CS Paper Radar

Outside the Purview of the Reward: Capability Drift in Multilingual Post-Training

Sutanshu Saksena Raj, Purusharth Malik, Pasunuti Prasanjith, Daksh Maan, Parijat Chatterjee, Avinash Benki

cs.CL cs.AI ·submitted 2026-11-11 16:55 UTC · v1
top venuebenchmark

Abstract

Post-training can improve rewarded behavior while weakening capabilities users still need. We study two GRPO branches from a shared supervised checkpoint in a multilingual mixture-of-experts model. The released branch gains 36.07 points on MATH but loses 19.32 F1 points on XQuAD-Indic, while XQuAD-English improves under the same scoring rule. An independently trained sibling adds Indic instruction following and language-aware shaping. It recovers the regressing Indic task scores and preserves MATH, but does not restore script compliance and trades off competition mathematics and code. In a later round, judge training is followed by a sharp extraction decline from the Indic branch; reward-free distillation from its earlier checkpoint restores most of the lost F1. A MATH score decomposition associates most of the measured gain with increased boxed-answer emission; accuracy among boxed answers changes little. These results show why benchmark scores should be read alongside the response forms they reward.

Comment
Accepted at the Foundations of LLM Post-Training in Changing Environments, NeurIPS 2026

Closest papers

Nearest neighbours by meaning across the whole index.

Indic-Task Gains and Regressions after Reinforcement Learning

cs.CL cs.AI ·2026-11-12 ·Sutanshu Saksena Raj, Daksh Maan, Avinash Benki, Parijat Chatterjee, Pasunuti Prasanjith +1 more
publishedbenchmark similarity 0.75
Abstract · Accepted at the GlobalSouthAI Workshop, NeurIPS 2026

We run reinforcement learning (RL) with GRPO across four environments, starting from a supervised fine-tuned 30B-A3B mixture-of-experts model covering English and eight Indic languages. We compare the model before and after RL across 34 benchmarks in one evaluation campaign. MATH improves by 36.07 points, and most Indic benchmarks also improve. Yet XQuAD-Indic F1 falls 19.32 points as English answers rise from 8.7% to 91.7%; IndicQA falls 8.64, while English XQuAD gains 11.71 F1. A second RL run from the same supervised initialization adds an Indic environment together with language and repetition rewards. Compared with the released model, it improves both passage-answering and instruction-following benchmarks, along with preserving MATH performance. But it loses ground on reasoning and coding benchmarks. Several training choices changed together, so we cannot isolate the effect of any one of the reward terms. Script compliance remains unresolved. Adding a language reward did not ensure use of the requested script. Passage answering and script compliance therefore need direct evaluation, even when other benchmark scores improve.

0.15

Impossible Code, Barely Changed Score: A Replay Study of Judge-Based Evaluation

cs.SE cs.AI ·2026-11-01 ·Masoud Fetanat
code/data similarity 0.72
Abstract

Language models are often used to judge generated code without running it. A useful evaluation should reveal substantial changes in the answers it scores. We examine this requirement by repeating an earlier evaluation on 421 coding problems, comparing Qwen-2.5-7B-Instruct with two versions trained using Group Relative Policy Optimization (GRPO). One receives rewards from execution tests (CE-trained), and the other from a strong language-model judge (SLJ-trained). The base model's syntax-failure rate remains stable, but the trained checkpoints produce invalid Python on 63%-67% of attempts. Despite this change, the strong judge's Pass@32 scores fall by only about 3 percentage points, from 96.67% to 93.95% for SLJ-trained. Within the later run, allowing more attempts makes the base and trained checkpoints look increasingly similar under the judge. With 128 attempts per question, the judge accepts at least one SLJ-trained answer on 96.44% of problems, compared with 28.27% under execution tests. We recommend reporting the share of answers that pass Python's syntax check alongside every judge-based score. This check requires no tests or code execution and flags invalid answers that the judge accepts, although it cannot establish functional correctness. Our findings show how a high judge-based score can hide substantial output damage and motivate checking an evaluation's sensitivity to change. Project page and code: https://curiousorbitnotes.github.io/evorobust-replay-audit/

0.31

CONCUR: Concordant Auxiliary Rewards for Multi-Task, Multi-Reward Reinforcement Learning

cs.CL ·2026-10-27 ·Jixuan Leng, Si Si, Fei Wang, Cho-Jui Hsieh, Jiong Zhang +2 more
similarity 0.68
Abstract

