arXiv subjects

Robert West

Publications and source records attributed to Robert West.

At least 19 recordsLinked to original sources

arXiv2026

Diff Mining: Logit Differences Reveal Finetuning Objectives

Finetuning has become the gold standard for refining existing behaviors and inducing new ones in language models, yet it often remains unclear exactly which behaviors emerge during this process. As models grow ever more capable, understanding finetuning better becomes increasingly important, particularly since unwanted behaviors may arise during finetuning. In this paper, we introduce Diff Mining, a simple yet effective framework for identifying what a finetuned model has learned by comparing its logits to those of its base model. Diff Mining effectively surfaces salient tokens that are amplified in the finetuned model, serving as a fingerprint of its training -- even on text unrelated to the finetuning domain. Unlike many existing model diffing methods which require model internals, Diff Mining only needs access to output logits and scales to large models. The framework consists of two modular stages: (i) extracting per-context logit differences between the finetuned and base models on a reference corpus, and (ii) aggregating the resulting signals to construct an interpretable token set representing the finetune. For aggregation, we explore both a simple Top-K frequency method and a Non-negative Matrix Factorization (NMF)-based approach for disentangling multiple finetuning objectives into distinct token clusters. Empirically, Diff Mining succeeds across diverse settings: on finetune domain detection, it significantly outperforms state-of-the-art model diffing methods both in identifying relevant tokens and in downstream performance when an interpretability agent is given access to the extracted token set; on models with injected biases, it identifies more than one third of the biases without targeted probing. Overall, our framework shows promise in developing auditing tools to detect finetuning objectives.

cs.LG↗

arXiv2026

Synthetic Persona Pretraining: Alignment from Token Zero

As language-model-based AI is increasingly deployed in autonomous settings, aligning its goals and values with those of humans becomes critical. Today, alignment, and the assistant identity itself, are typically introduced only after pretraining, once behavioral priors are already established. This can make values a thin overlay, rather than deeply rooted, and facilitate subsequent misalignment. Pursuing a different paradigm, we introduce Synthetic Persona Pretraining (SPP), which installs the desired assistant persona from token zero in pretraining. First, we annotate pretraining documents with value-aligned first-person reflections derived from a normative value constitution. Second, we pretrain via the standard cross-entropy loss on standard pretraining documents as well as their reflections, which installs the desired persona among a multitude of other personas. Finally, we post-train on user-assistant dialogue data, which binds this desired persona to the assistant identity, a process we call persona binding. By pretraining models up to 3B parameters on 500B tokens, we show that SPP improves constitution following and jailbreak robustness, and reduces the misalignment rate in out-of-distribution moral dilemmas, while preserving capabilities. Early intervention matters: compared with alignment from token zero, introducing SPP only at the end of pretraining yields weaker constitution adherence, does not shift value priorities, and leads to less aligned choices in dilemmas. This advantage depends on persona binding and, importantly, increases with pretraining budget. Overall, our results show that shaping values early is critical for alignment and establish pretraining-time persona interventions as an effective approach to do so.

cs.LG↗

arXiv2026

Tandem Reinforcement Learning with Verifiable Rewards

Reinforcement learning with verifiable rewards (RLVR) has significantly improved the reasoning capability of large language models, reaching expert or even superhuman performance in domains such as competition math. However, whether weaker agents and humans can actually harness this capability is far less certain, with RLVR documented to drift reasoning toward idiosyncratic patterns such as poor readability and language mixing. Tandem training is a recently introduced paradigm that targets this compatibility problem: a trained, stronger senior co-generates each rollout with a frozen, weaker junior, and the two are rewarded as a team, so the senior is pushed to reason in ways the junior can follow. Yet this paradigm has so far been demonstrated only in proof-of-concept settings, leaving open whether it scales to the long chains of thought of the modern RLVR pipeline. In this work, we propose Tandem Reinforcement Learning (TRL), which carries the tandem training paradigm into RLVR. In TRL, the senior and a frozen junior alternate stochastically to co-generate the reasoning, the resulting generation is rewarded, and the standard GRPO loss is applied to the senior. Training Qwen3-4B-Instruct on competition math, we find that TRL matches vanilla GRPO on solo reasoning capability while three properties emerge together from the same rollout structure: stronger handoff robustness with the junior, reduced distributional drift from the junior, and a chain-of-thought more legible to the junior. Our results demonstrate a promising route for RLVR with practical payoffs in multi-model communication and human compatibility.

