AtManRL: Towards Faithful Reasoning via Differentiable Attention Saliency

The core problem AtManRL addresses is subtler than it first appears: current chain-of-thought training can produce reasoning traces that look correct but are causally disconnected from what the model actually computed. The saliency reward is designed to penalize exactly that decorative reasoning, not just wrong reasoning.

This connects directly to a cluster of reliability concerns running through recent Modelwire coverage. The 'Diagnosing LLM Judge Reliability' piece from April 16 exposed that high aggregate consistency scores can mask widespread per-instance logical failures, which is structurally the same problem: surface-level coherence hiding internal incoherence. Meanwhile, IG-Search (also April 16) tackled a related flaw in RL-trained reasoning, finding that trajectory-level rewards cause gradient collapse and pushing toward step-level signals instead. AtManRL is working the same seam from a different angle, using attention saliency rather than information gain as the grounding signal. The AdaSplash-2 paper on differentiable sparse attention (April 16) is worth noting as adjacent infrastructure: making differentiable attention operations cheaper is a prerequisite for methods like AtManRL to scale without prohibitive training costs.

The meaningful test is whether AtManRL's faithfulness gains hold on tasks where the reasoning chain is long enough that attention diffusion becomes a real problem, such as multi-hop or mathematical competition benchmarks. If the saliency reward degrades performance on those relative to standard GRPO, the method may be trading accuracy for interpretability rather than genuinely capturing both.