Searcharxiv⌕ Search
Home/Archive/arXiv

arXiv · 2610.11012

Distillation for Incrimination and Distillation for Capabilities

Sebastian Prasanna·Jacqueline Tay·Alek Westover
View original source ↗Read source document ↗Export metadata ↓

Abstract

Powerful misaligned AI models might recognize alignment evaluations and strategically behave well on them, rendering direct audits uninformative. However, distilling such a model into a weaker benign student places the teacher in a Distillation Double Bind: if misalignment transfers, the student may conceal it less effectively, revealing evidence about the teacher; if it does not, the student may learn useful capabilities while remaining benign. We introduce two distinct distillation approaches, one targeting each outcome. Distillation for Incrimination (DFI) aims to transfer misalignment but not the ability to conceal it. Distilling AuditBench's secret-keeping models into their underlying instruction-tuned model produces students that are significantly more likely than their teachers to admit their hidden behavior when asked, suggesting that knowledge of the behavior transferred more readily than the propensity to conceal it. Confession gains largely disappear when the student does not share the teacher's pretrained base, so DFI should target the teacher's own pre-RL checkpoint, which is weaker than the teacher but shares its base model. Distillation for Capabilities (DFC) aims to transfer capabilities but not misalignment. Among several techniques we evaluate, two are effective: inoculation prompting and training for more epochs on fewer unique examples. Both preserve the capability gains of standard distillation while substantially reducing the subliminal transfer of an animal preference, our proxy for misalignment. Together, these findings demonstrate two ways distillation can be used for AI safety: incriminating misaligned models, and extracting their capabilities without their misalignment.

Explore related subjects

cs.AI

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Sebastian Prasanna, Jacqueline Tay, Alek Westover. 2026-10-07. Distillation for Incrimination and Distillation for Capabilities. https://arxiv.org/abs/2610.11012

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Failure-Driven Workflow Refinement

Optimizing LLM-based workflows is typically formulated as a global search, where candidate workflows are evaluated based on a scalar metric. This paradigm, however, suffers from a critical flaw: information collapse. By reducing rich, multi-step execution traces to simple success/failure signals, existing methods are rendered blind to the underlying structure of failures, fundamentally preventing them from modeling the workflow's failure distribution. We reconceptualize this challenge as a distributional problem. We propose a new paradigm where the optimization goal is not to maximize a scalar score, but to directly minimize a workflow's Expected Failure Mass, i.e., the integral of its failure probability density function defined over a high-dimensional Failure Signature Space (FSS). This distributional lens allows us to move from inefficient, zero-order optimization to a principled, gradient-like descent on the failure landscape itself. We introduce CE-Graph, a framework that operationalizes this paradigm through a novel, failure-driven refinement process. CE-Graph approximates the failure distribution from a pool of counterexamples, identifies its densest regions as recurring failure modes, and applies targeted, operator-constrained graph edits via a Propose-and-Verify mechanism to greedily reduce the failure mass. On math, code, and QA benchmarks, our CE-Graph achieves higher robustness at a significantly lower cost than strong baselines. This suggests that a system's reliability emerges not from avoiding failures, but from systematically learning and reshaping the geometric structure of its failure distributions.

cs.AI↗
arXiv2026

Payoff scaling shapes cooperation in LLM agents across languages

Large language models (LLMs) are increasingly deployed as autonomous agents that negotiate, coordinate, and make decisions on behalf of human users. Understanding how they cooperate is therefore important for AI governance and machine behaviour research. We examine two factors: payoff magnitude and interaction language. Classical game theory predicts that multiplying all payoffs by a positive constant should not change behaviour because the underlying strategic incentives remain unchanged. We test this prediction using five widely adopted commercial LLMs playing an iterated prisoner's dilemma across ten payoff scales and five languages (English, French, Vietnamese, Chinese and Arabic). Contrary to theoretical expectations, changing only the numerical scale of payoffs substantially alters cooperation. Models respond differently, often in opposite directions, meaning that averages across models can conceal important behavioural heterogeneity. Sensitivity to payoff scale also depends strongly on prompt language. Moreover, payoff scale modifies the effects of persona instructions: at sub-unit stakes, where no payoff shown exceeds 1, all five models cooperate more under a selfish than a cooperative persona, whereas Claude and Gemini reverse this pattern at larger scales. These findings show that cooperation measures depend critically on payoff scale and language and should not be treated as fixed model traits.

cs.AI↗
arXiv2026

An Alternative Trajectory for Generative AI

The generative artificial intelligence (AI) ecosystem is undergoing rapid transformations that threaten its sustainability. As models transition from research prototypes to high-traffic products, the energetic burden has shifted from one-time training to recurring, unbounded inference. This is exacerbated by reasoning models that inflate compute costs by orders of magnitude per query. The prevailing pursuit of artificial general intelligence through scaling of monolithic models is colliding with hard physical constraints: grid failures, water consumption, and diminishing returns on data scaling. This trajectory yields models with impressive factual recall but struggles in domains requiring in-depth reasoning, possibly due to insufficient abstractions in training data. Current large language models (LLMs) exhibit genuine reasoning depth only in domains like mathematics and coding, where rigorous, pre-existing abstractions provide structural grounding. In other fields, the current approach fails to generalize well. We propose an alternative trajectory based on domain-specific superintelligence (DSS). We argue for first constructing explicit symbolic abstractions (knowledge graphs, ontologies, and formal logic) to underpin synthetic curricula enabling small language models to master domain-specific reasoning without the model collapse problem typical of LLM-based synthetic data methods. Rather than a single generalist giant model, we envision "societies of DSS models": dynamic ecosystems where orchestration agents route tasks to distinct DSS back-ends. This paradigm shift decouples capability from size, enabling intelligence to migrate from energy-intensive data centers to secure, on-device experts. By aligning algorithmic progress with physical constraints, DSS societies move generative AI from an environmental liability to a sustainable force for economic empowerment.

cs.AI↗