cs.AI↗

arXiv2026

HF Etching and Silanization: Evidence for the Role of Surface Hydroxyl Groups in Silicon Nitride Resonator Loss

Silicon nitride $SiN_x$ nanomechanical resonators are central to sensing, quantum technologies, and fundamental physics experiments due to their exceptional mechanical quality factors (Q). However, as resonator thickness approaches the nano-scale, surface-related dissipation limits performance. Here, we investigate the role of surface chemistry in low-stress Si-rich SiNx membranes through a combination of hydrofluoric acid (HF) etching and trimethylchlorosilane (TMCS) silanization, correlated with surface characterization and mechanical measurements. Preliminary analysis by TEM-EELS, XPS, RBS/ERDA, and XRR reveals a native oxide surface layer (1-2 nm). Surface modification by HF and TMCS was subsequently evaluated using XPS, photothermal FTIR, contact-angle measurements, and intrinsic quality factor ($Q_{int}$) characterization. While HF etching effectively removes the native oxide and TMCS introduces hydrophobic $Si-(CH_3)_3$ termination, neither oxide thickness nor surface energy correlates with mechanical dissipation. TMCS treatments produce the largest enhancements, increasing $Q_{int}$ by up to 50%, whereas HF etching alone yields lower gains of 20-25%. These findings suggest surface hydroxyl groups as a key contributor to energy loss in $SiN_x$ resonators and demonstrate that chemical functionalization can substantially suppress surface dissipation.

cond-mat.mtrl-sci↗

arXiv2026

Tracing Persona Vectors Through LLM Pretraining

How large language models internally represent high-level behaviors is a core interpretability question with direct relevance to AI safety: it determines what we can detect, audit, or intervene on. Recent work has shown that traits such as evil or sycophancy correspond to linear directions in the internal activations, the so-called persona vectors. Although these vectors are now routinely utilized to inspect and steer model behavior in safety-relevant settings, how these representations are formed during training remains unknown. To address this gap, we trace persona vectors across the pretraining of OLMo-3-7B, finding that persona vectors form remarkably early -- within 0.22% of OLMo-3 pretraining -- and remain effective for steering the fully post-trained instruct models. Although core representations are formed early on, persona vectors continue to refine geometrically and semantically throughout pretraining. We further compare alternative elicitation strategies and find that all yield effective directions, with each strategy surfacing qualitatively distinct facets of the underlying persona. Replicating our analysis on Apertus-8B reveals that our findings transfer qualitatively beyond OLMo-3. Our results establish persona representations as stable features of early pretraining and open a path to studying how training forms, refines, and shapes them.

cs.CL↗

arXiv2026

Integrating Machine-Generated Short Descriptions into the Wikipedia Android App: A Pilot Deployment of Descartes

Short descriptions are a key part of the Wikipedia user experience, but their coverage remains uneven across languages and topics. In previous work, we introduced Descartes, a multilingual model for generating short descriptions. In this report, we present the results of a pilot deployment of Descartes in the Wikipedia Android app, where editors were offered suggestions based on outputs from Descartes while editing short descriptions. The experiment spanned 12 languages, with over 3,900 articles and 375 editors participating. Overall, 90% of accepted Descartes descriptions were rated at least 3 out of 5 in quality, and their average ratings were comparable to human-written ones. Editors adopted machine suggestions both directly and with modifications, while the rate of reverts and reports remained low. The pilot also revealed practical considerations for deployment, including latency, language-specific gaps, and the need for safeguards around sensitive topics. These results indicate that Descartes's short descriptions can support editors in reducing content gaps, provided that technical, design, and community guardrails are in place.

cs.CL↗

arXiv2025

TiME: Tiny Monolingual Encoders for Efficient NLP Pipelines

Today, a lot of research on language models is focused on large, general-purpose models. However, many NLP pipelines only require models with a well-defined, small set of capabilities. While large models are capable of performing the tasks of those smaller models, they are simply not fast enough to process large amounts of data or offer real-time responses. Furthermore, they often use unnecessarily large amounts of energy, leading to sustainability concerns and problems when deploying them on battery-powered devices. In our work, we show how to train small models for such efficiency-critical applications. As opposed to many off-the-shelf NLP pipelines, our models use modern training techniques such as distillation, and offer support for low-resource languages. We call our models TiME (Tiny Monolingual Encoders) and comprehensively evaluate them on a range of common NLP tasks, observing an improved trade-off between benchmark performance on one hand, and throughput, latency and energy consumption on the other. Along the way, we show that distilling monolingual models from multilingual teachers is possible, and likewise distilling models with absolute positional embeddings from teachers with relative positional embeddings.

