Pith. sign in

REVIEW 3 major objections 2 minor

Curating expert-annotated instruction data for verifiable rewards extends reinforcement learning to general reasoning tasks in language models.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

T0 review · grok-4.3

2026-05-10 17:12 UTC pith:2GWQW6DN

load-bearing objection SUPERNOVA shows that targeted selection from instruction datasets can lift RLVR performance on general reasoning benchmarks, but the 100+ experiments leave too many variables unaccounted for to credit the gains cleanly to the curation rules. the 3 major comments →

arxiv 2604.08477 v2 pith:2GWQW6DN submitted 2026-04-09 cs.AI cs.CLcs.LG

SUPERNOVA: Eliciting General Reasoning in LLMs with Reinforcement Learning on Natural Instructions

classification cs.AI cs.CLcs.LG
keywords reinforcement learninglarge language modelsdata curationgeneral reasoningverifiable rewardsinstruction tuningtask selectionbenchmark evaluation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents SUPERNOVA, a framework that selects and mixes tasks from existing instruction-tuning datasets to create training data for reinforcement learning with verifiable rewards. This targets the gap where RLVR has succeeded in math and code but struggles with broader skills such as causal inference and temporal reasoning. Over 100 controlled experiments show that source task selection based on individual target performance beats average-based mixing, and that synthetic quality improvements further help. Models trained this way outperform strong baselines on BBEH, Zebralogic, and MMLU-Pro, with relative gains reaching 52.8 percent on BBEH across sizes. The results indicate that principled curation of human-annotated data can systematically adapt natural instructions into effective RL signals for general reasoning.

Core claim

SUPERNOVA shows that instruction-tuning datasets containing expert-annotated ground-truth encode reasoning patterns that can be adapted into high-quality verifiable rewards for RLVR. Source task selection proves non-trivial and matters more when done per target task than by overall average performance. Task mixing strategies and synthetic interventions for data quality also affect outcomes. Models trained on the resulting data surpass baselines including Qwen3.5 on challenging general reasoning benchmarks, achieving up to 52.8 percent relative improvement on BBEH.

What carries the argument

SUPERNOVA, a data curation framework that examines source task selection, task mixing strategies, and synthetic interventions to convert natural instructions into RLVR training data.

Load-bearing premise

That expert-annotated instruction datasets contain reasoning patterns that can be reliably converted into verifiable reward signals without losing their effectiveness for general tasks.

What would settle it

Training models on SUPERNOVA-curated data and measuring no outperformance or even underperformance relative to baselines on BBEH, Zebralogic, and MMLU-Pro would falsify the central effectiveness claim.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Source task selection tailored to individual target tasks yields better downstream results than selection based on average performance across tasks.
  • Synthetic interventions improve the quality of adapted data for use in verifiable-reward training.
  • Principled curation of human-annotated resources enables RLVR to address general reasoning skills beyond formal domains.
  • Models trained under this approach deliver measurable gains on benchmarks requiring causal inference and temporal understanding.
  • The 100-plus controlled experiments isolate the impact of data design choices from other training variables.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Similar curation principles could be tested on other human-annotated resources to extend RLVR into domains with sparse verifiable signals.
  • The findings suggest that data design choices may interact with model scale in ways that warrant separate scaling experiments.
  • Automated versions of the per-target task selection process could reduce the manual effort needed to apply the framework.
  • The work implies that future RLVR pipelines might benefit from hybrid human-synthetic data pipelines rather than purely synthetic generation.
  • Extensions might explore whether the same curation logic applies when the base model has already undergone extensive instruction tuning.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 2 minor

Summary. The paper proposes SUPERNOVA, a data curation framework that adapts expert-annotated instruction-tuning datasets into verifiable rewards for RLVR to elicit general reasoning capabilities (causal inference, temporal understanding) in LLMs. It reports results from over 100 controlled RL experiments examining source task selection, task mixing strategies, and synthetic interventions, concluding that per-target performance-based task selection outperforms average-performance selection and that SUPERNOVA-trained models achieve relative gains of up to 52.8% on BBEH (plus improvements on Zebralogic and MMLU-Pro) over strong baselines such as Qwen3.5.

