📃 Enhancing Logical Reasoning in Language Models via Symbolically-Guided Monte Carlo Process Supervision. 🌐 @akhter_elahi .@xrysoflhs .@nikaletras.

- Symbolic ReAct-guided trajectories also improve multi-step reasoning LLMs over unguided trajectories (1%-4%) on out-of-domain evaluation (claim verification datasets that require deductive and abductive reasoning).

We found that:. - Symbolic ReAct-guided trajectories improve multi-step reasoning LLMs over unguided trajectories (2%-6%) on FOLIO and LogicAsker.

We investigate sampling trajectories in a guided way using Symbolic ReAct format for Monte Carlo estimation and sampling the trajectories for fine-tuning multi-step reasoning LLMs.

Can LLMs learn to reason in a sound and accurate way by fine-tuning on automatically sampled symbolic reasoning trajectories?.

However, the trajectories for fine-tuning the multi-step reasoning #LLMs and PRMs are unguided, leading to #LLMs generating verbose reasoning traces in informal languages.

Multi-step reasoning #LLMs are often fine-tuned on trajectories that are filtered with process reward models (PRMs) automatically trained on Monte Carlo estimation-generated pseudo labels.

Additionally, the system features conversation summarization capabilities to distill critical information from lengthy exchanges, as well as persona analysis to characterize the speakers involved.

We also implemented source identification methodologies to pinpoint key segments within the conversation that contribute to the LLM's analytical output.

Leveraging Large Language Models (LLMs) for detailed conversational analysis, the system provides explanatory justifications for identified instances of harmful speech.

In partnership with the UK Forensic Network, we have developed a system designed to analyze conversational data and identify instances of harmful speech, with a specific focus on Violence Against Women and Girls (VAWG) content.

If you are attending #NAACL2025 in person, you are welcome to check out our Safespeech DEMO in Hall 3 (9:00-10:30).

Welcome to join our oral talk on "Cascading Large Language Models for Salient Event Graph Generation" starting from 14:00 on Thursday at Albuquerque Convention Center Ruidoso room. #NAACL2025.

We implement source identification methods to pinpoint key segments within the conversation that contribute to the LLM's analytical output. This platform also serves as a research tool, enabling users to choose datasets and integrate diverse models for performance evaluation.

We leverage Large Language Models (LLMs) for detailed conversational analysis, the system provides explanatory justifications for identified instances of harmful speech.

In partnership with the UK Forensic Network, we have developed a system designed to analyze conversational data and identify instances of harmful speech, with a specific focus on Violence Against Women and Girls (VAWG) content. Checkout our #naacl2025 DEMO.