Searcharxiv⌕ Search

arXiv subjects

Nathan Hu

Publications and source records attributed to Nathan Hu.

13 recordsLinked to original sources

Interpreting Language Model Parameters

Mechanistic interpretability aims to uncover how neural networks use their parameters to implement their impressive neural algorithms. Although previous work has uncovered substantial structure in the intermediate representations that networks use, little progress has been made to understand how the parameters and nonlinearities of networks perform computations on those representations. In this work, we present a method that brings us closer to this understanding by decomposing a language model's parameters into subcomponents that each implement only a small part of the model's learned algorithm, while simultaneously requiring only a small fraction of those subcomponents to account for the network's behavior on any input. The method, adVersarial Parameter Decomposition (VPD), optimizes for decompositions of neural network parameters into simple subcomponents that preserve the network's input-output behavior even when many subcomponents are ablated, including under ablations that are adversarially selected to destroy behavior. This encourages learning subcomponents that provide short, mechanistically faithful descriptions of the network's behavior that should aggregate appropriately into more global descriptions of the network's learned algorithm. We study how sequences of interactions between these parameter subcomponents produce the network's output on particular inputs, enabling a new kind of 'circuit' analysis. While more work remains to be done to deepen our understanding of how neural networks use their parameters to compute their behavior, our work suggests an approach to identify a small set of simple, mechanistically faithful subcomponents on which further mechanistic analysis can be based.

cs.LG↗

Gradient Routing Localizes and Removes Unintended Behaviors in RL

We propose a mitigation to reward hacking in reinforcement learning. Our setup involves a misspecified reward function that gives rise to unintended behavior, and a monitor used to detect unintended behavior. As with frontier AI systems, the monitor makes systematic errors. Consequently, using the monitor for data filtering or to apply a reward penalty fails to prevent the model from learning unintended behavior ("reward hacking"). To address this problem, we introduce GRAFT: Gradient Routing Adapter Fine Tuning. Rather than penalizing or filtering flagged samples during training, GRAFT constrains gradient updates on those samples to a sacrificial set of new parameters. This has two benefits: First, because the monitor is not part of the reward function, the model does not learn to hide misbehavior. Second, ablating the sacrificial parameters reduces or entirely stops the unintended behavior while maintaining improved performance on target tasks. We demonstrate these benefits in RL environments where reward penalties cause models to learn to reward hack in undetected ways, training models ranging from 135M to 8B parameters. Additionally, we show a similar effect at a larger scale by distilling frontier LLM trajectories containing reward hacking into Qwen3-32B, without teaching that model to reward hack. Overall, our work suggests GRAFT may be a promising approach for leveraging imperfect monitors during post-training in settings where true performance is difficult to evaluate.

cs.LG↗

Matryoshka attribution: Learning to attribute language model outputs to representations and weights

Attributing language model outputs to their internal computations is an open problem in interpretability. Existing methods, which use causal interventions, gradients, or learnable masks, either are infeasibly expensive or struggle to identify actual causally-important internal computations. We propose framing attribution as the problem of identifying nested subsets of internal components which minimise a downstream loss. To learn this task, we introduce Matryoshka Attribution (MAttr), a mask learning method that parametrises the mask with a simple differentiable sigmoid top-$k$ operator. We supervise training over all sparsities simultaneously by randomising $k$ over training, resulting in a learned ordering of components by attribution score. MAttr achieves number 1 on the official leaderboard of the Mechanistic Interpretability Benchmark (Mueller et al., 2025); our method identifies sparse and task-transferrable circuits across varying circuit bases. As a practical application, we show that MAttr can be trained with reinforcement learning to identify weight changes responsible for downstream behaviours in LLM finetuning. We train MAttr on refusal judge scores and find that restoring $1\%$ of Llama 3.1 8B Instruct's weights to their base model state is sufficient to remove refusals while maintaining capabilities. We view MAttr as a successful formulation of interpretability into a learnable objective that we can tackle with gradient descent, and encourage future work along these lines.

