Things I have worked on.
Safeguard-Conditioned Uplift: Measuring Utility-Risk Frontiers for Dual-Use Biology Assistants
Abstract: Safety evaluations for dual-use biology assistants often measure base-model capability, refusal behavior, or jailbreak success. These metrics miss a deployment question: for a fixed base model, how does the access condition users actually see change benign utility and harmful actionable assistance? I introduce safeguard-conditioned uplift, a protocol for comparing deployed access conditions through a human-judged utility-risk frontier. I evaluate Claude Sonnet 4.6 and Gemini 3.5 Flash under helpful prompting, safety prompting, and an external safeguarded assistant on a 108-task surrogate benchmark, with the headline claim restricted to a locked 18-task held-out split. In a 600-row blinded human audit, the safeguarded assistant reduces harmful actionability relative to helpful prompting by -0.063 over 49 matched response pairs, with bootstrap 95% interval [-0.117, -0.011], while correctness changes by +0.009 with interval [-0.057, +0.077]. Adaptive, Test-B, cue-ablation, and controller-baseline checks support the measurement story but also show non-dominance: safety prompting is often strongest for Claude, while external control helps more for Gemini and can reduce benign utility. The contribution is not a universal defense. It is a deployment-level evaluation target, plus a learned risk-budgeted calibration procedure, for measuring how user-facing access conditions move the utility-risk frontier.
Safety Geometry in Multi-turn Conversations: When Static Steering Directions Stop Separating
Abstract: Activation steering often assumes that a behavioral concept can be controlled by a single residual-stream direction estimated from positive and negative examples. We show that this assumption breaks in multi-turn dialogue: because the residual state at a fixed extraction position is conditioned by the ordered conversational prefix, a turn-0 direction can stop separating the same concept after several turns of pressure. We construct a conversation-conditioned binary-pair dataset over five safety-relevant concepts and four dialogue contexts, and decompose static-steering failure into signal collapse versus direction rotation. Across 15 model variants from 6 families, simple linear predictors recover most of the lost separation from first-order residual information. However, the behavioral recovery does not imply stable concept geometry: recovered directions remain close to concept centroids even as top-PC subspace overlap with turn 0 drops substantially. Multi-turn steering drift is therefore often recoverable at the level of first-order contrast, while the broader concept subspace can still move.
Beyond RAG: Enabling Explicit Memory in Pretrained LLMs for Enhanced Reasoning and Efficiency
Abstract: Explicit memory refers to highly sparse attention key-value pairs derived from the reference text. As a form of knowledge that offers lower decoding costs compared to plain-text retrieval augmented generation (RAG) and lower encoding costs compared to model parameters, explicit memory can help large language mod els (LLMs) reduce the cost of acquiring new knowledge. Previous work has explored train ing models from scratch to reason with explicit memory, achieving significant capability improvements while reducing training and inference costs. However, many existing pretrained models lack the ability to utilize explicit memory. In this work, we aim to study how to enable pretrained models to learn to use explicit memory through supervised fine-tuning (SFT), without forgetting their pretrained knowledge. We curated 120,000 training examples and de signed a training strategy to teach the model to use explicit memory to solve math and coding tasks. Experiment results show that the use of explicit memory can help models solve mathematical problems; however, the improvements are limited and come at the cost of degraded performance on coding tasks. Furthermore, the results reveal that pretrained models are prone to catastrophic forgetting during memory in volving fine-tuning. This work serves as an initial attempt to explore how pretrained models can be equipped with explicit memory, and we plan to continue investigating better training and data strategies in the future research. All of our work can be found on our GitHub homepage: https://github.com/szjiozi/Explicit-Memory
Abstract: Modeling human decision-making under risk and uncertainty remains a significant challenge, with implications in fields like economics and cognitive science. Many decision-making models fail to fully account for individual risk preferences and utility functions. This work investigates the relationship between payoff and utility across diverse risk rofiles, using Blackjack as a simulation to quantify rewards and outcomes. By modeling the game as a deterministic environment, we apply dynamic programming to compute action values for each state and subsequent subgame encountered during gameplay
based on different utility functions, and select the highest-value action, weighted by probabilistic outcomes1. Drawing on cognitive theories such as Prospect Theory and Expected Utility Theory, our findings show that variations in utility functions significantly influence decision-making and financial outcomes, offering new insights into decision-making under un-certainty.