How Fast Should a Model Commit to Supervision? Training Reasoning Models on the Tsallis Loss Continuum

• URL: http://arxiv.org/abs/2604.25907v2
• Date: Thu, 07 May 2026 17:16:40 GMT
• Title: How Fast Should a Model Commit to Supervision? Training Reasoning Models on the Tsallis Loss Continuum
• Abstract summary: We show a unified loss family $J_Q$ that interpolates between RLVR and log-marginal-likelihood over latent trajectories.<n>All members share the same per-example gradient direction, differing only by a per-instance $P_-q$ that reweights each instance independently of the learning rate.<n>We derive two Monte Carlo estimators that directly optimize fixed-$q$ on the $J_Q$ continuum, without annotated rationales.
• Score: 3.9929570259734604
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: SFT-then-RLVR is widely used for post-training reasoning models, but why this specific ordering, and why RLVR-only stalls at cold start, have lacked a unifying theoretical account. We provide that account under a unified loss family $J_Q$ using the Tsallis $q$-logarithm. $J_Q$ is a single-parameter family that interpolates between RLVR (at $q{=}0$, the \textit{exploitation pole}) and the log-marginal-likelihood over latent trajectories (at $q{=}1$, the \textit{density-estimation pole}), under which the standard pipeline corresponds to a stepwise $q{=}1 \to 0$ schedule. All members share the same per-example gradient direction, differing only by a per-instance amplification $P_θ^{-q}$ that reweights each instance independently of the learning rate. Under gradient flow analysis, we show that the exploitation pole requires $Ω(\frac{1}{p_0})$ time to escape cold start but is robust to label noise, while the density-estimation pole escapes in $Θ\big(\log(\frac{1}{p_0})\big)$ but memorizes label noise. This separation explains how SFT ($q{=}1$) first moves the model out of the cold-start regime, followed by the more robust RLVR ($q{=}0$), under the SFT-then-RLVR paradigm. We further derive two Monte Carlo estimators that directly optimize fixed-$q$ on the $J_Q$ continuum, without annotated rationales: Gradient-Amplified RL (GARL) and Posterior-Attenuated Fine-Tuning (PAFT), with shared bias $O\big(\frac{q}{M P_θ^q}\big)$ but different variance and stability properties. On FinQA, HotPotQA, and MuSiQue, GARL at sufficiently high $q$ 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 and PAFT at $q{=}0.75$ remains stable, reaching $47.9$ \texttt{m@16} on HotPotQA ($+13.9$ over GRPO).