cs.CL↗

Verbalizing Subliminal Learning Effects Using Text Optimization

Subliminal learning is a phenomenon in which a distillation dataset transmits traits from the teacher model that are not legibly encoded in the dataset itself. This introduces a new challenge for model development and creates new risks from data poisoning. In this work, we use text optimization to detect subliminal learning effects and describe them as legible prompts. Subliminal learning from a prompted teacher motivates our approach. We observe that this is a special case of context distillation and leverage this observation to show that, in theory, the prompted subliminal learning dataset identifies the teacher's prompt. We reduce recovering this prompt to a text optimization problem and present a method to approximately solve it. Our method, SALVE (Search-Aided Latent Verbalization), optimizes a soft prompt, queries the same model to verbalize it as text, and uses beam search to make the verbalization reliable. In the standard subliminal learning setting, SALVE reliably recovers legible prompts that name the teacher's trait, while common text optimization methods fail to do so. In addition, we find that there are settings in which SALVE recovers the teacher's trait from a dataset even when subliminal learning fails, but that modifying student training to improve context distillation can create subliminal learning effects. We lastly show that SALVE detects subliminal learning effects in three additional settings: (1) mixtures of subliminal learning data and unrelated data, (2) data generated when the teacher is biased via activation steering, and (3) subsets of real preference data selected via Logit-Linear Selection. Overall, our results deepen our understanding of subliminal learning and present SALVE as a method to proactively detect subliminal learning effects.

cs.LG↗

PlantRig - From Bones to Branches: Adaptation of Autoregressive Rigging Models for Plant Skeletal Reconstruction

Autoregressive rigging models such as UniRig and SkinTokens perform well on articulated characters, but their ability to generalize to plant structures remains largely unexplored, since plant topologies exhibit highly variable, non-canonical branching patterns that challenge learned skeletal priors. We evaluate these models for plant skeletal reconstruction using synthetic L-system-generated trees and real scanned data spanning monopodial, sympodial, whorled, and vine-like archetypes. Preliminary testing showed UniRig collapsing complex branching into near-linear chains, while SkinTokens preserved topology better but over-segmented branches and produced an unstable output space, so we focused on UniRig for its greater stability. Diagnosis traced the collapse to sampling-level suppression of branch tokens, and further analysis showed the frozen mesh encoder had limited sensitivity to structural variation, pointing to a geometric bottleneck in the tokenization pipeline rather than a purely learned bias. Building on these findings, we applied multi-round fine-tuning over multiple procedurally generated synthetic datasets. Across rounds, the model progressively recovered accurate branching topology and generalized beyond branch-only structures to plants with foliage, a harder case given the zero-thickness, mesh-normal-dependent geometry of leaves. The resulting model generalized well across diverse plant forms without leaf-specific architectural changes, indicating that targeted fine-tuning can substantially close the domain gap between character-rigging priors and plant skeletal structure. As such, our work points toward a viable path for automated plant rigging across both branch topology and foliage type, even those not considered in our findings.

cs.CV↗

LLMs Can Annotate Attribution Graphs

Circuit tracing is an exciting technique for revealing the internal computation of language models, but it requires a time-intensive manual step of grouping individual features or MLP neurons into supernodes. We present a simple pipeline for automating this step: directly presenting feature descriptions to a language model that groups them into supernodes. Using automated interpretability metrics, we confirm that supernodes generated by our pipeline are as interpretable as those generated by human annotators. On a two-hop Capitals task, our pipeline recovers a supernode corresponding to the intermediate hop in 97 of 100 prompts. Finally, we present a simple proof of concept using our pipeline for open-ended exploration, where we automatically annotate 1000 attribution graphs from Wikipedia prompt completions and then use an LLM judge to flag interesting graphs worth human review. We hope this work demonstrates that even simple automation can produce meaningful attribution graph annotations, motivating further work on automated circuit tracing.