cs.CL↗

arXiv2025

Carrot, stick, or both? Price incentives for sustainable food choice in competitive environments

Meat consumption is a major driver of global greenhouse gas emissions. While pricing interventions have shown potential to reduce meat intake, previous studies have focused on highly constrained environments with limited consumer choice. Here, we present the first large-scale field experiment to evaluate multiple pricing interventions in a real-world, competitive setting. Using a sequential crossover design with matched menus in a Swiss university campus, we systematically compared vegetarian-meal discounts (-2.5 CHF), meat surcharges (+2.5 CHF), and a combined scheme (-1.2 CHF=+1.2 CHF) across four campus cafeterias. Only the surcharge and combined interventions led to significant increases in vegetarian meal uptake--by 26.4% and 16.6%, respectively--and reduced CO2 emissions per meal by 7.4% and 11.3%, respectively. The surcharge, while effective, triggered a 12.3% drop in sales at intervention sites and a corresponding 14.9% increase in non-treated locations, hence causing a spillover effect that completely offset environmental gains. In contrast, the combined approach achieved meaningful emission reductions without significant effects on overall sales or revenue, making it both effective and economically viable. Notably, pricing interventions were equally effective for both vegetarian-leaning customers and habitual meat-eaters, stimulating change even within entrenched dietary habits. Our results show that balanced pricing strategies can reduce the carbon footprint of realistic food environments, but require coordinated implementation to maximize climate benefits and avoid unintended spillover effects.

econ.GN↗

arXiv2025

Cmprsr: Abstractive Token-Level Question-Agnostic Prompt Compressor

Motivated by the high costs of using black-box Large Language Models (LLMs), we introduce a novel prompt compression paradigm, under which we use smaller LLMs to compress inputs for the larger ones. We present the first comprehensive LLM-as-a-compressor benchmark spanning 25 open- and closed-source models, which reveals significant disparity in models' compression ability in terms of (i) preserving semantically important information (ii) following the user-provided compression rate (CR). We further improve the performance of gpt-4.1-mini, the best overall vanilla compressor, with Textgrad-based compression meta-prompt optimization. We also identify the most promising open-source vanilla LLM - Qwen3-4B - and post-train it with a combination of supervised fine-tuning (SFT) and Group Relative Policy Optimization (GRPO), pursuing the dual objective of CR adherence and maximizing the downstream task performance. We call the resulting model Cmprsr and demonstrate its superiority over both extractive and vanilla abstractive compression across the entire range of compression rates on lengthy inputs from MeetingBank and LongBench as well as short prompts from GSM8k. The latter highlights Cmprsr's generalizability across varying input lengths and domains. Moreover, Cmprsr closely follows the requested compression rate, offering fine control over the cost-quality trade-off.

cs.CL↗

arXiv2025

From Model Training to Model Raising

Current AI training methods align models with human values only after their core capabilities have been established, resulting in models that are easily misaligned and lack deep-rooted value systems. We propose a paradigm shift from "model training" to "model raising", in which alignment is woven into a model's development from the start. We identify several key components for this paradigm, all centered around redesigning the training corpus: reframing training data from a first-person perspective, recontextualizing information as lived experience, simulating social interactions, and scaffolding the ordering of training data. We expect that this redesign of the training corpus will lead to an early commitment to values from the first training token onward, such that knowledge, skills, and values are intrinsically much harder to separate. In an ecosystem in which large language model capabilities start overtaking human capabilities in many tasks, this seems to us like a critical need.

cs.AI↗

arXiv2025

The Collaboration Gap: Exploration and Benchmarking of Open-World Agentic Cooperation

The trajectory of AI development suggests that we will increasingly rely on agent-based systems powered by language models, composed of independently developed agents with different information, privileges, and tools. The success of these systems will depend critically on effective collaboration among these heterogeneous agents, even under partial observability. Despite intense interest, the literature lacks empirical studies evaluating agentic collaboration without relying on fixed communication protocols, limiting insights for open-world deployments. We propose an illustrative collaborative maze-solving benchmark that (i) isolates collaborative capabilities, (ii) modulates problem complexity, (iii) enables scalable automated grading, and (iv) imposes no output-format constraints, benchmarking unguided, natural communication. Using this benchmark, we evaluate 32 leading open- and closed-source models in solo, homogeneous, and heterogeneous pairings. Our results reveal a surprising collaboration gap: models that perform well solo often degrade substantially when required to collaborate. We identify mitigations that show remarkable influence. For example, a small nudge via a relay inference approach, where a stronger agent leads before handing off to a weaker one, closes much of the gap. Our findings argue for (1) collaboration-aware evaluation, (2) training strategies to enhance collaborative capabilities, and (3) deliberate interaction design to elicit agents' collaboration skills, principles relevant to AI-AI and human-AI settings where agents must establish common ground.

