. Further, to elucidate the underlying physical mechanisms governing the behavior of photo-excited carriers, CAPs, and strain pulses, we performed carrier density and temperature-dependent (7-294 K) studies on BSTS films with thicknesses of 22 nm and 192 nm. In the 22 nm film, the both fast and slow decay processes increase with carrier density at room temperature but decrease with temperature at a carrier density of 1.7*10^{19} cm^{-3}. A detailed analysis suggests that the faster decay arises from electron-phonon scattering and carrier diffusion, while the slower decay likely results from defect-assisted and phonon-assisted recombination. Furthermore, increasing the sample temperature leads to anharmonic decay induced softening of ~ 14 % in the phonon frequency and an anomalous ~ 48 % decrease in the phonon damping parameter due to reduced Dirac surface electron and acoustic phonon scattering.

cond-mat.mtrl-sci↗

arXiv2024

Are Language Models Agnostic to Linguistically Grounded Perturbations? A Case Study of Indic Languages

Pre-trained language models (PLMs) are known to be susceptible to perturbations to the input text, but existing works do not explicitly focus on linguistically grounded attacks, which are subtle and more prevalent in nature. In this paper, we study whether PLMs are agnostic to linguistically grounded attacks or not. To this end, we offer the first study addressing this, investigating different Indic languages and various downstream tasks. Our findings reveal that although PLMs are susceptible to linguistic perturbations, when compared to non-linguistic attacks, PLMs exhibit a slightly lower susceptibility to linguistic attacks. This highlights that even constrained attacks are effective. Moreover, we investigate the implications of these outcomes across a range of languages, encompassing diverse language families and different scripts.

cs.CL↗

arXiv2024

A Morphology-Based Investigation of Positional Encodings

Contemporary deep learning models effectively handle languages with diverse morphology despite not being directly integrated into them. Morphology and word order are closely linked, with the latter incorporated into transformer-based models through positional encodings. This prompts a fundamental inquiry: Is there a correlation between the morphological complexity of a language and the utilization of positional encoding in pre-trained language models? In pursuit of an answer, we present the first study addressing this question, encompassing 22 languages and 5 downstream tasks. Our findings reveal that the importance of positional encoding diminishes with increasing morphological complexity in languages. Our study motivates the need for a deeper understanding of positional encoding, augmenting them to better reflect the different languages under consideration.

cs.CL↗