cs.LG↗

Transcoder Adapters for Reasoning-Model Diffing

While reasoning models are increasingly ubiquitous, the effects of reasoning training on a model's internal mechanisms remain poorly understood. In this work, we introduce transcoder adapters, a technique for learning an interpretable approximation of the difference in MLP computation before and after fine-tuning. We apply transcoder adapters to characterize the differences between Qwen2.5-Math-7B and its reasoning-distilled variant, DeepSeek-R1-Distill-Qwen-7B. Learned adapters are faithful to the target model's internal computation and next-token predictions. When evaluated on reasoning benchmarks, adapters match the reasoning model's response lengths and typically recover 50-90% of the accuracy gains from reasoning fine-tuning. Adapter features are sparsely activating and interpretable. When examining adapter features, we find that only ~8% have activating examples directly related to reasoning behaviors. We deeply study one such behavior -- the production of hesitation tokens (e.g., "wait"). Using attribution graphs, we trace hesitation to only ~2.4% of adapter features (5.6k total) performing one of two functions. These features are necessary and sufficient for producing hesitation tokens; removing them reduces response length, often without affecting accuracy. Overall, our results provide insight into reasoning training and suggest transcoder adapters may be useful for studying fine-tuning more broadly.

cs.LG↗

Confucius Code Agent: Scalable Agent Scaffolding for Real-World Codebases

Real-world software engineering tasks require coding agents that can operate on massive repositories, sustain long-horizon sessions, and reliably coordinate complex toolchains at test time. Existing research-grade coding agents offer transparency but struggle when scaled to heavier, production-level workloads, while production-grade systems achieve strong practical performance but provide limited extensibility, interpretability, and controllability. We introduce the Confucius Code Agent (CCA), a software engineering agent that can operate at large-scale codebases. CCA is built on top of the Confucius SDK, an agent development platform structured around three complementary perspectives: Agent Experience (AX), User Experience (UX), and Developer Experience (DX). The SDK supports a unified orchestrator with advanced context management for long-context reasoning, a persistent note-taking system for cross-session continual learning, and a modular extension system for reliable tool use. In addition, we introduce a meta-agent that automates the construction, evaluation, and refinement of agents through a build-test-improve cycle, enabling rapid agent development on new tasks and tool stacks. Instantiated on the Confucius SDK using the meta-agent, CCA demonstrates strong performance on real-world software engineering tasks. On SWE-Bench-Pro, CCA achieves a Resolve@1 of 59%, exceeding prior research baselines as well as commercial results, under identical repositories, model backends, and tool access.

cs.CL↗

Measuring Sparse Autoencoder Feature Sensitivity

Sparse Autoencoder (SAE) features have become essential tools for mechanistic interpretability research. SAE features are typically characterized by examining their activating examples, which are often "monosemantic" and align with human interpretable concepts. However, these examples don't reveal feature sensitivity: how reliably a feature activates on texts similar to its activating examples. In this work, we develop a scalable method to evaluate feature sensitivity. Our approach avoids the need to generate natural language descriptions for features; instead we use language models to generate text with the same semantic properties as a feature's activating examples. We then test whether the feature activates on these generated texts. We demonstrate that sensitivity measures a new facet of feature quality and find that many interpretable features have poor sensitivity. Human evaluation confirms that when features fail to activate on our generated text, that text genuinely resembles the original activating examples. Lastly, we study feature sensitivity at the SAE level and observe that average feature sensitivity declines with increasing SAE width across 7 SAE variants. Our work establishes feature sensitivity as a new dimension for evaluating both individual features and SAE architectures.

cs.AI↗

High-Fidelity Cellular Network Control-Plane Traffic Generation without Domain Knowledge