When a Large Language Model (LLM) is trained with Reinforcement Learning (RL) on multiple tasks, a task's primary reward can gate auxiliary rewards such as a length reward: on a failing response, an auxiliary cannot exceed the primary. However, group-relative advantage estimators treat a gated auxiliary like any other component, so it can reverse the primary's order, and at the primary's ceiling only auxiliaries order responses. Keeping gated auxiliaries hurts some tasks, while dropping them hurts others. We propose CONCUR, an advantage estimator that replaces the standardized scores of each gated auxiliary with the closest scores that satisfy two conditions: responses with higher primary rewards cannot receive lower scores, and responses at the primary's ceiling receive equal scores. CONCUR-T applies both conditions throughout training, and CONCUR-A lifts the second for an auxiliary that agrees with the primary's order more often than it reverses it among passing responses. On three models from two families and across math, code, instruction following, and writing, both variants score higher on average than all baselines and than when all the gated auxiliaries are dropped.

0.18

What Does Post-Training Change in Multilingual Reasoning?

cs.CL cs.AI ·2026-09-29 ·Hongyang Li, Xiao Li, Caesar Wu, Grégoire Danoy, Pascal Bouvry
code/data similarity 0.67
Abstract · 20 pages, 9 figures, 21 tables. Main paper and supplementary material in one document

Open-source reasoning models provide unequal access to reasoning capability across languages. When a model can solve a problem but cannot deliver a complete solution in the user's language, language becomes an access barrier rather than merely a source of performance variation. We audit Qwen3 checkpoints on competition-mathematics tasks in eleven languages. Across the ten non-English languages, only 15.4-17.9% of problems receive a correct, terminating solution with visible reasoning in the requested language in any of 16 samples, compared with 92.9% in English. To identify the source of this disparity, we evaluate thirteen endpoints from one model family, spanning released checkpoints, multilingual supervised fine-tuning (SFT) at two scales, controlled SFT ablations, and three reinforcement-learning (RL) reward formulations. We jointly track correctness, language adherence, termination, and delivery efficiency. The dominant bottleneck shifts across post-training stages. Released models often reason in English. Multilingual SFT restores target-language reasoning, but accuracy declines across multilingual, English-only, and single-language SFT runs, showing that this cost is not specific to multilingual mixing; non-English reasoning traces additionally become prone to non-terminating loops. RL restores termination in both arms at no cost in accuracy, but only the arm whose reward includes a language term delivers: rewarding correctness alone returns the model to English. Together, these stages establish a constructive post-training path from English-pivoted capability to multilingual reasoning that is reliably delivered.

0.33

Scaling Laws for Generative Reward Models

cs.AI ·2026-10-22 ·Angel Raychev, Nathan Lile, Rafael Rafailov
code/data similarity 0.66
Abstract · 28 pages. Code and research artifacts: https://github.com/RaychevAngel/generative-reward-models

We study generative reward models (GenRMs) as online preference providers for policy optimization in creative writing. Across Qwen3 reward-model sizes from 0.6B to 14B, Thinking judges trained with GRPO have higher static Gold-label accuracy than Answer-Only judges trained with SFT. This ranking often reverses when the judges train policies through online DPO: Answer-Only rewarders yield higher peak and top-five-average policy Elo in seven of eight matched settings under a fixed, learned 32B Gold evaluator. Intermediate Thinking-judge checkpoints can also outperform later, more statically accurate checkpoints as rewarders. Sweeps over rewarder size, policy size, training duration and DPO regularization show substantial benefits from reward-model capacity, including judges larger than the policy. We analyze trajectories with evaluator-specific Elo arenas, conditional match-resampling intervals and descriptive scaling fits. These results demonstrate an evaluator-rewarder gap in the studied pipelines: fixed-distribution judging accuracy does not determine downstream optimization quality. The comparison changes judge-training objective as well as response format, and the main ranking is conditional on its learned reference rather than a universal human-quality result. A limited follow-up using a different learned judge is inconclusive overall.

