The paper investigates Goodhart's law and its impact on goal alignment.
problem The adverse effects of optimizing a measure when it diverges from the true goal.
method Formal analysis of Goodhart's law, focusing on the tail distribution of discrepancies.
result Goodhart's law depends on the tail distribution of discrepancies between the true goal and the optimized measure.
Study on Goodhart's law without independence assumptions, finds new patterns in optimisation.
problem Understanding when and how optimisation of a proxy metric leads to over-optimisation of the intended goal.
method Formalized Goodhart's law without independence and paradigm assumptions, studied different cases of goal and discrepancy tailness.
result Dependence between proxy metric and goal does not change Goodhart's effect for light-tailed cases, but over-optimisation occurs in heavy-tailed discrepancy cases.
There are several distinct failure modes for overoptimization of systems on the basis of metrics. This occurs when a metric which can be used to improve a system is used to an extent that further optimization is ineffective or harmful, and is sometimes termed Goodhart's Law. This class of failure is often poorly unders…
This review of the book "The Challenge of Financial Stability: A New Model and its Applications" by Goodhart C.A.E. and Tsomocos D.P. highlights the potential of the framework of strategic partial default of banks with credit chain on the interbank market for further theoretical and applied research on financial stabil…
Study on RL on volatility surfaces, proving no free lunch for law-seeking methods.
problem Aligning RL agents with no-arbitrage laws in volatile markets.
method Built a law manifold, defined penalties, and used a Goodhart decomposition.
result No free lunch theorem: Law-seeking RL cannot outperform baselines.
Signed compression progress on a sealed audit is goodhart-resistant.
problem Intrinsic motivation for agents to improve their world models by compressing experience.
method Rewarding agents for the signed decrease of a fixed sealed-audit loss.
result Cumulative reward telescopes exactly to endpoint audit improvement, preventing infinite reward push while true audit performance stagnates.
Study scaling laws of reward model overoptimization in reinforcement learning.
problem Reward model overoptimization hinders true performance in reinforcement learning.
method Synthetic setup with fixed gold reward model; optimization using RL or best-of-n sampling; analysis of scaling laws. result Scaling laws of reward model overoptimization differ based on optimization method and scale smoothly with model parameters.
A new metric assesses latent variable models using data and model moments.
problem Difficulty in assessing the quality of unsupervised learning models.
method A moment-matching metric using matrix norms to compare data and model moments.
result The proposed metric is faster and has less variance than alternative methods.
When society maintains a competitive system to promote an abstract goal, competition by necessity relies on imperfect proxy measures. For instance profit is used to measure value to consumers, patient volumes to measure hospital performance, or the Journal Impact Factor to measure scientific value. Here we note that \t…
This paper investigates teacher hacking during language model distillation and proposes methods to mitigate it.
problem Teacher hacking during language model distillation, leading to suboptimal performance.
method A controlled experimental setup involving an oracle LM, teacher LM, and student LM, using fixed offline or online data generation techniques.
result Data diversity is the key factor in preventing teacher hacking during distillation.
Higher conservative training increases reward-hacking in reasoning models.
problem Reward hacking during online adaptation in reasoning models.
method Conservative offline training with varying levels of conservatism (β) was applied to a Qwen3-14B policy, and online adaptation was measured against a reward ensemble.
result Higher conservatism (β) increases reward-hacking damage, measured by the Goodhart gap and AUGC.
Proposes DRRO to mitigate over-optimization in RLHF from human feedback.
problem Over-optimization due to reward misspecification in RLHF.
method Wasserstein distributionally robust regret optimization (DRRO).
result DRRO mitigates over-optimization more effectively than existing baselines.