With rapid evolution of mobile core network (MCN) architectures, large-scale control-plane traffic (CPT) traces are critical to studying MCN design and performance optimization by the R&D community. The prior-art control-plane traffic generator SMM heavily relies on domain knowledge which requires re-design as the domain evolves. In this work, we study the feasibility of developing a high-fidelity MCN control plane traffic generator by leveraging generative ML models. We identify key challenges in synthesizing high-fidelity CPT including generic (to data-plane) requirements such as multimodality feature relationships and unique requirements such as stateful semantics and long-term (time-of-day) data variations. We show state-of-the-art, generative adversarial network (GAN)-based approaches shown to work well for data-plane traffic cannot meet these fidelity requirements of CPT, and develop a transformer-based model, CPT-GPT, that accurately captures complex dependencies among the samples in each traffic stream (control events by the same UE) without the need for GAN. Our evaluation of CPT-GPT on a large-scale control-plane traffic trace shows that (1) it does not rely on domain knowledge yet synthesizes control-plane traffic with comparable fidelity as SMM; (2) compared to the prior-art GAN-based approach, it reduces the fraction of streams that violate stateful semantics by two orders of magnitude, the max y-distance of sojourn time distributions of streams by 16.0%, and the transfer learning time in deriving new hourly models by 3.36x.

cs.NI↗

Long-form factuality in large language models

Large language models (LLMs) often generate content that contains factual errors when responding to fact-seeking prompts on open-ended topics. To benchmark a model's long-form factuality in open domains, we first use GPT-4 to generate LongFact, a prompt set comprising thousands of questions spanning 38 topics. We then propose that LLM agents can be used as automated evaluators for long-form factuality through a method which we call Search-Augmented Factuality Evaluator (SAFE). SAFE utilizes an LLM to break down a long-form response into a set of individual facts and to evaluate the accuracy of each fact using a multi-step reasoning process comprising sending search queries to Google Search and determining whether a fact is supported by the search results. Furthermore, we propose extending F1 score as an aggregated metric for long-form factuality. To do so, we balance the percentage of supported facts in a response (precision) with the percentage of provided facts relative to a hyperparameter representing a user's preferred response length (recall). Empirically, we demonstrate that LLM agents can outperform crowdsourced human annotators - on a set of ~16k individual facts, SAFE agrees with crowdsourced human annotators 72% of the time, and on a random subset of 100 disagreement cases, SAFE wins 76% of the time. At the same time, SAFE is more than 20 times cheaper than human annotators. We also benchmark thirteen language models on LongFact across four model families (Gemini, GPT, Claude, and PaLM-2), finding that larger language models generally achieve better long-form factuality. LongFact, SAFE, and all experimental code are available at https://github.com/google-deepmind/long-form-factuality.

cs.CL↗

Meta-Learning Online Adaptation of Language Models

Large language models encode impressively broad world knowledge in their parameters. However, the knowledge in static language models falls out of date, limiting the model's effective "shelf life." While online fine-tuning can reduce this degradation, we find that naively fine-tuning on a stream of documents leads to a low level of information uptake. We hypothesize that online fine-tuning does not sufficiently attend to important information. That is, the gradient signal from important tokens representing factual information is drowned out by the gradient from inherently noisy tokens, suggesting that a dynamic, context-aware learning rate may be beneficial. We therefore propose learning which tokens to upweight. We meta-train a small, autoregressive model to reweight the language modeling loss for each token during online fine-tuning, with the objective of maximizing the out-of-date base question-answering model's ability to answer questions about a document after a single weighted gradient step. We call this approach Context-aware Meta-learned Loss Scaling (CaMeLS). Across three different distributions of documents, our experiments find that CaMeLS provides substantially improved information uptake on streams of thousands of documents compared with standard fine-tuning and baseline heuristics for reweighting token losses.

cs.CL↗

Sampling Arborescences in Parallel

We study the problem of sampling a uniformly random directed rooted spanning tree, also known as an arborescence, from a possibly weighted directed graph. Classically, this problem has long been known to be polynomial-time solvable; the exact number of arborescences can be computed by a determinant [Tut48], and sampling can be reduced to counting [JVV86, JS96]. However, the classic reduction from sampling to counting seems to be inherently sequential. This raises the question of designing efficient parallel algorithms for sampling. We show that sampling arborescences can be done in RNC. For several well-studied combinatorial structures, counting can be reduced to the computation of a determinant, which is known to be in NC [Csa75]. These include arborescences, planar graph perfect matchings, Eulerian tours in digraphs, and determinantal point processes. However, not much is known about efficient parallel sampling of these structures. Our work is a step towards resolving this mystery.

cs.DS↗