arXiv subjects

Noah Y. Siegel

Publications and source records attributed to Noah Y. Siegel.

12 recordsLinked to original sources

arXiv2026

A Positive Case for Faithfulness: LLM Self-Explanations Help Predict Model Behavior

LLM self-explanations are often presented as a promising tool for AI oversight, yet their faithfulness to the model's true reasoning process is poorly understood. Existing faithfulness metrics have critical limitations, typically relying on identifying unfaithfulness via adversarial prompting or detecting reasoning errors. These methods overlook the predictive value of explanations. We introduce Normalized Simulatability Gain (NSG), a general and scalable metric based on the idea that a faithful explanation should allow an observer to learn a model's decision-making criteria, and thus better predict its behavior on related inputs. We evaluate 18 frontier proprietary and open-weight models, e.g., Gemini 3, GPT-5.2, and Claude 4.5, on 7,000 counterfactuals from popular datasets covering health, business, and ethics. We find self-explanations substantially improve prediction of model behavior (11-37% NSG). Self-explanations also provide more predictive information than explanations generated by external models, even when those models are stronger. This implies an advantage from self-knowledge that external explanation methods cannot replicate. Our approach also reveals that, across models, 5-15% of self-explanations are highly misleading. Despite their imperfections, we show a positive case for self-explanations: they encode information that helps predict model behavior.

cs.AI↗

arXiv2026

Debate Training Reduces Reward Hacking in RLAIF

We demonstrate that RL finetuning an LLM using debate, a two-player adversarial game between a generator and a critic adjudicated by a weaker LLM judge, reduces reward hacking compared to a reinforcement learning from AI feedback (RLAIF) baseline. Reward hacking is a central obstacle in RLAIF: as training progresses, the policy learns to exploit systematic errors in its AI judge, degrading task performance, a problem that worsens precisely when the judge is weaker than the policy, the setting most relevant to overseeing increasingly capable AI systems. We study mathematics tasks, where final-answer correctness is verifiable, allowing us to measure reward hacking dynamics. We train a Gemini~2.5 Flash-class policy with a frozen, weaker Gemini~2.5 Flash Lite judge, comparing a single-player RLAIF baseline against debate. While the baseline quickly hacks the judge, debate maintains judge performance throughout training, leading to a higher peak validation accuracy (45\% performance gap recovered) that persists through many RL steps. Additional experiments show that: 1) further weakening the judge leads to faster hacking, but this can be compensated by adding an additional debate round; 2) debate incentives override prompted misalignment; 3) RL using an LLM judge has a smaller train/validation reward gap than RL from verifiable rewards; 4) learning to critique to convince the judge using ground truth labels is possible but slow. Taken together, our results are a positive update on the feasibility of debate, while highlighting that balancing multi-agent training is critical: without player constraints, adversarial training risks defaulting to critic judge-hacking. We show that critique word limits (effective up to 150 words) successfully balance the game and avoid judge hacking, though this introduces a trade-off by restricting critic expressive clarity.

cs.LG↗

arXiv2026

Training Large Language Models for Self-Explanation Faithfulness

We propose a Reinforcement Learning (RL) method to directly optimize the faithfulness of self-explanations - the extent to which a model's generated reasoning accurately reflects its internal decision-making process. While existing work focuses on evaluating faithfulness or using inference-time prompting frameworks to improve an LLM's self-explanation's tractability, these approaches do not provide a mechanism to directly optimize a model's parameters to generate faithful self-explanations. We bridge this gap by modifying existing faithfulness metrics into an RL training objective. We investigate (1) if models can be trained to accurately detect factors that affect their decisions, and (2) whether RL can directly optimize for the disclosure of these factors thereby improving LLM self-explanations' faithfulness. We experiment with two intervention types: random-word insertions and user-bias insertions, using a per-sample reward derived from the Phi-CCT correlation metric. RL fine-tuned Llama3.1-8B and Qwen3-8B show substantial improvements on the Phi-CCT faithfulness metric, with in-distribution scores rising from near-zero to as high as 0.664, and out-of-distribution scores reaching up to 0.691 on held-out tasks such as StrategyQA. Cross-intervention generalization is weaker but more interesting: a priori we would not expect a model trained only on random word insertions to generalize to user-bias phrases, yet Llama3.1-8B shows non-zero transfer in this direction. The reverse direction and Qwen3-8B do not replicate this, indicating model-dependent and setup-dependent effects we cannot yet explain. Lastly we analyze model behavior to rule out reward gaming behaviors that often plague RL training. Ultimately, we show that models can be trained to implicitly identify influential factors and disclose them, offering a scalable path toward reducing unfaithful reasoning in LLMs.

cs.LG↗

arXiv2026

Verbosity Tradeoffs and the Impact of Scale on the Faithfulness of LLM Self-Explanations