Significance. If the reported gains prove robust under full experimental controls, the work would offer concrete, actionable guidance for extending RLVR from formal domains to general reasoning by leveraging existing human-annotated resources, with the open-sourced code and data providing a useful starting point for the community.

major comments (3)
  1. [Experimental section] Experimental section (description of the 100+ RL runs): the manuscript states that experiments are 'controlled' and isolate the effects of source task selection, mixing, and synthetic interventions, yet supplies no complete hyperparameter table, multi-seed variance statistics, or explicit verification that reward functions (exact match vs. LLM judge) and other non-data variables remained identical across all compared configurations. Without these, performance deltas cannot be confidently attributed to the three studied factors.
  2. [Results on BBEH] Results on BBEH (and related benchmark tables): the headline relative improvement of up to 52.8% is presented without accompanying statistical significance tests, standard-error bars, or precise baseline implementation details (e.g., whether Qwen3.5 was instruction-tuned with the same RLVR setup or merely prompted). This leaves open the possibility that gains arise from unaccounted confounds rather than the proposed data-curation principles.
  3. [Task-selection ablation] Task-selection ablation: the claim that 'selecting tasks based on their performance for individual target tasks outperforms strategies based on overall average performance' requires clarification on how per-target performance was measured (on which split, with which metric) and whether selection was performed on held-out data to prevent leakage into the final evaluation.
minor comments (2)
  1. [Abstract] Abstract: the acronym RLVR is used before its expansion ('Reinforcement Learning with Verifiable Rewards') is given; a parenthetical expansion on first use would improve readability.
  2. [Data-curation description] Data-curation description: the precise mapping from expert-annotated ground-truth answers to automatic verifiable reward functions is described at a high level; a short pseudocode or explicit example for one task type would clarify how 'rich reasoning patterns' are preserved versus introducing shortcuts.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive and detailed feedback. The comments highlight important aspects of experimental rigor and clarity that we will address in the revision. Below we respond point-by-point to the major comments.

read point-by-point responses
  1. Referee: [Experimental section] Experimental section (description of the 100+ RL runs): the manuscript states that experiments are 'controlled' and isolate the effects of source task selection, mixing, and synthetic interventions, yet supplies no complete hyperparameter table, multi-seed variance statistics, or explicit verification that reward functions (exact match vs. LLM judge) and other non-data variables remained identical across all compared configurations. Without these, performance deltas cannot be confidently attributed to the three studied factors.

    Authors: We agree that additional documentation will strengthen the experimental section. In the revised manuscript we will add a complete hyperparameter table in the appendix covering all RL runs. All original experiments used a fixed random seed for reproducibility; we will also report variance statistics from a subset of multi-seed replications (three seeds) for the key configurations. We will insert explicit statements confirming that reward functions (exact match versus LLM judge) and all non-data variables were held identical across compared runs, consistent with the controlled design described in the paper. These additions will make the attribution of performance differences to the studied factors fully transparent. revision: partial

  2. Referee: [Results on BBEH] Results on BBEH (and related benchmark tables): the headline relative improvement of up to 52.8% is presented without accompanying statistical significance tests, standard-error bars, or precise baseline implementation details (e.g., whether Qwen3.5 was instruction-tuned with the same RLVR setup or merely prompted). This leaves open the possibility that gains arise from unaccounted confounds rather than the proposed data-curation principles.

    Authors: We will strengthen the results presentation by adding standard-error bars and statistical significance tests (paired t-tests across seeds) to the BBEH and related benchmark tables. For the baseline, Qwen3.5 was trained with the identical RLVR procedure and hyperparameters as the SUPERNOVA models (not merely prompted); we will clarify this explicitly in the text and provide the corresponding implementation details. These revisions will reduce the possibility of unaccounted confounds and allow readers to assess the robustness of the reported relative gains. revision: yes

  3. Referee: [Task-selection ablation] Task-selection ablation: the claim that 'selecting tasks based on their performance for individual target tasks outperforms strategies based on overall average performance' requires clarification on how per-target performance was measured (on which split, with which metric) and whether selection was performed on held-out data to prevent leakage into the final evaluation.

    Authors: We will expand the task-selection ablation section to provide the requested details. Per-target performance was measured on a held-out validation split (separate from all test benchmarks) using exact-match accuracy as the metric. Task selection was performed exclusively on this validation split before any final evaluation, thereby preventing leakage. The revised text will include the exact split sizes, the validation metric, and the full selection procedure to make the ablation fully reproducible and free of data leakage concerns. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical results from controlled experiments

full rationale

The paper reports outcomes from 100+ RL training runs on curated instruction data, with performance measured on external benchmarks (BBEH, Zebralogic, MMLU-Pro). No equations, fitted parameters renamed as predictions, self-definitional constructs, or load-bearing self-citations appear in the provided text. Claims rest on direct experimental deltas rather than any derivation that reduces to its own inputs by construction. This is self-contained empirical work.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on one domain assumption about the content of existing datasets; no free parameters or invented entities are introduced.

