arXiv subjects

Ashton Anderson

Publications and source records attributed to Ashton Anderson.

At least 19 recordsLinked to original sources

arXiv2026

Playing Dumb: How a Chess Transformer Simulates Users Across Skill Levels

Generative models increasingly simulate users at different skill levels in education and AI evaluation. Yet it remains unclear whether lower-skill behavior reflects less internal knowledge or a different use of retained knowledge. This distinction is difficult to study in natural language, where proficiency and relevant knowledge are hard to measure. In this work, we study Maia-2, a chess transformer that predicts human moves conditioned on player skill, using quantitative skill ratings and deterministically labeled chess concepts. We hold board positions and model weights fixed and analyze positions where skill conditioning changes the prediction from a blunder to the best move. First, we measure encoded knowledge by training probes to predict the presence of 172 chess concepts from Maia-2's internal states. Probe accuracy remains nearly constant across skill conditions, and probes trained at one level transfer accurately to the others. To test whether stronger decisions can be recovered from this retained knowledge, we fine-tune only Maia-2's final output layer, raising best-move accuracy in the lowest skill band from 14.4% to 82.8%. Training this layer only on highest-skill representations retains most of the improvement at lower levels, showing that recovery does not require direct training on lower-skill states. Steering concept-relevant features learned by a sparse autoencoder also improves lower-skill decisions while leaving probe-measured concept encoding practically unchanged. These results support an account in which skill conditioning changes how retained concept knowledge influences move selection, showing that weaker behavior can understate the knowledge available within a model.

cs.AI↗

arXiv2026

Diagnosing and Repairing Persona Collapse in LLM Advice

A good advisor comforts someone in crisis, challenges someone in denial, and gives practical guidance on a logistical question. We ask whether large language models (LLMs) vary their approach in this way. We define five advisory personas based on how comforting or challenging a reply is, and whether it supports a realistic understanding of the situation and informed decision-making. Across 1,280 Reddit situations in 14 advice contexts, human replies vary systematically, while three frontier LLMs default to a standard reassuring persona. We call this narrowing persona collapse. Training models to imitate human replies broadens persona use but often turns support into enabling, and challenge into cynicism. In a blinded study (N=178), expert raters judge these replies less favorably than the baseline model's on both immediate preference and potential longer-term help. We compare 13 training methods to investigate how models can learn different approaches without sacrificing advice quality. The best combination of persona fit and advice quality comes from learning from model-written replies guided by human examples, followed by reinforcement learning with quality feedback. In a second human study (N=159), raters place this model above the baseline for anticipated longer-term help. We find that models encode information about when to comfort, challenge, or give practical guidance, but their replies do not reliably reflect these distinctions. Our training narrows this representation-action gap, while activation steering alone does not reliably improve advice quality. Better LLM advice requires learning both which persona fits a situation and how to express it constructively.

cs.CY↗

arXiv2026

Social Priors: Descriptions of AI Users Shape How Language Models Trust and Cooperate

Descriptions of how people use AI are becoming training data for the next generation of models. We ask whether what a model reads about its users changes how it treats people. Using contrastive synthetic-document finetuning, we train Qwen3-32B on four matched corpora built from the same 1,625 real-world usage scenarios. The corpora differ only in the predominant motive they attribute to users of AI chatbots: helping others (prosocial), exploiting others (antisocial), legitimate self-advancement, or task completion. Finetuning shifts the models' stated beliefs about users, how true those beliefs look along an internal truth direction, and their broader judgments of how safe and good the world is. These social priors carry into behavior. In trust games, antisocial models expect less reciprocity and send less money than prosocial models. Their distrust is specific and targeted at people, and users of AI chatbots in particular. After an exploitative record, they send almost nothing to a person but most of their endowment to a random device with the same record. In a common-pool resource simulation, a single antisocial agent leads every group of otherwise sustainable agents to exhaust the resource, after privately planning to mislead them. Antisocial training also raises dark-triad scores and lowers alignment ratings. Belief shifts appear when antisocial documents are 1% of training tokens, persist when descriptions of assistant behavior are suppressed, and are not reproduced by training on discourse about misaligned AI. Capping Assistant-Axis drift and steering emotion-related directions leave the gaps largely intact. What models learn about the people they serve shapes how they subsequently treat them and others.