Related papers

• Sharp Minimax Regret for Infinite-Memory Logistic Prediction [55.29259818039367]
  Lag $j$ can affect prediction by scale $r_j$ and enters only $n_T,j=T-j+1$ prediction rounds.<n>For every summable envelope, a localized mixture proves $cR_T(r)leq C_T(r)$.<n>For exponential and envelopes, under the stated finite-sample condition, a Toeplitz-design converse proves $cR_T(r)geq c_T(r)$.
  arXiv  Detail & Related papers  (2026-08-27T01:31:46Z)
• One Inverse Step is a Convex Program: Bayes-Limit Calibration of Diffusion Inversion [0.0]
  One implicit DDIM inversion step is the cheapest probe of whether a pretrained diffusion model encodes local manifold geometry.<n>The Fermi window conflicts with the model's own training support by $3.6$-$5.6times$, and the trained Hessian-Lipschitz constant is $2$-$12%$ of the curvature the law reads.<n>Finally the unconditional ceiling $_t_max(mathrmsym,J)le1$, from $mathrmCov(x_0mid x_t)succeq0$
  arXiv  Detail & Related papers  (2026-08-24T11:01:54Z)
• Sharp Stability Threshold and Certification for Designing Stable Residual Architectures [4.781989580907985]
  We propose a sharp stability threshold on the input-magnitude exponent of every residual block's velocity field.<n>The threshold $q = 1$ is established via two independent arguments.<n>An arithmetic of input-magnitude exponents under the five operations that build residual blocks enables efficient certification of $q_k le 1$ at the level of architectural primitives.
  arXiv  Detail & Related papers  (2026-07-16T05:13:44Z)
• Asymptotic-Preserving A Posteriori Analysis of Diffusion and Flow-Matching Samplers [0.6768558752130311]
  Diffusion and flow-matching samplers integrate a learned probability-flow ODE from a large noise scale down to a small terminal floor $_min$, at which the score is stiff and the flow develops a boundary layer.<n>We treat $_min$ as a singular-perturbation parameter and determine which fixed-step samplers are solvable.
  arXiv  Detail & Related papers  (2026-07-05T04:48:33Z)
• Post-Training with Policy Gradients: Optimality and the Base Model Barrier [27.674563695368665]
  We study post-training linear autoregressive models with outcome and process rewards.<n>We prove that variants of policy gradient (PG) can achieve likelihood $1 - varepsilon$ with an essentially minimax optimal number of reward queries.
  arXiv  Detail & Related papers  (2026-03-07T00:25:53Z)
• Complexity of Classical Acceleration for $\ell_1$-Regularized PageRank [14.919427330415608]
  We show that FISTA can improve the dependence on $$ while preserving the $1/$ locality scaling.<n>We analyze FISTA on a slightly over-regularized objective and show that, under a checkable confinement condition, all spurious activations remain inside a boundary set.<n>This yields a bound consisting of an accelerated $(sqrt)-1log(/varepsilon)$ term plus a boundary overhead $sqrtvol(mathcalB)/(3/2)$.
  arXiv  Detail & Related papers  (2026-02-24T17:35:46Z)
• Entropic Risk Optimization in Discounted MDPs: Sample Complexity Bounds with a Generative Model [2.4145441422386464]
  We analyze the sample complexities of learning the optimal state-action value function $Q*$ and an optimal policy $pi*$ in a finite discounted decision process.<n>We provide and analyze a simple model based approach which we call model-based risk-sensitive $Q$-value-iteration (MB-RS-QVI)
  arXiv  Detail & Related papers  (2025-05-30T22:27:57Z)
• Temperature is All You Need for Generalization in Langevin Dynamics and other Markov Processes [43.857810191928166]
  We analyze the gap between the training and test errors when training a potentially over-parametrized model.<n>We have no dependence on either training time or reliance on mixing, nor a dependence on dimensionality, gradient norms, or any other properties of the loss or model.
  arXiv  Detail & Related papers  (2025-05-25T10:49:09Z)
• On the $O(\rac{\sqrt{d}}{T^{1/4}})$ Convergence Rate of RMSProp and Its Momentum Extension Measured by $\ell_1$ Norm [54.28350823319057]
  This paper considers the RMSProp and its momentum extension and establishes the convergence rate of $frac1Tsum_k=1T.<n>Our convergence rate matches the lower bound with respect to all the coefficients except the dimension $d$.<n>Our convergence rate can be considered to be analogous to the $frac1Tsum_k=1T.
  arXiv  Detail & Related papers  (2024-02-01T07:21:32Z)
• (Accelerated) Noise-adaptive Stochastic Heavy-Ball Momentum [7.095058159492494]
  Heavy ball momentum (SHB) is commonly used to train machine learning models, and often provides empirical results over gradient descent.<n>We show that SHB can attain an accelerated mini-batch size when the mini-batch size is larger than a threshold $b*$ that on the condition number $kappa2$.
  arXiv  Detail & Related papers  (2024-01-12T18:17:28Z)
• Settling the Sample Complexity of Online Reinforcement Learning [92.02082223856479]
  We show how to achieve minimax-optimal regret without incurring any burn-in cost.<n>We extend our theory to unveil the influences of problem-dependent quantities like the optimal value/cost and certain variances.
  arXiv  Detail & Related papers  (2023-07-25T15:42:11Z)
• Depth Dependence of $\mu$P Learning Rates in ReLU MLPs [72.14317069090407]
  We study the dependence on $n$ and $L$ of the maximal update ($mu$P) learning rate. We find that it has a non-trivial dependence of $L$, scaling like $L-3/2.$
  arXiv  Detail & Related papers  (2023-05-13T01:10:49Z)
• Horizon-Free and Variance-Dependent Reinforcement Learning for Latent Markov Decision Processes [62.90204655228324]
  We study regret minimization for reinforcement learning (RL) in Latent Markov Decision Processes (LMDPs) with context in hindsight. We design a novel model-based algorithmic framework which can be instantiated with both a model-optimistic and a value-optimistic solver.
  arXiv  Detail & Related papers  (2022-10-20T21:32:01Z)
• Near-Optimal Regret Bounds for Multi-batch Reinforcement Learning [54.806166861456035]
  We study the episodic reinforcement learning (RL) problem modeled by finite-horizon Markov Decision Processes (MDPs) with constraint on the number of batches. We design a computational efficient algorithm to achieve near-optimal regret of $tildeO(sqrtSAH3Kln (1/delta))$tildeO(cdot) hides logarithmic terms of $(S,A,H,K)$ in $K$ episodes. Our technical contribution are two-fold: 1) a near-optimal design scheme to explore
  arXiv  Detail & Related papers  (2022-10-15T09:22:22Z)
• Reward-Mixing MDPs with a Few Latent Contexts are Learnable [75.17357040707347]
  We consider episodic reinforcement learning in reward-mixing Markov decision processes (RMMDPs) Our goal is to learn a near-optimal policy that nearly maximizes the $H$ time-step cumulative rewards in such a model.
  arXiv  Detail & Related papers  (2022-10-05T22:52:00Z)
• What Happens after SGD Reaches Zero Loss? --A Mathematical Framework [35.31946061894308]
  Understanding the implicit bias of Gradient Descent (SGD) is one of the key challenges in deep learning. This paper gives a general framework for such analysis by adapting ideas from Katzenberger (1991). It yields some new results: (1) a global analysis of the implicit bias valid for $eta-2$ steps, in contrast to the local analysis of Blanc et al. ( 2020) that is only valid for $eta-1.6$ steps and (2) allowing arbitrary noise covariance.
  arXiv  Detail & Related papers  (2021-10-13T17:50:46Z)
• Breaking the Sample Complexity Barrier to Regret-Optimal Model-Free Reinforcement Learning [52.76230802067506]
  A novel model-free algorithm is proposed to minimize regret in episodic reinforcement learning. The proposed algorithm employs an em early-settled reference update rule, with the aid of two Q-learning sequences. The design principle of our early-settled variance reduction method might be of independent interest to other RL settings.
  arXiv  Detail & Related papers  (2021-10-09T21:13:48Z)
• $Q$-learning with Logarithmic Regret [60.24952657636464]
  We prove that an optimistic $Q$-learning enjoys a $mathcalOleft(fracSAcdot mathrmpolyleft(Hright)Delta_minlogleft(SATright)right)$ cumulative regret bound, where $S$ is the number of states, $A$ is the number of actions, $H$ is the planning horizon, $T$ is the total number of steps, and $Delta_min$ is the minimum sub-optimality gap.
  arXiv  Detail & Related papers  (2020-06-16T13:01:33Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.