When asked to explain their decisions, LLMs can often give explanations which sound plausible to humans. But are these explanations faithful, i.e. do they convey the factors actually responsible for the decision? In this work, we analyse counterfactual faithfulness across 75 models from 13 families. We analyze the tradeoff between conciseness and comprehensiveness, how correlational faithfulness metrics assess this tradeoff, and the extent to which metrics can be gamed. This analysis motivates two new metrics: the phi-CCT, a simplified variant of the Correlational Counterfactual Test (CCT) which avoids the need for token probabilities while explaining most of the variance of the original test; and F-AUROC, which eliminates sensitivity to imbalanced intervention distributions and captures a model's ability to produce explanations with different levels of detail. Our findings reveal a clear scaling trend: larger and more capable models are consistently more faithful on all metrics we consider. Our code is available at https://github.com/google-deepmind/corr_faith.

cs.CL↗

arXiv2025

LLMs Can Covertly Sandbag on Capability Evaluations Against Chain-of-Thought Monitoring

Trustworthy evaluations of dangerous capabilities are increasingly crucial for determining whether an AI system is safe to deploy. One empirically demonstrated threat is sandbagging - the strategic underperformance on evaluations by AI models or their developers. A promising defense is to monitor a model's chain-of-thought (CoT) reasoning, as this could reveal its intentions and plans. In this work, we measure the ability of models to sandbag on dangerous capability evaluations against a CoT monitor by prompting them to sandbag while being either monitor-oblivious or monitor-aware. We show that both frontier models and small open-sourced models can covertly sandbag against CoT monitoring 0-shot without hints. However, they cannot yet do so reliably: they bypass the monitor 16-36% of the time when monitor-aware, conditioned on sandbagging successfully. We qualitatively analyzed the uncaught CoTs to understand why the monitor failed. We reveal a rich attack surface for CoT monitoring and contribute five covert sandbagging policies generated by models. These results inform potential failure modes of CoT monitoring and may help build more diverse sandbagging model organisms.

cs.CR↗

arXiv2025

Advancing Event Forecasting through Massive Training of Large Language Models: Challenges, Solutions, and Broader Impacts

Many recent papers have studied the development of superforecaster-level event forecasting LLMs. While methodological problems with early studies cast doubt on the use of LLMs for event forecasting, recent studies with improved evaluation methods have shown that state-of-the-art LLMs are gradually reaching superforecaster-level performance, and reinforcement learning has also been reported to improve future forecasting. Additionally, the unprecedented success of recent reasoning models and Deep Research-style models suggests that technology capable of greatly improving forecasting performance has been developed. Therefore, based on these positive recent trends, we argue that the time is ripe for research on large-scale training of superforecaster-level event forecasting LLMs. We discuss two key research directions: training methods and data acquisition. For training, we first introduce three difficulties of LLM-based event forecasting training: noisiness-sparsity, knowledge cut-off, and simple reward structure problems. Then, we present related ideas to mitigate these problems: hypothetical event Bayesian networks, utilizing poorly-recalled and counterfactual events, and auxiliary reward signals. For data, we propose aggressive use of market, public, and crawling datasets to enable large-scale training and evaluation. Finally, we explain how these technical advances could enable AI to provide predictive intelligence to society in broader areas. This position paper presents promising specific paths and considerations for getting closer to superforecaster-level AI technology, aiming to call for researchers' interest in these directions.

cs.LG↗

arXiv2024

On scalable oversight with weak LLMs judging strong LLMs

Scalable oversight protocols aim to enable humans to accurately supervise superhuman AI. In this paper we study debate, where two AI's compete to convince a judge; consultancy, where a single AI tries to convince a judge that asks questions; and compare to a baseline of direct question-answering, where the judge just answers outright without the AI. We use large language models (LLMs) as both AI agents and as stand-ins for human judges, taking the judge models to be weaker than agent models. We benchmark on a diverse range of asymmetries between judges and agents, extending previous work on a single extractive QA task with information asymmetry, to also include mathematics, coding, logic and multimodal reasoning asymmetries. We find that debate outperforms consultancy across all tasks when the consultant is randomly assigned to argue for the correct/incorrect answer. Comparing debate to direct question answering, the results depend on the type of task: in extractive QA tasks with information asymmetry debate outperforms direct question answering, but in other tasks without information asymmetry the results are mixed. Previous work assigned debaters/consultants an answer to argue for. When we allow them to instead choose which answer to argue for, we find judges are less frequently convinced by the wrong answer in debate than in consultancy. Further, we find that stronger debater models increase judge accuracy, though more modestly than in previous studies.

cs.LG↗

arXiv2024

The Probabilities Also Matter: A More Faithful Metric for Faithfulness of Free-Text Explanations in Large Language Models