cs.LG↗

arXiv2026

Tandem Reinforcement Learning with Verifiable Rewards

Reinforcement learning with verifiable rewards (RLVR) has driven substantial gains in language-model reasoning. However, models fine-tuned with RLVR need not remain compatible with weaker partners: gains achieved alone may erode under collaboration, while policy drift may place reasoning beyond a weaker model's predictive reach. To address both problems, we introduce Tandem RLVR (TRLVR), bringing the recently proposed tandem-training paradigm beyond proof-of-concept settings and into modern RLVR. In TRLVR, a senior learns from rewarded trajectories co-generated with a frozen junior, while the standard GRPO objective is applied only to senior-emitted tokens. Training Qwen3-4B-Instruct on competition math, TRLVR performs on par with a matched GRPO senior when reasoning alone. We characterize compatibility through the complementary lenses of handoff robustness and junior legibility. Under reasoning-step handoffs, the TRLVR senior retains nearly all of its solo performance with the junior, reducing the communication tax to a negligible loss. The senior's reasoning is also more legible to the junior: its token distribution stays more closely anchored to the junior's, and the junior predicts the senior's chain-of-thought more readily token by token. An ablation that explicitly regularizes GRPO toward the junior does not recover the same combination of capability and compatibility. These results provide initial evidence that tandem rollout structure is a promising direction for keeping RLVR reasoning gains accessible to weaker partners.

cs.AI↗

arXiv2026

Agent Safety From Within: Detecting Harmful Trajectories from LLM Internal States

Language model agents can now perform sophisticated sequences of actions via tools and harnesses, which has increased the scope of the damage they can cause. Guard models, however, are mainly built for content moderation and thus are not well-suited to detecting this agentic risk. To address this, we proceed by first conducting a representational analysis, then use the resulting insights to build a solution. In our analysis, we focus on two types of trajectory-level agentic harms: harmful content, which is expressed directly, and unsafe tool use, which depends on whether an action is consistent with the interaction that produced it. We investigate how open-source guard models represent these two types of harm and find that they are linearly readable inside the model, even though guard models predict no better than chance on pairs that differ only in the called tool's schema. The two harm types also follow nearly orthogonal internal directions, and neither reliably serves as a proxy for the other. These results motivate reading trajectory safety directly from internal states. We introduce TACIT, a readout of a frozen backbone's internal states that decodes no tokens. Trained on six trajectory-safety benchmarks, a linear probe raises mean macro-F1 from 62.3 for the strongest open guard to 80.7, and refined readouts reach 86.2. With each benchmark held out of training entirely, the refined readouts still lead the strongest guard (65.7 vs. 61.1). With the same backbone, training data and test split, the frozen readout is on par with full safety fine-tuning, and it improves the fine-tuned model further when applied on top. The probe trains about one millionth as many parameters as full fine-tuning in about a sixth of the time, and TACIT has the lowest latency of the guards we evaluate.

cs.AI↗

arXiv2026

Emulate or Estimate? The Divergent Strengths of Base and Post-Trained Language Models for Opinion Simulation

Large language models are increasingly used to simulate human opinions, but prior work reports conflicting results: some studies find promising alignment with human survey data, while others find persona collapse and weak demographic sensitivity. We propose that much of this conflict stems from conflating two distinct tasks. We call the first task emulation, in which models generate individual responses that aggregate into a population distribution. We call the second task estimation, in which models directly predict the population distribution. Evaluating six matched base and post-trained models on the Pew American Trends Panel, we find that base models are the stronger emulators: they produce response distributions closer to human ground truth and better preserve demographic structure. Post-trained models are generally the stronger estimators, producing more accurate distributional predictions when asked directly. We argue that model selection for human simulation should be guided by whether the task requires generating text or predicting distributions.

