Test-time RL alignment exposes task familiarity artifacts in LLM benchmarks

• URL: http://arxiv.org/abs/2603.12875v1
• Date: Fri, 13 Mar 2026 10:24:19 GMT
• Title: Test-time RL alignment exposes task familiarity artifacts in LLM benchmarks
• Abstract summary: The train-before-test approach controls for task familiarity by giving each model task-relevant training before evaluation.<n>We propose a two-stage test-time reinforcement learning (RL) alignment method for train-before-test.
• Score: 20.988724384459392
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Direct evaluation of LLMs on benchmarks can be misleading because comparatively strong performance may reflect task familiarity rather than capability. The train-before-test approach controls for task familiarity by giving each model task-relevant training before evaluation, originally through supervised finetuning. However, suitable training data is often hard to come by, and evaluation results vary with the data chosen. In this paper, we propose a two-stage test-time reinforcement learning (RL) alignment method for train-before-test. First, RL with a single sample provides a first alignment of the model to the task format, and second, test-time RL with majority-voting reward aligns the model to the benchmark distribution. Our test-time RL alignment method aligns similarly well as SFT-based train-before test, but without requiring a task-specific training set. On a domain-specific benchmark without training data, we show that direct evaluation underestimates base models which perform substantially better once aligned, yielding a more faithful evaluation of their capabilities. Moreover, for reasoning tasks, the performance gap between fine-tuned models and their base models largely disappears after alignment, suggesting that many gains from RLVR/SFT reported in the literature are not a difference in reasoning capability, but rather artifacts of task familiarity.

Related papers

• Beyond Solvability: Task Learnability as a Static Prior for LLM RL Post-Training [52.78248352851176]
  Reinforcement learning (RL) has become a central post-training paradigm for eliciting reasoning capabilities in large language models.<n>We study this residual axis as task learnability: a regime-conditional measure of expected positive response to continued training.<n>We propose TrajVal, a lightweight probe-based estimator that approximates per-task learnability from a short probe run and two endpoint evaluations.
  arXiv  Detail & Related papers  (2026-08-10T07:43:05Z)
• Unified Data Selection for LLM Reasoning [66.0118347918729]
  Training Large Language Models (LLMs) for complex, long-CoT reasoning is often bottlenecked by the need for massive high-quality reasoning data.<n>We propose High-Entropy Sum (HES), a training-free metric that quantifies reasoning quality by summing only the entropy of the top (e.g., 0.5%) highest-entropy tokens in each reasoning sample.
  arXiv  Detail & Related papers  (2026-05-21T12:21:41Z)
• DAJ: Data-Reweighted LLM Judge for Test-Time Scaling in Code Generation [30.131052926559956]
  We propose DAJ, a reasoning-based LLM judge trained with rewards under a bi-level data-reweighted learning framework.<n>Our approach automatically emphasizes hard problems, in-distribution samples, and trajectory-aligned data, without relying on hand-crafted verifiables.
  arXiv  Detail & Related papers  (2026-01-29T19:04:24Z)
• Self-Improving LLM Agents at Test-Time [49.9396634315896]
  One paradigm of language model (LM) fine-tuning relies on creating large training datasets.<n>In practice, gathering large sets of data is inefficient, and training on them is prohibitively expensive.<n>We study two variants of this approach: Test-Time Self-Improvement (TT-SI) and Test-Time Distillation (TT-D)
  arXiv  Detail & Related papers  (2025-10-09T06:37:35Z)
• Mirage or Method? How Model-Task Alignment Induces Divergent RL Conclusions [22.83151273022573]
  counterintuitive phenomena have been reported in large language models (LLMs)<n>We identify a key factor that differentiates RL observations: whether the pretrained model already exhibits strong Model-Task Alignment.<n>Our findings show that while standard RL training remains consistently robust across settings, many of these counterintuitive results arise only when the model and task already exhibit strong model-task alignment.
  arXiv  Detail & Related papers  (2025-08-28T20:02:10Z)
• Test-time Offline Reinforcement Learning on Goal-related Experience [50.94457794664909]
  Research in foundation models has shown that performance can be substantially improved through test-time training.<n>We propose a novel self-supervised data selection criterion, which selects transitions from an offline dataset according to their relevance to the current state.<n>Our goal-conditioned test-time training (GC-TTT) algorithm applies this routine in a receding-horizon fashion during evaluation, adapting the policy to the current trajectory as it is being rolled out.
  arXiv  Detail & Related papers  (2025-07-24T21:11:39Z)
• Continuous Self-Improvement of Large Language Models by Test-time Training with Verifier-Driven Sample Selection [6.471199527741301]
  We introduce a new framework called VDS-TTT - Verifier-Driven Sample Selection for Test-Time Training.<n>We use a learned verifier to score a pool of generated responses and select only from high ranking pseudo-labeled examples for fine-tuned adaptation.<n>We fine-tune only low-rank LoRA adapter parameters, ensuring adaptation efficiency and fast convergence.
  arXiv  Detail & Related papers  (2025-05-26T03:54:47Z)
• Bridging SFT and DPO for Diffusion Model Alignment with Self-Sampling Preference Optimization [67.8738082040299]
  Self-Sampling Preference Optimization (SSPO) is a new alignment method for post-training reinforcement learning.<n>SSPO eliminates the need for paired data and reward models while retaining the training stability of SFT.<n>SSPO surpasses all previous approaches on the text-to-image benchmarks and demonstrates outstanding performance on the text-to-video benchmarks.
  arXiv  Detail & Related papers  (2024-10-07T17:56:53Z)
• Training Language Models to Self-Correct via Reinforcement Learning [98.35197671595343]
  Self-correction has been found to be largely ineffective in modern large language models (LLMs) We develop a multi-turn online reinforcement learning approach, SCoRe, that significantly improves an LLM's self-correction ability using entirely self-generated data. We find that SCoRe achieves state-of-the-art self-correction performance, improving the base models' self-correction by 15.6% and 9.1% respectively on MATH and HumanEval.
  arXiv  Detail & Related papers  (2024-09-19T17:16:21Z)
• Making Pre-trained Language Models both Task-solvers and Self-calibrators [52.98858650625623]
  Pre-trained language models (PLMs) serve as backbones for various real-world systems. Previous work shows that introducing an extra calibration task can mitigate this issue. We propose a training algorithm LM-TOAST to tackle the challenges.
  arXiv  Detail & Related papers  (2023-07-21T02:51:41Z)
• Listen, Adapt, Better WER: Source-free Single-utterance Test-time Adaptation for Automatic Speech Recognition [65.84978547406753]
  Test-time Adaptation aims to adapt the model trained on source domains to yield better predictions for test samples. Single-Utterance Test-time Adaptation (SUTA) is the first TTA study in speech area to our best knowledge.
  arXiv  Detail & Related papers  (2022-03-27T06:38:39Z)

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.