AI summaryⓘ
The authors study a problem where training reasoning models gets stuck when the chance of initial success is very low. They create a new loss function family that balances between pure reinforcement learning and likelihood-based methods, helping the model escape this 'cold start' problem faster. They introduce two practical algorithms to approximate this loss, called GARL and PAFT, which trade off variance and bias in different ways. Their tests on reasoning benchmarks show that GARL helps the model start learning better in hard cases, while PAFT provides more stable training in some scenarios. Overall, their work addresses training stalls by cleverly adjusting how the learning signal is amplified during fine-tuning.
reinforcement learningTsallis q-logarithmloss functioncold startgradient amplificationMonte Carlo estimationposterior resamplingreasoning modelsfine-tuninglatent trajectories
Abstract
Adapting reasoning models to new tasks during post-training with only output-level supervision stalls under reinforcement learning from verifiable rewards (RLVR) when the initial success probability $p_0$ is small. Using the Tsallis $q$-logarithm, we define a loss family $J_Q$ that interpolates between RLVR (at $q{=}0$, the exploitation pole) and the log-marginal-likelihood over latent trajectories (at $q{=}1$, the density-estimation pole). All members share the same per-example gradient direction, differing only by a scalar amplification $P_{θ^{-q}}$ that reweights each instance independently of the learning rate. This amplification is the mechanism that addresses cold-start stalling: under gradient flow, the exploitation pole requires $Ω(\frac{1}{p_0})$ time to escape cold start, while the density-estimation pole escapes in $Θ\big(\log(\frac{1}{p_0})\big)$; intermediate $q$ trades escape speed against noise memorization. Because $P_θ$ is intractable, we derive two Monte Carlo estimators from the two factorizations of the gradient: Gradient-Amplified RL (GARL) samples from the prior and amplifies the RL gradient, and Posterior-Attenuated Fine-Tuning (PAFT) importance-resamples from the posterior and runs standard SFT. Both have bias $O\big(\frac{q}{M P_θ^{q+1}}\big)$; GARL has lower variance, PAFT has semantically coherent gradients. On FinQA, HotPotQA, and MuSiQue, GARL at $q{=}0.75$ substantially mitigates cold-start stalling, escaping cold start where GRPO fails entirely. In warm start, GARL at low $q$ dominates FinQA where training is stable; on HotPotQA and MuSiQue, GARL destabilizes during training, and PAFT at $q{=}0.75$ provides stable gradients (best overall on HotPotQA at 47.9 maj@16, $+14.4$ over GRPO).