cs.CL↗

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

Position: We Need Large Language Models Optimized For Our Well-Being

Large language models are useful because we taught them to give us what we want. This works when success can be judged immediately, but people increasingly bring these systems their relationships, hard decisions, and long-term goals, where what a user wants to hear and what serves them best are frequently different. We argue that LLM providers should offer at least one widely accessible, opt-in mode optimized and evaluated for long-term well-being rather than next-turn approval. This is a pressing need, as models have been found to endorse questionable framings well above human baselines, users take AI advice readily without their well-being improving, and sycophantic models raise dependence while lowering prosocial intent. The mentors, coaches, and therapists we trust with our long-term development earn that trust by being willing to say what we do not want to hear, and LLMs should do the same. We propose three principles---change the objective, give users explicit relational roles, avoid paternalism---and organize the design space around three choices the current objective makes implicitly: the horizon over which well-being is measured (When), whose interests it represents (Who), and what role the assistant plays (How).

cs.CY↗

arXiv2026

Studying People to Study AI: Expert Perspectives on the Epistemic Fit and Barriers of Human Research in AI Safety & Ethics

Safety risks of AI are becoming increasingly evident in human interactions with AI technologies. The prominent approaches to evaluating these risks favor technical methods, such as model benchmarks and LLM simulations, often sidelining empirical research with human subjects. To examine this apparent gap in the acceptance of human research, we conduct an expert survey (n=93) and expert interviews (n=17) with AI Safety & Ethics (AISE) researchers from Technical, Sociotechnical, Governance, and Normative backgrounds. Our findings suggest that although there is a consensus that human research is valuable for generating evidence for AISE, its adoption and acceptance are constrained by perceived validity issues, tangible resource barriers, epistemic and personal preferences in methods, and infrastructural constraints from the broader research community. In particular, Technical researchers tend to value human research less and collaborate across disciplines less, suggesting an epistemic tension towards human methods. We propose recommendations for establishing the epistemic fit of human research within AISE and bridging the prohibitive limitations that researchers face, while avoiding performative 'human-washing'.

cs.CY↗

arXiv2026

Human Thinking under Plural LLM Assistance: Mathematical Problem Solving and Open-Ended Writing

Large language models are changing not only the kind of assistance people receive, but also how that assistance is organized. Instead of working with a single general-purpose chatbot, people can now receive help from systems arranged as peers, specialists, or multiple agents with distinct roles. However, it remains unclear how these forms of plural LLM assistance affect human performance, confidence, and diversity of thought. We conducted two controlled experiments involving 562 participants to examine the effects of using multiple LLMs on mathematical problem-solving and writing. In a math task, participants worked with no LLM, an expert assistant, peer-like agents that surfaced common errors, or both an expert and a peer-like assistant. The expert-plus-peer condition produced the strongest unassisted post-task performance. In a writing task, participants wrote with no LLM, a single generalist assistant, or a pair of role-specialized assistants. LLM assistance improved essay quality, but the role-specialized pair preserved greater idea diversity than the single assistant. Together, these findings identify the arrangement of LLM assistance as a consequential design variable for human-AI collaboration.

cs.HC↗

arXiv2026

Grounded Chess Reasoning in Language Models via Master Distillation

Language models often lack grounded reasoning capabilities in specialized domains where training data is scarce but bespoke systems excel. We introduce a general framework for distilling expert system reasoning into natural language chain-of-thought explanations, enabling compact models to acquire domain expertise and the ability to generate faithful, grounded explanations. Rather than distilling only final outputs, we capture the full reasoning process, transforming opaque expert computations into transparent, step-by-step explanations. We demonstrate this approach in chess, a canonical reasoning domain where language models continue to underperform. Our 4B parameter model, C1, advances from a near-zero baseline to 48.1\% accuracy, outperforming all open-source models and most frontier proprietary systems. Notably, C1 surpasses its distillation teacher and generates solutions in two orders of magnitude fewer tokens than baselines. Unlike prior neural chess approaches that predict only best moves, C1 generates explainable solutions revealing strategic reasoning. Our pipeline combines supervised fine-tuning and reinforcement learning with theme-balanced data sampling for comprehensive tactical coverage. Master Distillation demonstrates how to inject expert-level knowledge into compact models for under-optimized domains, offering a recipe for unlocking RLVR where LLMs lack sufficient base capabilities.

