Brendan Hogan Rappazzo

Xiwen Chen, Wenhui Zhu, Jingjing Wang et al.

Aligning Large Language Models (LLMs) with human preferences is often formulated via Direct Preference Optimization (DPO). However, the standard Bradley-Terry instantiation of DPO is limited in modeling common departures from transitivity in human preferences. To address this, recent work has introduced Self-Play Preference Optimization (SPPO), which iteratively refines the policy by training on self-generated win-lose pairs. Our investigation, however, reveals a critical instability in SPPO: the optimization is prone to policy degeneration when the preference oracle assigns overly confident wins to semantically indistinguishable responses. To mitigate this, we propose S-SPPO, a dual-space semantic calibration framework comprising: i) Supervision Calibration via semantic gating, which anneals win rate targets toward the maximum-entropy baseline as semantic overlap increases; and ii) Representation Calibration via latent repulsion to enforce geometric diversity to prevent manifold collapse and maintain latent diversity between chosen and rejected samples. Theoretically, we show that the calibration preserves the constant-sum game structure, facilitating convergence to a Nash Equilibrium. Empirically, S-SPPO avoids the performance degradation seen in prior methods, achieving 52.19% win rate and 47.46% length-controlled win rate on AlpacaEval 2.0 with Llama-3-8B, without using additional human-annotated preferences during training. The code will be available at https://github.com/xiwenc1/s-sppo.

Wenhui Zhu, Xiwen Chen, Zhipeng Wang et al.

Long-horizon LLM agents require memory systems that remain accurate under fixed context budgets. However, existing systems struggle with two persistent challenges in long-term dialogue: (i) \textbf{disconnected evidence}, where multi-hop answers require linking facts distributed across time, and (ii) \textbf{state updates}, where evolving information (e.g., schedule changes) creates conflicts with older static logs. We propose AriadneMem, a structured memory system that addresses these failure modes via a decoupled two-phase pipeline. In the \textbf{offline construction phase}, AriadneMem employs \emph{entropy-aware gating} to filter noise and low-information message before LLM extraction and applies \emph{conflict-aware coarsening} to merge static duplicates while preserving state transitions as temporal edges. In the \textbf{online reasoning phase}, rather than relying on expensive iterative planning, AriadneMem executes \emph{algorithmic bridge discovery} to reconstruct missing logical paths between retrieved facts, followed by \emph{single-call topology-aware synthesis}. On LoCoMo experiments with GPT-4o, AriadneMem improves \textbf{Multi-Hop F1 by 15.2\%} and \textbf{Average F1 by 9.0\%} over strong baselines. Crucially, by offloading reasoning to the graph layer, AriadneMem reduces \textbf{total runtime by 77.8\%} using only \textbf{497} context tokens. The code is available at https://github.com/LLM-VLM-GSL/AriadneMem.

Brendan Hogan Rappazzo, Yingheng Wang, Aaron Ferber et al.

The ability to form, retrieve, and reason about memories in response to stimuli serves as the cornerstone for general intelligence - shaping entities capable of learning, adaptation, and intuitive insight. Large Language Models (LLMs) have proven their ability, given the proper memories or context, to reason and respond meaningfully to stimuli. However, they are still unable to optimally encode, store, and retrieve memories - the ability to do this would unlock their full ability to operate as AI agents, and to specialize to niche domains. To remedy this, one promising area of research is Retrieval Augmented Generation (RAG), which aims to augment LLMs by providing them with rich in-context examples and information. In question-answering (QA) applications, RAG methods embed the text of interest in chunks, and retrieve the most relevant chunks for a prompt using text embeddings. Motivated by human memory encoding and retrieval, we aim to improve over standard RAG methods by generating and encoding higher-level information and tagging the chunks by their utility to answer questions. We introduce Graphical Eigen Memories For Retrieval Augmented Generation (GEM-RAG). GEM-RAG works by tagging each chunk of text in a given text corpus with LLM generated ``utility'' questions, connecting chunks in a graph based on the similarity of both their text and utility questions, and then using the eigendecomposition of the memory graph to build higher level summary nodes that capture the main themes of the text. We evaluate GEM-RAG, using both UnifiedQA and GPT-3.5 Turbo as the LLMs, with SBERT, and OpenAI's text encoders on two standard QA tasks, showing that GEM-RAG outperforms other state-of-the-art RAG methods on these tasks. We also discuss the implications of having a robust RAG system and future directions.

Brendan Hogan Rappazzo, Aaron Ferber, Carla Gomes

Modern neural network classifiers achieve remarkable performance across a variety of tasks; however, they frequently exhibit overconfidence in their predictions due to the cross-entropy loss. Inspired by this problem, we propose the \textbf{Cr}i\textbf{t}ic Loss for Image \textbf{Cl}assification (CrtCl, pronounced Critical). CrtCl formulates image classification training in a generator-critic framework, with a base classifier acting as a generator, and a correctness critic imposing a loss on the classifier. The base classifier, acting as the generator, given images, generates the probability distribution over classes and intermediate embeddings. The critic model, given the image, intermediate embeddings, and output predictions of the base model, predicts the probability that the base model has produced the correct classification, which then can be back propagated as a self supervision signal. Notably, the critic does not use the label as input, meaning that the critic can train the base model on both labeled and unlabeled data in semi-supervised learning settings. CrtCl represents a learned loss method for accuracy, alleviating the negative side effects of using cross-entropy loss. Additionally, CrtCl provides a powerful way to select data to be labeled in an active learning setting, by estimating the classification ability of the base model on unlabeled data. We study the effectiveness of CrtCl in low-labeled data regimes, and in the context of active learning. In classification, we find that CrtCl, compared to recent baselines, increases classifier generalization and calibration with various amounts of labeled data. In active learning, we show our method outperforms baselines in accuracy and calibration. We observe consistent results across three image classification datasets.