arXiv subjects

Michael Li

Publications and source records attributed to Michael Li.

15 recordsLinked to original sources

arXiv2026

How Much Do Circuits Tell Us? Measuring the Consistency and Specificity of Language Model Circuits

The circuits framework in mechanistic interpretability aims to identify sparse subgraphs of model components that are causally responsible for a behavior, typically evaluated by measuring necessity and sufficiency. But these criteria say little about whether a circuit consistently captures how a model performs a task, or if it is specific to that task. We study these two properties, consistency and specificity, across six tasks and five models, extracting circuits at the component level (attention heads and MLP blocks) and at the level of individual MLP neurons. We find that component-level circuits are highly consistent and causally important on most tasks, but they are not specific: ablating one task's circuit damages another task's performance about as much as that task's own circuit does. Neuron-level circuits, on the other hand, exhibit higher task-specificity but are far less consistent within tasks. This is explained by circuit overlap: component-level circuits share most of their components across all task pairs, related or not, while neuron-level circuits overlap only between closely related tasks. In a case study of the components shared by the task circuits of Llama-3.2-3B, we show that they consist mostly of MLP blocks, while the few attention heads within turn out to be generic attention-sink heads. Overall, our findings raise questions about the degree to which circuits can support targeted understanding of, and intervention on, model behavior.

cs.CL↗

arXiv2026

Beyond the Answer: Decoding the Behavior of LLMs as Scientific Reasoners

As Large Language Models (LLMs) achieve increasingly sophisticated performance on complex reasoning tasks, current architectures serve as critical proxies for the internal heuristics of frontier models. Characterizing emergent reasoning is vital for long-term interpretability and safety. Furthermore, understanding how prompting modulates these processes is essential, as natural language will likely be the primary interface for interacting with AGI systems. In this work, we use a custom variant of Genetic Pareto (GEPA) to systematically optimize prompts for scientific reasoning tasks, and analyze how prompting can affect reasoning behavior. We investigate the structural patterns and logical heuristics inherent in GEPA-optimized prompts, and evaluate their transferability and brittleness. Our findings reveal that gains in scientific reasoning often correspond to model-specific heuristics that fail to generalize across systems, which we call "local" logic. By framing prompt optimization as a tool for model interpretability, we argue that mapping these preferred reasoning structures for LLMs is an important prerequisite for effectively collaborating with superhuman intelligence.

cs.AI↗

arXiv2026

Towards Fluent Interaction with Cyber-Physical Architecture

What happens when your walls begin to move? This paper explores the design of human-robot interaction for architectural-scale, shape-changing environments. We present findings from two studies: (1) a series of speculative design workshops (N=20) that uncovered aspirational visions for these spaces, and (2) a task-based Wizard-of-Oz elicitation study (N=12) that grounded these visions in the challenges of practical interaction. Our workshop findings reveal a complex landscape of user desires, exposing critical tensions between proactive automation and the preservation of user autonomy, and between personalization and public ownership. Our elicitation study reveals a set of core interaction challenges related to multimodal collaboration; and, most critically: suggests the need for a modality-agnostic model of evolving user intent. We conclude with a set of grounded proposals for creating robotic environments that are collaborative and trusted partners in everyday life.

cs.HC↗

arXiv2026

Step 3.5 Flash: Open Frontier-Level Intelligence with 11B Active Parameters

We introduce Step 3.5 Flash, a sparse Mixture-of-Experts (MoE) model that bridges frontier-level agentic intelligence and computational efficiency. We focus on what matters most when building agents: sharp reasoning and fast, reliable execution. Step 3.5 Flash pairs a 196B-parameter foundation with 11B active parameters for efficient inference. It is optimized with interleaved 3:1 sliding-window/full attention and Multi-Token Prediction (MTP-3) to reduce the latency and cost of multi-round agentic interactions. To reach frontier-level intelligence, we design a scalable reinforcement learning framework that combines verifiable signals with preference feedback, while remaining stable under large-scale off-policy training, enabling consistent self-improvement across mathematics, code, and tool use. Step 3.5 Flash demonstrates strong performance across agent, coding, and math tasks, achieving 85.4% on IMO-AnswerBench, 86.4% on LiveCodeBench-v6 (2024.08-2025.05), 88.2% on tau2-Bench, 69.0% on BrowseComp (with context management), and 51.0% on Terminal-Bench 2.0, comparable to frontier models such as GPT-5.2 xHigh and Gemini 3.0 Pro. By redefining the efficiency frontier, Step 3.5 Flash provides a high-density foundation for deploying sophisticated agents in real-world industrial environments.

cs.CL↗

arXiv2025

Generative Modeling for Robust Deep Reinforcement Learning on the Traveling Salesman Problem

