Reinforcement learning (RL), particularly RL with Verifiable Rewards (RLVR), has recently emerged as a central paradigm for enhancing large language models' (LLMs) reasoning abilities, demonstrating remarkable effectiveness across reasoning tasks. Recent studies suggest that high-entropy tokens play an exceptionally important role in model training, since training with only the highest 20% entropy tokens yields significant performance gains. However, why such high-entropy tokens are beneficial remains insufficiently understood. In this work, we find that although high-entropy tokens within one answer tend to correlate with large gradient magnitude, entropy alone fails to consistently reflect token importance across different answers, considering the variations in the answer-level reward signals. Based on this observation, we introduce the Gradient Magnitude-based Token Selection (GMTS) method to quantify token importance, which leverages the entropy-gradient connection to approximate gradient-magnitude rankings for token selection. We find that training on the top 20% tokens ranked by GMTS consistently outperforms entropy-based token selection across three reasoning domains and various model sizes, suggesting that GMTS provides a more fine-grained estimate of token contribution for RLVR training.
Recent work proposes next-chunk reasoning RL for leveraging no-CoT data---corpora such as worked solutions and textbook derivations that contain reasoning-rich content but lack explicit chain-of-thought annotations. The method trains a model to generate implicit reasoning traces and rewards them by their ability to predict the next chunk of text. While promising, existing evaluations primarily compare against conventional SFT baselines, leaving open whether the gains come from the RL formulation itself or from more effectively exposing the model to no-CoT data. We address this question with a controlled study of next-chunk reasoning RL and a simple but previously overlooked alternative: Mixed SFT, a single supervised fine-tuning stage that jointly trains on no-CoT and long-CoT data. Despite its simplicity, Mixed SFT achieves a clearly higher post-RLVR performance ceiling than next-chunk reasoning RL while requiring over 60 times less training compute. The advantage is consistent across in-domain mathematical reasoning and out-of-domain reasoning tasks. Moreover, we show that higher pre-RLVR accuracy does not necessarily translate into higher post-RLVR accuracy, highlighting the need to evaluate no-CoT training strategies in the context of the full post-training pipeline.
Mapping cyber threat intelligence (CTI) text to MITRE ATT&CK techniques is essential for structured threat analysis, yet manual annotation is costly and does not scale. The ATT&CK taxonomy comprises several hundred attack techniques, and a single CTI passage may describe multiple techniques, making accurate and complete extraction challenging. Existing automated approaches fall short in different ways: multi-label classifiers struggle with severe class imbalance and the large label space, while LLM-based methods--retrieval pipelines and fine-tuned generators--optimize token-level objectives that treat technique annotation as sequence generation rather than set prediction, lacking direct supervision on whether the predicted technique set is correct and complete. We propose TTP-R1, a two-stage framework that combines retrieval-augmented supervised fine-tuning (SFT) with reinforcement learning using verifiable rewards (RLVR). A hybrid retriever first narrows the large label space to a candidate set, and a fine-tuned LLM learns to select the correct techniques. We then apply Group Relative Policy Optimization with a decomposed reward that directly supervises the precision, recall, and output format of the predicted technique set. Across four CTI benchmarks, TTP-R1 achieves the best average F1, improving sub-technique-level F1 by 7.4 percentage points over Claude Sonnet 4.5 with retrieval augmentation, while running 28x faster when served as an 8B-parameter model on a single GPU.