axioms (1)
  • domain assumption Instruction-tuning datasets containing expert-annotated ground-truth encode rich reasoning patterns that can be systematically adapted for RLVR
    This is the key insight stated in the abstract that enables the entire data curation framework.

pith-pipeline@v0.9.0 · 5638 in / 1446 out tokens · 69565 ms · 2026-05-10T17:12:37.357342+00:00 · methodology

0 comments
read the original abstract

Reinforcement Learning with Verifiable Rewards (RLVR) has substantially improved reasoning in formal domains such as mathematics and code, but extending these gains beyond STEM remains challenging. Extending RLVR beyond STEM is fundamentally constrained by the lack of high-quality verifiable training data. In this work, we introduce SUPERNOVA, a framework for curating RLVR data from natural instruction datasets, which are a rich source of expert-annotated data but are underexplored for RLVR training. Through 100+ controlled RL experiments, we systematically study how to utilize these dataset for RLVR and how data curation decisions affect downstream reasoning performance . In particular, we investigate three data designs: (a) source task selection, (b) task mixing, and (c) synthetic interventions. Our analysis reveals that source task selection has a significant impact on downstream reasoning performance. Moreover, selecting tasks based on their performance for individual target tasks outperforms strategies based on overall average performance and synthetic interventions do not improve reasoning. Guided by these insights, we construct SUPERNOVA, a high-quality RLVR dataset of 25K instances curated from natural instruction datasets. We show that training Qwen3-0.6B on SUPERNOVA outperforms the base Qwen3-0.6B, yielding a relative gain of 64.4pp on BigBench Extra Hard (BBEH), a challenging benchmark comprising 23 complex reasoning tasks. Importantly, we find that gains from SUPERNOVA generalize to unseen benchmarks, larger model scales, and newer model families. Overall, our findings provide practical insights for curating human-annotated resources to extend RLVR to general reasoning. Models, Data, Code at https://github.com/asuvarna31/supernova.

Figures

Figures reproduced from arXiv: 2604.08477 by Ashima Suvarna, Hritik Bansal, Kendrick Phan, Mehrab Beikzadeh, Saadia Gabriel.

Figure 1
Figure 1. Figure 1: SuperNova elicits strong general reasoning. We show that training with our curated SUPERNOVA data leads to consistent pass@k improvements across varying values of k on a challenging benchmark, BBEH-test. We highlight that SUPERNOVA is effective on various models sizes from Qwen3 family. 1 arXiv:2604.08477v1 [cs.AI] 9 Apr 2026 [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: SUPERNOVA Framework: In this work, we curate reasoning data from natural instruction to enhance general reasoning capabilities in LLMs. First, we study the impact of task selection on downstream reasoning performance. Then, we explore strategies to mix diverse tasks in source data. Finally, we examine whether synthetic data interventions can enhance data quality and improve downstream reasoning. reasoning … view at source ↗
Figure 3
Figure 3. Figure 3 [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparison with Other Reasoning Datasets. We report the relative gains achieved by training Qwen3-0.6B on SUPERNOVA and existing reasoning datasets. 8 [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: (Left) We train Qwen3.5-2B and LLaMA3.2-3B-Instruct with SUPERNOVA and show relatives gains over the baseline model on BBEH-mini. (Right) We show the performance comparison between the baseline model and SUPERNOVA-0.6B by scaling values of k till 128 on BBEH-mini. 4B outperforms Qwen3-4B by 21pp. This suggests that training on SUPERNOVA enhances logical reasoning capabilities, particularly for constraint s… view at source ↗
Figure 6
Figure 6. Figure 6: Pass@k accuracy of task-specific models across various values of k. [PITH_FULL_IMAGE:figures/full_fig_p016_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: (a)We categorize the source tasks based on target reasoning skill and task type. We [PITH_FULL_IMAGE:figures/full_fig_p016_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: (Left) Correlation between semantic similarity and task performance. (Middle) Correlation between lexical similarity and task performance. (Right) Correlation between win rate and task performance. We observe weak correlation for all three approaches with downstream task performance on BBEH-mini [PITH_FULL_IMAGE:figures/full_fig_p019_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Distribution of reasoning skills in SUPERNOVA. H Similarity and Task Difficulty We find that semantic similarity and lexical similarity between the tasks and validation benchmark are poor predictors of task utility for RLVR. As shown in [PITH_FULL_IMAGE:figures/full_fig_p019_9.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.