The Traveling Salesman Problem (TSP) is a classic NP-hard combinatorial optimization task with numerous practical applications. Classic heuristic solvers can attain near-optimal performance for small problem instances, but become computationally intractable for larger problems. Real-world logistics problems such as dynamically re-routing last-mile deliveries demand a solver with fast inference time, which has led researchers to investigate specialized neural network solvers. However, neural networks struggle to generalize beyond the synthetic data they were trained on. In particular, we show that there exist TSP distributions that are realistic in practice, which also consistently lead to poor worst-case performance for existing neural approaches. To address this issue of distribution robustness, we present Combinatorial Optimization with Generative Sampling (COGS), where training data is sampled from a generative TSP model. We show that COGS provides better data coverage and interpolation in the space of TSP training distributions. We also present TSPLib50, a dataset of realistically distributed TSP samples, which tests real-world generalization ability without conflating this issue with instance size. We evaluate our method on various synthetic datasets as well as TSPLib50, and compare to state-of-the-art neural baselines. We demonstrate that COGS improves distribution robustness, with most performance gains coming from worst-case scenarios.

cs.LG↗

arXiv2025

Model Internal Sleuthing: Finding Lexical Identity and Inflectional Features in Modern Language Models

Large transformer-based language models dominate modern NLP, yet our understanding of how they encode linguistic information relies primarily on studies of early models like BERT and GPT-2. We systematically probe 25 models from BERT Base to Qwen2.5-7B focusing on two linguistic properties: lexical identity and inflectional features across 6 diverse languages. We find a consistent pattern: inflectional features are linearly decodable throughout the model, while lexical identity is prominent early but increasingly weakens with depth. Further analysis of the representation geometry reveals that models with aggressive mid-layer dimensionality compression show reduced steering effectiveness in those layers, despite probe accuracy remaining high. Pretraining analysis shows that inflectional structure stabilizes early while lexical identity representations continue evolving. Taken together, our findings suggest that transformers maintain inflectional features across layers, while trading off lexical identity for compact, predictive representations. Our code is available at https://github.com/ml5885/model_internal_sleuthing

cs.CL↗

arXiv2024

Non-Stabilizing Parallel Chip-Firing Games

In 2010, Kominers and Kominers proved that any parallel chip-firing game on $G(V,\,E)$ with $|\sigma|\geq 4|E|-|V|$ chips stabilizes. Recently, Bu, Choi, and Xu made the bound exact: all games with $|\sigma|< |E|$ chips or $|\sigma|> 3|E|-|V|$ chips stabilize. Meanwhile, Levine found a "devil's staircase'' pattern in the plot of the activity of parallel chip-firing games against their density of chips. The stabilizing bound of Bu, Choi, and Xu corresponds to the top and bottom stairs of this staircase, in which the activity is 1 and 0, respectively. In this paper, we analyze the middle stair of the staircase, corresponding to activity $\frac{1}{2}$. We prove that all parallel chip-firing games with $2|E|-|V|< |\sigma|< 2|E|$ have period $T\neq 3,\,4$. In fact, this is exactly the range of $|\sigma|$ for which all games are non-stabilizing. We conjecture that all parallel chip-firing games with $2|E|-|V|< |\sigma|<2|E|$ have $T=2$ and thus activity $\frac{1}{2}$. This conjecture has been proven for trees by Bu, Choi, and Xu, cycles by Dall'asta, and complete graphs by Levine. We extend Levine's method of conjugate configurations to prove the conjecture on complete bipartite graphs $K_{a,a}$.

math.CO↗

arXiv2024

Periods and atomic firing sequences of parallel chip-firing games on directed graphs

In 1992, Bitar and Goles introduced the parallel chip-firing game on undirected graphs. Two years later, Prisner extended the game to directed graphs. While the properties of parallel chip-firing games on undirected graphs have been extensively studied, their analogs for parallel chip-firing games on directed graphs have been sporadic. In this paper, we prove the outstanding analogs of the core results of parallel chip-firing games on undirected graphs for those on directed graphs. We find the possible periods of a parallel chip-firing game on a directed simple cycle and introduce the method of Gauss-Jordan elimination on a Laplacian-like matrix to establish a lower bound on the maximum period of a parallel chip-firing game on an orientation of an undirected complete graph and an undirected complete bipartite graph. Finally, we expand the method of motors by Jiang, Scully, and Zhang to directed graphs to show that a binary string $s$ can be the atomic firing sequence of a vertex in a parallel chip-firing game on a strongly connected directed graph if and only if $s$ contains $1$ or $s=0$.

math.CO↗

arXiv2024

View From Above: A Framework for Evaluating Distribution Shifts in Model Behavior

When large language models (LLMs) are asked to perform certain tasks, how can we be sure that their learned representations align with reality? We propose a domain-agnostic framework for systematically evaluating distribution shifts in LLMs decision-making processes, where they are given control of mechanisms governed by pre-defined rules. While individual LLM actions may appear consistent with expected behavior, across a large number of trials, statistically significant distribution shifts can emerge. To test this, we construct a well-defined environment with known outcome logic: blackjack. In more than 1,000 trials, we uncover statistically significant evidence suggesting behavioral misalignment in the learned representations of LLM.