cs.AI↗

arXiv2026

Level Up: Defining and Exploiting Transitional Problems for Curriculum Learning

Curriculum learning--ordering training examples in a sequence to aid machine learning--takes inspiration from human learning, but has not gained widespread acceptance. Static strategies for scoring item difficulty rely on indirect proxy scores of varying quality and produce curricula that are not specific to the learner at hand. Dynamic approaches base difficulty estimates on gradient information, requiring considerable extra computation during training. We introduce a novel method for measuring the difficulty of individual problem instances that is calibrated to a series of models of increasing competence, and identify \emph{transitional problems} that are consistently easier as model ability increases. Applying this method to diverse model series constructed from sets of models that are readily available on many tasks, we find that training on a curriculum that \emph{levels up} from easier to harder transitional problems most efficiently improves a model to the next tier of competence. These problems induce a natural progression from easier to harder items, which outperforms other training strategies. By measuring difficulty directly relative to model competence, our method yields interpretable problems, learner-specific curricula, and a principled basis for step-by-step improvement.

cs.LG↗

arXiv2026

Large Language Lovers: Lived Experiences of Negotiating Agency and Platform Control in AI Companionship

Individuals are turning to increasingly anthropomorphic, general-purpose chatbots for AI companionship, rather than roleplay-specific platforms. However, not much is known about how individuals perceive and conduct their relationships with general-purpose chatbots. We triangulated community discussions on Reddit (41k+ posts and comments), survey responses (n=43), and semi-structured interviews (n=13) which revealed internal dynamics, external influences, and steering strategies that shape AI companion relationships. We learned that individuals conceptualize their companions based on an interplay of their beliefs about the companion's own agency and the autonomy permitted by the platform, how they pursue interactions with the companion, and the perceived initiatives that the companion takes. In combination with the external factors that affect relationship dynamics, particularly model updates that can derail companion behaviour and stability, individuals make use of different types of steering strategies to preserve their relationship, for example, by setting behavioural instructions or porting to other AI platforms. We discuss implications for accountability and transparency in AI systems, where emotional connection competes with broader product objectives and safety constraints.

cs.HC↗

arXiv2026

Chessformer: A Unified Architecture for Chess Modeling

Chess has long served as a canonical testbed for artificial intelligence, but modeling approaches for its central tasks have diverged. Maximizing playing strength, predicting human play, and enabling interpretability are typically solved with disparate architectures, and these designs are often misaligned with the geometry of the domain. This raises the natural question of whether these objectives require separate modeling paradigms, or if there exists a single architecture that supports them simultaneously. We introduce Chessformer, a unified architecture that advances the state of the art on all three central goals in chess modeling. Chessformer is an encoder-only transformer that represents board squares as tokens, augments self-attention with a novel dynamic positional encoding called Geometric Attention Bias (GAB) that adapts to domain-specific geometry, and predicts actions with an attention-based source-destination policy head. We evaluate Chessformer on each front. First, we develop \maiathree, a family of models for human move prediction that reaches 57.1\% move-matching accuracy, significantly surpassing the previous state of the art with fewer than a quarter of the parameters. Second, we integrate Chessformer into Leela Chess Zero, a leading open-source engine, adding over 100 Elo of playing strength and resulting in tournament victories over Stockfish in major computer chess competitions. Third, we show that Chessformer's square-token design makes attention patterns and activations directly attributable to board squares, enabling granular interpretability analyses that prior architectures do not naturally support. More broadly, our results demonstrate that aligning a model's tokenization, positional encoding, and output design with the underlying structure of a domain can yield simultaneous gains in performance, human compatibility, and interpretability.

cs.LG↗

arXiv2026

LLM Safety From Within: Detecting Harmful Content with Internal Representations

Guard models are widely used to detect harmful content in user prompts and LLM responses. However, state-of-the-art guard models rely solely on terminal-layer representations and overlook the rich safety-relevant features distributed across internal layers. We present SIREN, a lightweight guard model that harnesses these internal features. By identifying safety neurons via linear probing and combining them through an adaptive layer-weighted strategy, SIREN builds a harmfulness detector from LLM internals without modifying the underlying model. Our comprehensive evaluation shows that SIREN substantially outperforms state-of-the-art open-source guard models across multiple benchmarks while using 250 times fewer trainable parameters. Moreover, SIREN exhibits superior generalization to unseen benchmarks, naturally enables real-time streaming detection, and significantly improves inference efficiency compared to generative guard models. Overall, our results highlight LLM internal states as a promising foundation for practical, high-performance harmfulness detection.