0.52

Rubric Dropout: A Simple Way to Mitigate Reward Hacking in Rubric-Based RL

cs.LG cs.AIcs.CL ·2026-08-12 ·Minglai Yang, Xinyu Guo, Zhengliang Shi, Utkarsh Tyagi, Mian Zhang +6 more
highly cited similarity 0.66
Abstract · 24 pages, 14 figures, 5 tables

Rubric-based reinforcement learning, in which an LLM judge scores responses against prompt-specific criteria, has become a standard way to post-train language models on tasks without verifiable answers. The rubric, however, is only a proxy for response quality, and a policy trained against it long enough learns to exploit the difference. We train models with GRPO on medical and science rubrics and grade their responses on out-of-distribution (OOD) benchmarks throughout training with two judges: the training judge and a stronger OOD judge. The two scores first rise together and then diverge. The policy is learning to satisfy the training judge without writing better responses, which is a form of reward hacking. To mitigate it, we propose Rubric Dropout, which applies dropout to rubric criteria: at each training step, a random subset of each rubric's criteria is removed from the reward, so the rewarded subset of criteria is resampled at each step. All responses sampled for the same prompt are scored on the same subset, so their rewards remain comparable, and evaluation always uses the full rubric. We also introduce Block Dropout, which drops semantically similar criteria together so that a near duplicate of a dropped criterion cannot keep rewarding the same behavior. Across both benchmarks and models from 1.7B to 8B, dropout improves the gold judge's score on OOD benchmarks by up to 3 percent on HealthBench-Hard and 7 percent on ResearchQA, and it reduces both the gap between the two judges and over-crediting. It keeps in-domain reward close to the baseline, and dropout rates of 30-50 percent work best. Block Dropout improves the gold judge's OOD score further at the same dropout rate. Changing which criteria are rewarded at each step is a simple, inexpensive way to make rubric rewards harder to exploit.

0.50

Prompting Strategy Effects Depend on the Judge: A Multilingual Study of LLM-as-a-Judge Robustness

cs.CL cs.AIcs.LG ·2026-10-22 ·Asal Mehrabi, Mohammad Aliannejadi, Clemencia Siro
benchmark similarity 0.66
Abstract · 23 pages,

LLM-as-a-Judge evaluation is increasingly applied to multilingual tasks, yet prompting strategies such as Chain-of-Thought and Role Conditioning are recommended based solely on English benchmarks. Whether their effectiveness transfers across judge models and languages remains unknown: prompt-sensitivity studies are largely English-centric, while existing multilingual judge studies fix the prompt template. In this work, we address this gap by measuring this interaction through a benchmark of 9 prompting strategies, 4 open-weight judges, 27 languages, and 4 tasks across 13 datasets, yielding over 1M pointwise scores on gold-reference outputs, validated against human judgments. We find that prompting strategies are not transferable: the choice of the judge model dominates the choice of prompting strategy, strategy effects reverse direction across judges, and stability degrades in low-resource languages, where these effects are concealed by aggregate reporting. These findings show that prompting effects in multilingual evaluation are conditioned on the judge, task, and language. Consequently, prompting strategies should be validated for the intended evaluation setting, and cross-study comparisons should account for both judge and prompt configuration.

0.25

Training Negatives Still Matter After Calibrating Answerability Readouts

cs.CL ·2026-10-19 ·Jinchen Zhang, Zheyuan Liu, Yihan Zhu, Meng Jiang
similarity 0.66
Abstract · 20 pages, 5 figures

Answerability readouts predict whether a passage supplies an answer, but whether calibration can compensate for removing negative passages during training remains unclear. We present a controlled study of how training-negative construction affects calibrated prediction, ranking, and downstream answer retention. We compare native training, which retains negative passages, with empty training, which removes them, holding questions, labels, positive examples, and model weights fixed. Both are calibrated on the same labeled native examples. On SQuAD 2.0 and RepLiQA, native-trained logistic readouts at selected depths achieve lower Brier loss for Qwen2.5-7B-Instruct and Gemma-2-9B-it, with or without output scores. Native fusion also outperforms calibrated output alone. Post-hoc AUROC analysis shows better ranking under native training, beyond what threshold adjustment alone can explain. On the official SQuAD 2.0 development set, native readout policies improve F1 over empty policies but retain more answers to unanswerable questions. The advantage varies by estimator and depth, while RepLiQA's text shortcuts limit claims about contextual understanding. Unfinished generations prevent complete answer-level evaluation, although worst-case bounds preserve the native-empty F1 gains. Our findings show that training negatives affect ranking and downstream decisions even after common calibration, without establishing a universal advantage for native training.

0.19