In order to oversee advanced AI systems, it is important to understand their underlying decision-making process. When prompted, large language models (LLMs) can provide natural language explanations or reasoning traces that sound plausible and receive high ratings from human annotators. However, it is unclear to what extent these explanations are faithful, i.e., truly capture the factors responsible for the model's predictions. In this work, we introduce Correlational Explanatory Faithfulness (CEF), a metric that can be used in faithfulness tests based on input interventions. Previous metrics used in such tests take into account only binary changes in the predictions. Our metric accounts for the total shift in the model's predicted label distribution, more accurately reflecting the explanations' faithfulness. We then introduce the Correlational Counterfactual Test (CCT) by instantiating CEF on the Counterfactual Test (CT) from Atanasova et al. (2023). We evaluate the faithfulness of free-text explanations generated by few-shot-prompted LLMs from the Llama2 family on three NLP tasks. We find that our metric measures aspects of faithfulness which the CT misses.

cs.CL↗

arXiv2024

The Effect of Model Size on LLM Post-hoc Explainability via LIME

Large language models (LLMs) are becoming bigger to boost performance. However, little is known about how explainability is affected by this trend. This work explores LIME explanations for DeBERTaV3 models of four different sizes on natural language inference (NLI) and zero-shot classification (ZSC) tasks. We evaluate the explanations based on their faithfulness to the models' internal decision processes and their plausibility, i.e. their agreement with human explanations. The key finding is that increased model size does not correlate with plausibility despite improved model performance, suggesting a misalignment between the LIME explanations and the models' internal processes as model size increases. Our results further suggest limitations regarding faithfulness metrics in NLI contexts.

cs.CL↗

arXiv2024

Learning Agile Soccer Skills for a Bipedal Robot with Deep Reinforcement Learning

We investigate whether Deep Reinforcement Learning (Deep RL) is able to synthesize sophisticated and safe movement skills for a low-cost, miniature humanoid robot that can be composed into complex behavioral strategies in dynamic environments. We used Deep RL to train a humanoid robot with 20 actuated joints to play a simplified one-versus-one (1v1) soccer game. The resulting agent exhibits robust and dynamic movement skills such as rapid fall recovery, walking, turning, kicking and more; and it transitions between them in a smooth, stable, and efficient manner. The agent's locomotion and tactical behavior adapts to specific game contexts in a way that would be impractical to manually design. The agent also developed a basic strategic understanding of the game, and learned, for instance, to anticipate ball movements and to block opponent shots. Our agent was trained in simulation and transferred to real robots zero-shot. We found that a combination of sufficiently high-frequency control, targeted dynamics randomization, and perturbations during training in simulation enabled good-quality transfer. Although the robots are inherently fragile, basic regularization of the behavior during training led the robots to learn safe and effective movements while still performing in a dynamic and agile way -- well beyond what is intuitively expected from the robot. Indeed, in experiments, they walked 181% faster, turned 302% faster, took 63% less time to get up, and kicked a ball 34% faster than a scripted baseline, while efficiently combining the skills to achieve the longer term objectives.

cs.RO↗

arXiv2021

From Motor Control to Team Play in Simulated Humanoid Football

Intelligent behaviour in the physical world exhibits structure at multiple spatial and temporal scales. Although movements are ultimately executed at the level of instantaneous muscle tensions or joint torques, they must be selected to serve goals defined on much longer timescales, and in terms of relations that extend far beyond the body itself, ultimately involving coordination with other agents. Recent research in artificial intelligence has shown the promise of learning-based approaches to the respective problems of complex movement, longer-term planning and multi-agent coordination. However, there is limited research aimed at their integration. We study this problem by training teams of physically simulated humanoid avatars to play football in a realistic virtual environment. We develop a method that combines imitation learning, single- and multi-agent reinforcement learning and population-based training, and makes use of transferable representations of behaviour for decision making at different levels of abstraction. In a sequence of stages, players first learn to control a fully articulated body to perform realistic, human-like movements such as running and turning; they then acquire mid-level football skills such as dribbling and shooting; finally, they develop awareness of others and play as a team, bridging the gap between low-level motor control at a timescale of milliseconds, and coordinated goal-directed behaviour as a team at the timescale of tens of seconds. We investigate the emergence of behaviours at different levels of abstraction, as well as the representations that underlie these behaviours using several analysis techniques, including statistics from real-world sports analytics. Our work constitutes a complete demonstration of integrated decision-making at multiple scales in a physically embodied multi-agent setting. See project video at https://youtu.be/KHMwq9pv7mg.

cs.AI↗

arXiv2020

Keep Doing What Worked: Behavioral Modelling Priors for Offline Reinforcement Learning

Off-policy reinforcement learning algorithms promise to be applicable in settings where only a fixed data-set (batch) of environment interactions is available and no new experience can be acquired. This property makes these algorithms appealing for real world problems such as robot control. In practice, however, standard off-policy algorithms fail in the batch setting for continuous control. In this paper, we propose a simple solution to this problem. It admits the use of data generated by arbitrary behavior policies and uses a learned prior -- the advantage-weighted behavior model (ABM) -- to bias the RL policy towards actions that have previously been executed and are likely to be successful on the new task. Our method can be seen as an extension of recent work on batch-RL that enables stable learning from conflicting data-sources. We find improvements on competitive baselines in a variety of RL tasks -- including standard continuous control benchmarks and multi-task learning for simulated and real-world robots.

cs.LG↗