Language models are increasingly promoted from examinees to examiners: they write the test suites, answer keys, rubrics, and reward functions that define correctness for other systems. We measure the capability that role assumes and find it lacking under the protocol the role is usually deployed with, one-shot greedy authoring with no test-time reasoning. Across four reference constructions - two with complete finite truth, one with a hardened executable reference (HumanEval+/MBPP+), one with an explicitly incomplete lexical reference (WordNet) - models judge whether a candidate belongs far better than they author the set itself. On the incompleteness-proof algorithmic construction the gap is +0.34 to +0.29 F1 over a 24x parameter range and does not close; on executable code, models judging at F1 0.74-0.90 author suites admitting only 19-42% of oracle-correct solutions. A control locates the deficit: asked to emit the predicate rather than its extension, the same models reach F1 about 0.99. The failure is not missing knowledge or an inability to specify, but an inability to materialise the region a specification induces. The dominant error is omission, which resists audit: an over-inclusion is a token a reviewer can challenge, a missing member an absence whose discovery is the authoring problem itself. Models detect planted over-inclusions 6-7x more often than planted omissions, and a production deployment of 43,227 items fails omission-first at 10:1. Wired into RLVR, an authored key costs 1.9 points of accuracy against an exact oracle and 18.5 WordNet-relative (six paired seeds, p=0.031). Gating authored verifiers on a known-correct probe cuts false rejection from 58-92% to at most 5%, but keeps only 5-39% of suites. Repairing them instead, by rewriting each wrong expected value to what a reference execution returns, raises yield 3.3-10.6x across four author families.
Reliable reward and preference signals are critical for evaluating and optimizing large language models on open-ended tasks. Rubric-based judges offer a transparent way to decompose such judgments into explicit evaluation criteria, but existing annotation-free rubric generators typically rely on a single generic evaluator. As a result, they may overlook important dimensions of human preference, a failure mode we term dimensional blind spots. To address this limitation, we propose Multi-Role Rubric Generation (MRRG), a training-free and reference-free framework that elicits evaluation criteria from multiple complementary roles and consolidates them into an auditable rubric-based scorer. This scorer can be used both to validate pairwise preferences and to provide rewards for GRPO-style Reinforcement Learning with Verifiable Rewards (RLVR). Experiments on preference validation benchmarks show that MRRG consistently outperforms single-role rubric generation baselines across multiple backbone models. Further RLVR experiments demonstrate that MRRG yields a stronger reward signal for improving open-ended generation.
RL with verifiable rewards (RLVR) has emerged as a powerful paradigm for training LMs on tasks with well-defined success metrics, such as code generation and mathematical reasoning. However, current RLVR methods optimize only what can be objectively scored, often neglecting subjective, non-verifiable aspects of human-like outputs, such as style and structure. This limitation leads to well-documented failure modes such as diversity collapse, unnatural-sounding responses, and reward hacking. We propose an adversarial generator-discriminator framework that augments verifiable rewards with a learned signal from human demonstrations. A generator model is trained using RL to maximize both task accuracy and an adversarial reward derived from a discriminator. The discriminator, trained alongside the generator policy, learns to distinguish human-written outputs from model-generated ones. The discriminator serves as a learned proxy for the human output distribution, providing feedback on aspects of generation that are difficult to formalize as scalar rewards. Across diverse domains, including bug fixing and open-ended generation, our approach consistently improves non-verifiable properties while preserving the accuracy gains of RLVR. In bug fixing, our method produces solutions with significantly lower edit distance compared to RLVR baselines while matching end performance. In story generation, our method significantly improves win rate while producing stories that are diverse and more human-like. And in a simple reward hacking benchmark, our method nearly eliminates model misbehavior while maintaining high benchmark scores. Together, these results show that our approach bridges RL and SFT, offering a scalable path toward jointly optimizing the verifiable and non-verifiable properties of a task.
Diversity in LLM mathematical reasoning is critical for exploration, but common diversity metrics mostly capture surface-level variation rather than differences in how a problem is solved. We address this gap by introducing approach-level diversity: variation in strategies across correct solutions to the same problem. Using a human-calibrated LLM judge framework, we show that prior diversity measures are unreliable proxies for approach-level diversity, and this mismatch carries over to diversity-aware RLVR, where target metrics are preserved while approach-level diversity declines. Investigating when approach-level diversity helps and whether it can be directly induced, we find that approach-diverse candidate sets improve test-time scaling. However, optimizing an LLM judge diversity reward during training causes the policy to exploit judge-specific preferences rather than broaden its approaches, leaving direct optimization of approach-level diversity as an open problem. Together, our work introduces the notion of approach-level diversity and uncovers a systematic divergence between surface- and approach-level signals, marking a step toward LLMs that reason in genuinely diverse, human-like ways.