cs.CL↗

arXiv2024

Increasingly Many Bounded Eigenvalues of the Graph of Whitehead Moves

In this paper, we investigate the eigenvalues of the Laplacian matrix of the "graph of graphs", in which cubic graphs of order n are joined together using Whitehead moves. Our work follows recent results from arXiv:2303.13923 , which discovered a significant "bottleneck" in the graph of graphs. We found that their bottleneck implies an eigenvalue of order at most O(1). In fact, our main contribution is to expand upon this result by showing that the graph of graphs has increasingly many bounded eigenvalues as n increases to infinity. We also show that these eigenvalues are unusually small, in the sense that they are much smaller than the eigenvalues of a random regular graph with an equal number of vertices and a similar degree.

math.CO↗

arXiv2023

NAS-X: Neural Adaptive Smoothing via Twisting

Sequential latent variable models (SLVMs) are essential tools in statistics and machine learning, with applications ranging from healthcare to neuroscience. As their flexibility increases, analytic inference and model learning can become challenging, necessitating approximate methods. Here we introduce neural adaptive smoothing via twisting (NAS-X), a method that extends reweighted wake-sleep (RWS) to the sequential setting by using smoothing sequential Monte Carlo (SMC) to estimate intractable posterior expectations. Combining RWS and smoothing SMC allows NAS-X to provide low-bias and low-variance gradient estimates, and fit both discrete and continuous latent variable models. We illustrate the theoretical advantages of NAS-X over previous methods and explore these advantages empirically in a variety of tasks, including a challenging application to mechanistic models of neuronal dynamics. These experiments show that NAS-X substantially outperforms previous VI- and RWS-based methods in inference and model learning, achieving lower parameter error and tighter likelihood bounds.

cs.LG↗

arXiv2023

Toward a comprehensive system for constructing compartmental epidemic models

Compartmental models are valuable tools for investigating infectious diseases. Researchers building such models typically begin with a simple structure where compartments correspond to individuals with different epidemiological statuses, e.g., the classic SIR model which splits the population into susceptible, infected, and recovered compartments. However, as more information about a specific pathogen is discovered, or as a means to investigate the effects of heterogeneities, it becomes useful to stratify models further -- for example by age, geographic location, or pathogen strain. The operation of constructing stratified compartmental models from a pair of simpler models resembles the Cartesian product used in graph theory, but several key differences complicate matters. In this article we give explicit mathematical definitions for several so-called ``model products'' and provide examples where each is suitable. We also provide examples of model stratification where no existing model product will generate the desired result.

q-bio.PE↗

arXiv2022

Evaluating undercounts in epidemics: response to Maruotti et al. 2022

Maruotti et al. 2022 used a mark-recapture approach to estimate bounds on the true number of monkeypox infections in various countries. These approaches are fundamentally flawed; it is impossible to estimate undercounting based solely on a single stream of reported cases. Simulations based on a Richards curve for cumulative incidence show that, for reasonable epidemic parameters, the proposed methods estimate bounds on the ascertainment ratio of $\approx 0.2-0.5$ roughly independently of the true ascertainment ratio. These methods should not be used.

q-bio.PE↗

arXiv2021

Generative Probabilistic Image Colorization

We propose Generative Probabilistic Image Colorization, a diffusion-based generative process that trains a sequence of probabilistic models to reverse each step of noise corruption. Given a line-drawing image as input, our method suggests multiple candidate colorized images. Therefore, our method accounts for the ill-posed nature of the colorization problem. We conducted comprehensive experiments investigating the colorization of line-drawing images, report the influence of a score-based MCMC approach that corrects the marginal distribution of estimated samples, and further compare different combinations of models and the similarity of their generated images. Despite using only a relatively small training dataset, we experimentally develop a method to generate multiple diverse colorization candidates which avoids mode collapse and does not require any additional constraints, losses, or re-training with alternative training conditions. Our proposed approach performed well not only on color-conditional image generation tasks using biased initial values, but also on some practical image completion and inpainting tasks.

cs.CV↗

arXiv2021

Minimax Optimal Conditional Density Estimation under Total Variation Smoothness

This paper studies the minimax rate of nonparametric conditional density estimation under a weighted absolute value loss function in a multivariate setting. We first demonstrate that conditional density estimation is impossible if one only requires that $p_{X|Z}$ is smooth in $x$ for all values of $z$. This motivates us to consider a sub-class of absolutely continuous distributions, restricting the conditional density $p_{X|Z}(x|z)$ to not only be H\"older smooth in $x$, but also be total variation smooth in $z$. We propose a corresponding kernel-based estimator and prove that it achieves the minimax rate. We give some simple examples of densities satisfying our assumptions which imply that our results are not vacuous. Finally, we propose an estimator which achieves the minimax optimal rate adaptively, i.e., without the need to know the smoothness parameter values in advance. Crucially, both of our estimators (the adaptive and non-adaptive ones) impose no assumptions on the marginal density $p_Z$, and are not obtained as a ratio between two kernel smoothing estimators which may sound like a go to approach in this problem.

math.ST↗