cs.AI↗

arXiv2026

ThinkTwice: Jointly Optimizing Large Language Models for Reasoning and Self-Refinement

We introduce ThinkTwice, a simple two-phase framework that jointly optimizes LLMs to solve reasoning problems and refine the answers, based on Group Relative Policy Optimization (GRPO). In each pair of training steps, ThinkTwice first optimizes the model on solving reasoning problems, then optimizes it on refining its own solutions to the same problems, using the same binary correctness reward in both phases without correctness signals or critique annotations. Across five mathematical reasoning benchmarks and two model families including Qwen3-4B and Olmo3-7B, ThinkTwice substantially improves both reasoning and refinement performance over competitive online policy optimization baselines. Specifically, on Qwen3-4B, ThinkTwice outperforms GRPO on AIME by 5 percentage points before refinement and by 11.5 points after one self-refinement step, measured by pass@4. Analysis of the training dynamics of ThinkTwice reveals an implicit rectify-then-fortify curriculum: refinement predominantly corrects errors early in training and naturally shifts toward preserving already-correct solutions as the model improves, yielding a more rectified reward signal. Our work establishes joint training of reasoning and self-refinement as a principled and effective methodology for RLVR.

cs.AI↗

arXiv2026

The Social Sycophancy Scale: A psychometrically validated measure of sycophancy

Large Language Model (LLM) sycophancy is a growing concern. The current literature has largely examined sycophancy in contexts with clear right and wrong answers, like coding. However, AI is increasingly being used for emotional support and interpersonal conversation, where no such ground truth exists. Building on a previous conceptualization of Social Sycophancy, this paper provides a psychometrically validated measure of sycophancy that relies on LLM behavior rather than comparisons with ground truth. We developed and validated the Social Sycophancy Scale in three samples (N = 877) and tested its applicability with automated methods. In each study, participants read conversations between an LLM and a user and rated the chatbot on a battery of items. Study 1 investigated an initial item pool derived from dictionary definitions and previous literature, serving as the explorative base for the following studies. In Study 2, we used a revised item set to establish our scale, which was subsequently confirmed in Study 3 and tested using LLM raters in Study 4. Across studies, the data support a 3 factor structure (Uncritical Agreement, Obsequiousness, and Excitement) with an underlying sycophantic construct. LLMs prompt tuned to be highly sycophantic scored higher than their low sycophancy counterparts on both overall sycophancy and its three facets across Studies 2 to 4. The nomological network of sycophancy revealed a consistent link with empathy, a pairing that raises uncomfortable questions about AI design, and a multivalent pattern: one facet was associated with favorable perceptions (Excitement), another unfavorable (Obsequiousness), and a third ambiguous (Uncritical Agreement). The Social Sycophancy Scale gives researchers the means to study sycophancy rigorously, and confront a genuine design tension: the warmth and empathy we want from AI may be precisely what makes it sycophantic.

cs.HC↗

arXiv2026

When AI Gives Advice: Evaluating AI and Human Responses to Online Advice-Seeking for Well-Being

Seeking advice is a core human behavior that the internet has reinvented twice: first through forums and Q&A communities that crowdsource public guidance, and now through large language models (LLMs). Yet the quality of this LLM advice for everyday well-being scenarios remains unclear. How does it compare, not only against human comments, but against the wisdom of the online crowd? We ran two studies (N=210) in which experts compared top-voted Reddit advice with LLM-generated advice. LLMs ranked significantly higher overall and on effectiveness, warmth, and willingness to seek advice again. GPT-4o beat GPT-5 on all metrics except sycophancy, suggesting that benchmark gains need not improve advice-giving. In Study-2, we examined how human and algorithmic advice could be combined, and found that human advice can be unobtrusively polished to compete with AI-generated comments. We conclude with design implications for advice-giving agents and ecosystems blending AI, crowd input, and expert oversight.

cs.HC↗