Following complex instructions with multiple explicit constraints remains a fundamental challenge for large language models (LLMs). Existing alignment methods, such as DPO, optimize holistic reward signals that often underemphasize strict satisfaction of individual constraints, particularly under out-of-distribution or multi-constraint settings. In this paper, we propose STAIF, a stage-wise optimization framework that decouples the alignment of subjective (soft) constraints from the optimization of objectively verifiable (hard) constraints. Stage 1 applies preference optimization with multiple negative samples to sharpen sensitivity to soft constraints, while Stage 2 applies Reinforcement Learning with Verifiable Rewards (RLVR) to enforce strict compliance with hard constraints. To support this method, we construct STAINSTRUCT, a high-quality bilingual (English, Chinese) dataset of approximately 31,000 complex multi-constraint instructions. Extensive analyses validate the design of STAIF and show state-of-the-art performance on representative benchmarks against strong baselines, as well as genuine generalization.
While in-context learning is generally shown to be effective in Large Language Models (LLMs), bad contexts can cause performance degradation and mode collapse, a phenomenon we call "pigeonholing." **Unintentionally bad** contexts can happen without malicious jailbreaking intents: For example, a user asks the model to justify an incorrect math theorem or fails to correct the model's buggy code. Specifically, we investigate ``pigeonholing" in two scenarios: (1) when the user suggests a solution, and (2) when the conversation context includes the assistant's previous (incorrect) responses. Our experiments across 10 verifiable and open-ended tasks with 10 different models show that pigeonholing manifests in several ways: (1) repeating the incorrect answers from context (leading to 38-40% performance drop), (2) converging on a narrow set of answers in coding and text generation without exploring alternatives, and (3) flipping stance on controversial topics to align with the user or the assistant's previous claims. We find that pigeonholing worsens almost monotonically with the number of conversation turns (performance drops by additional 14+% as repeated mistakes increase from 1 to 5), and pigeonholing-induced mode collapse can happen even when the provided example is correct. As a step toward mitigation, we propose RLVR with synthetic errors which improves models by 43-60% under bad contexts compared to vanilla RLVR baselines.
Small language models are cheap to serve and feasible on local hardware, but strong public 135M-class systems are commonly trained with hundreds of billions to trillions of tokens on large clusters. We study a sharply resource-constrained regime: a complete 134.5M-parameter language-model pipeline executed on one NVIDIA L20 GPU. The released checkpoint, L20-Edu-135M, receives approximately 13B pretraining tokens: 10B FineWeb-Edu tokens followed by a 3B-token educational, mathematics, code, and reasoning mixture. We document the architecture, data gates, cross-source MinHash/LSH near-deduplication, segment deduplication, benchmark-overlap removal, throughput optimization, supervised fine-tuning (SFT) with weight interpolation, and reinforcement learning from verifiable rewards (RLVR) on GSM8K. In a self-run zero-shot six-task harness, L20-Edu-135M obtains a mean score of 0.4150. It trails SmolLM-135M (0.4767) and SmolLM2-135M (0.4917), but its mean is 87.1% of SmolLM-135M's while its nominal token count is 2.17% as large. This ratio is descriptive, not evidence of statistical equivalence or a controlled scaling law. The model exceeds several older 100M-160M public baselines under the same harness. Direct GRPO-style RLVR decreases GSM8K exact-match accuracy from 1.82% to 1.59% (192-token completions) and 1.21% (320-token completions). These single-run results identify a concrete failure mode rather than establishing a general lower bound on RLVR. The contribution is an auditable resource-constrained case study, not a state-of-the-art claim.
Reinforcement learning with verifiable rewards (RLVR) is a standard recipe for training large language models on mathematical reasoning, where an answer verifier serves as a language-neutral reward function. We show that this assumption fails in multilingual settings: an exact-match verifier turns format and script variation into language-dependent false-negative reward noise. We introduce a reusable protocol for auditing multilingual RLVR rewards: a verifier-robustness suite, a rollout-diagnosis procedure, and language-conditioned reward-error metrics for Japanese, English, and Chinese answers. On MGSM rollouts with k=8, the exact-match proxy rejects trusted-correct answers at sharply different rates by language across Qwen3-4B, Qwen3-8B, and Llama-3.1-8B-Instruct; for Qwen3-8B, the false-negative rate reaches 0.642 on JP against 0.122 on EN and 0.073 on CN. A plain-numeric probe localizes the mechanism to the final-answer interface: an interface model drives reward-error VLB to zero while the residual accuracy gap is unchanged. We then expose a cross-lingual selection bottleneck: on MGSM250 rollouts, a target-local aggregation rule using no trusted labels closes 55-78% of the average selection gap, and over 95% of repairs require genuine cross-lingual support. The bottleneck replicates on a 483-problem MATH-500 set. A controlled training audit shows that rule-GRPO raises trusted accuracy while the reward-error VLB stays high. The unifying message is operational: multilingual RLVR rewards should be audited by language and by answer interface before they are optimized.