cs.AI↗

arXiv2025

SynthWorlds: Controlled Parallel Worlds for Disentangling Reasoning and Knowledge in Language Models

Evaluating the reasoning ability of language models (LMs) is complicated by their extensive parametric world knowledge, where benchmark performance often reflects factual recall rather than genuine reasoning. Existing datasets and approaches (e.g., temporal filtering, paraphrasing, adversarial substitution) cannot cleanly separate the two. We present SynthWorlds, a framework that disentangles task reasoning complexity from factual knowledge. In SynthWorlds, we construct parallel corpora representing two worlds with identical interconnected structure: a real-mapped world, where models may exploit parametric knowledge, and a synthetic-mapped world, where such knowledge is meaningless. On top of these corpora, we design two mirrored tasks as case studies: multi-hop question answering and page navigation, which maintain equal reasoning difficulty across worlds. Experiments in parametric-only (e.g., closed-book QA) and knowledge-augmented (e.g., retrieval-augmented) LM settings reveal a persistent knowledge advantage gap, defined as the performance boost models gain from memorized parametric world knowledge. Knowledge acquisition and integration mechanisms reduce but do not eliminate this gap, highlighting opportunities for system improvements. Fully automatic and scalable, SynthWorlds provides a controlled environment for evaluating LMs in ways that were previously challenging, enabling precise and testable comparisons of reasoning and memorization.

cs.CL↗

arXiv2025

Tandem Training for Language Models

As language models continue to rapidly improve, we can expect their actions and reasoning to become difficult or impossible for weaker agents and humans to follow, undermining interpretability and oversight. With an eye on long-term futures, we pursue methods that encourage models to produce solutions that remain intelligible to weaker collaborators. We formalize intelligibility as handoff robustness: a strong model's solution is intelligible to a weaker model if randomly handing off control to the weaker model along the solution path does not cause failure. Building on this criterion, we introduce tandem training for language models, a reinforcement learning (RL) paradigm in which rollout tokens are intermittently and randomly sampled from a frozen weak model rather than the strong model being trained. Because rollouts succeed only when the strong model's actions and reasoning process can be continued by the weak model -- when the two can co-construct a successful solution -- optimizing standard RL objectives with tandem training implicitly incentivizes both correctness and intelligibility. In the GSM8K math reasoning task, tandem training reliably teaches models to abandon jargon and adapt their language to weaker partners while keeping task accuracy high. Our results demonstrate a promising route to building AI systems that remain auditable by weaker agents, with implications for human--AI collaboration and multi-agent communication.

cs.AI↗

arXiv2025

Narrow Finetuning Leaves Clearly Readable Traces in Activation Differences

Finetuning on narrow domains has become an essential tool to adapt Large Language Models (LLMs) to specific tasks and to create models with known unusual properties that are useful for research. We show that narrow finetuning creates strong biases in LLM activations that can be interpreted to understand the finetuning domain. These biases can be discovered using simple tools from model diffing - the study of differences between models before and after finetuning. In particular, analyzing activation differences on the first few tokens of random text and steering by adding this difference to the model activations produces text similar to the format and general content of the finetuning data. We demonstrate that these analyses contain crucial information by creating an LLM-based interpretability agent to understand the finetuning domain. With access to the bias, the agent performs significantly better compared to baseline agents using simple prompting. Our analysis spans synthetic document finetuning for false facts, emergent misalignment, subliminal learning, and taboo word guessing game models across different architectures (Gemma, LLaMA, Qwen) and scales (1B to 32B parameters). We suspect these biases reflect overfitting and find that mixing pretraining data into the finetuning corpus largely removes them, though residual risks may remain. Our work (1) demonstrates that narrowly finetuned models have salient traces of their training objective in their activations and suggests ways to improve how they are trained, (2) warns AI safety and interpretability researchers that the common practice of using such models as a proxy for studying broader finetuning (e.g., chat-tuning) might not be realistic, and (3) highlights the need for deeper investigation into the effects of narrow finetuning and development of truly realistic case studies for model-diffing, safety and interpretability research.

cs.CL↗

arXiv2025

Valid Survey Simulations with Limited Human Data: The Roles of Prompting, Fine-Tuning, and Rectification

Surveys provide valuable insights into public opinion and behavior, but their execution is costly and slow. Large language models (LLMs) have been proposed as a scalable, low-cost substitute for human respondents, but their outputs are often biased and yield invalid estimates. We study the interplay between synthesis methods that use LLMs to generate survey responses and rectification methods that debias population estimates, and explore how human responses are best allocated between them. Using two panel surveys with questions on nutrition, politics, and economics, we find that synthesis alone introduces substantial bias (24-86%), whereas combining it with rectification reduces bias below 5% and increases effective sample size by up to 14%. Overall, we challenge the common practice of using all human responses for fine-tuning, showing that under a fixed budget, allocating most to rectification results in far more effective estimation.

cs.CL↗

arXiv2025

GRAD: Generative Retrieval-Aligned Demonstration Sampler for Efficient Few-Shot Reasoning

Large Language Models (LLMs) achieve strong performance across diverse tasks, but their effectiveness often depends on the quality of the provided context. Retrieval-Augmented Generation (RAG) enriches prompts with external information, but its reliance on static databases constrains adaptability and can result in irrelevant demonstrations. In this work, we propose a Generative Retrieval-Aligned Demonstrator (GRAD), a dynamic demonstration-based approach where an LLM model is trained to generate input-specific concise demonstrations. By tailoring demonstrations to each input, our method offers better contextual support than traditional RAG approaches. We demonstrate the superiority of GRAD under budget constraints, where we limit both the number of tokens used per demonstration and the number of tokens used for the final output. Trained solely on a math dataset, GRAD consistently outperforms strong baselines on Qwen2.5-14B across mathematical reasoning and advanced STEM questions, highlighting GRAD's robust generalization to out-of-distribution (OOD) domains such as physics, chemistry, and computer science. Furthermore, we show that demonstrations generated by trained smaller models can effectively guide larger target models, reducing training costs while maintaining competitive accuracy. Overall, this work introduces a scalable demonstration generator model presenting the first step toward a dynamic few-shot learning paradigm in resource-constrained settings. We release the code used for the project.

cs.CL↗

arXiv2025

Interactive Evaluation of Large Language Models for Multi-Requirement Software Engineering Tasks

Standard single-turn, static benchmarks fall short in evaluating the nuanced capabilities of Large Language Models (LLMs) on complex tasks such as software engineering. In this work, we propose a novel interactive evaluation framework that assesses LLMs on multi-requirement programming tasks through structured, feedback-driven dialogue. Each task is modeled as a requirement dependency graph, and an ``interviewer'' LLM, aware of the ground-truth solution, provides minimal, targeted hints to an ``interviewee'' model to help correct errors and fulfill target constraints. This dynamic protocol enables fine-grained diagnostic insights into model behavior, uncovering strengths and systematic weaknesses that static benchmarks fail to measure. We build on DevAI, a benchmark of 55 curated programming tasks, by adding ground-truth solutions and evaluating the relevance and utility of interviewer hints through expert annotation. Our results highlight the importance of dynamic evaluation in advancing the development of collaborative code-generating agents.

cs.AI↗

arXiv2025

Ensemble learning of pathology foundation models for precision oncology

Histopathology is essential for cancer diagnosis and treatment selection, and pathology foundation models learn visual representations from whole-slide images (WSIs). However, existing foundation models are trained on disparate datasets using varying strategies, leading to inconsistent performance and limited generalizability. Here, we introduce ELF (Ensemble Learning of Foundation models), which integrates five pretrained pathology foundation models into unified slide-level representations. Trained on 53,699 WSIs spanning 20 anatomical sites, ELF leverages ensemble learning to capture complementary information across models. ELF's slide-level architecture is designed for data-efficient downstream evaluation, including settings with limited data such as therapeutic response prediction. We evaluate ELF for disease classification, biomarker detection, as well as anticancer and immunotherapy response prediction across multiple cancer types. ELF achieves higher performance than the evaluated constituent and slide-level foundation models across the tested tasks, supporting further evaluation of ensemble learning for pathology applications in oncology.

cs.CV↗