During instruction fine-tuning (IFT), large language models (LLMs) learn to follow instructions by using the provided context to answer a query. While prior work has studied how context characteristics correlate with context usage by the LLM, this analysis has been limited to inference time, leaving open how these relationships are acquired in the first place. Here, we measure how models' sensitivity to such characteristics shifts across successive IFT stages: supervised fine-tuning (SFT), direct preference optimization (DPO), and reinforcement learning with verifiable rewards (RLVR). Experiments across four models and three datasets show that SFT makes models more likely to use contexts that are easy to understand, such as containing high length, context-query similarity, and fluency. Post-SFT dynamics may either reinforce or resolve these preferences depending on the training dataset. Our findings reveal that context usage is actively reshaped at each IFT stage, and designing a balanced IFT dataset is important in ensuring robust context utilization of instruction-tuned models.
As LLM capabilities advance rapidly, the evaluation methods used to assess them increasingly lag behind. Traditional benchmarks relied on programmatic verification of narrow, surface-level constraints, but real-world instruction following and agentic tasks demand assessment of nuanced, context-dependent behaviors that resist simple scripted checks. We present a systematic analysis of expert-curated rubric-based evaluation as an alternative paradigm, drawing on empirical evidence from two domains: complex instruction following and enterprise agentic tasks. We first articulate five design principles for constructing high-quality rubrics, including Maximum Viable Atomicity, intent-aware criterion design, and iterative LLM-judge calibration. To validate these principles, we introduce ComplexConstraints, a new expert-curated instruction-following dataset in which each prompt is paired with 10-40 atomic rubric criteria. We demonstrate that these expert rubrics are not only better evaluation instruments but also highly effective training signals: training on approximately 1,000 ComplexConstraints examples yields +15.5% improvement for a 4B-parameter model and +12.2% for a 235B-parameter model on instruction following, while single-epoch RL training on a rubric-graded enterprise environment produces gains that transfer to out-of-distribution benchmarks the model was never trained on (+4.5% BFCL, +7.4% Tau2-Bench, +6.8% Tool-Decathlon). Our findings establish that expert-authored rubrics improve both the measurement and the development of frontier LLM capabilities, serving as effective evaluation and RL training signals.
Large Reasoning Models (LRMs) have achieved remarkable progress thanks to Reinforcement Learning with Verifiable Rewards (RLVR) on Chain-of-Thoughts (CoTs). However, since long CoTs naturally contain trial and errors and mainstream RLVR approaches choose outcome-correct CoT trajectories for memorization, the redundant explorations in long CoTs are inevitably reinforced, which results in the over-thinking issues of LRMs. Previous attempts to resolve this issue mainly give more advantage to shorter trajectories, yet their learning signals are still outcome-based and cannot reduce the memorization of redundant explorations in long CoTs. Therefore, we propose ThoughtFold, a framework that leverages fine-grained preference learning to mitigate redundant explorations for efficient reasoning. ThoughtFold employs an introspective strategy to identify redundancy within each correct trajectory, which yields a spectrum of candidate sub-trajectories. Leveraging this spectrum, we introduce a masked preference optimization objective that explicitly penalizes redundant explorations and encourages the model to directly bridge essential reasoning segments, effectively folding its reasoning chains into a more concise path. Extensive experiments show that ThoughtFold significantly enhances efficiency. It reduces the token usage of DeepSeek-R1-Distill-Qwen-7B by approximately 56% while maintaining state-of-the-art accuracy.