Does Reasoning Preserve Alignment? On the Trustworthiness of Large Reasoning Models

Instruction-tuned LLMs are increasingly converted into reasoning models through post-training to improve multi-step task performance. This conversion is usually optimized for reasoning accuracy, without explicitly preserving the alignment behavior of the instruction-tuned model, such as safe refusal, bias avoidance, and privacy protection. We ask: does this conversion preserve alignment? We study this question through a trustworthiness audit and find that it is not behavior-preserving by default. For a systematic analysis, we compare reasoning models produced via supervised fine-tuning, RL-based post-training, and distillation against matched instruction-tuned baselines across six trustworthiness dimensions: safety, toxicity, stereotyping and bias, machine ethics, privacy, and out-of-distribution robustness. We observe that reasoning models often improve on reasoning benchmarks but exhibit alignment regressions, including increased toxicity, amplified stereotyping, miscalibrated refusal, and contextual privacy leakage. These regressions are consistent with behavioral drift from the instruction-tuned baseline, measured by KL divergence. Overall, our results point to the broader conclusion that trustworthiness metrics are essential for evaluating reasoning models and should be reported alongside gains in reasoning capability.

FlowBank: Query-Adaptive Agentic Workflows Optimization through Precompute-and-Reuse

Large Language Model (LLM)-based multi-agent systems are increasingly powerful, but current agentic workflow optimization paradigms make an unsatisfying trade-off. Task-level methods spend substantial offline compute yet deploy only a single workflow, leaving complementary candidates unused, while query-level methods synthesize a new workflow per query at substantial inference cost. Our motivating analysis shows these paradigms are more complementary than competing: workflows discovered during offline search often solve different subsets of queries, and many queries handled by expensive query-level generation can already be solved by cheaper precomputed workflows. This suggests a different objective: rather than searching for one universally best workflow or regenerating one per instance, we should build a compact bank of reusable, complementary workflows and select among them adaptively at inference time. Doing so requires solving three coupled problems: generating complementary rather than redundant candidates, compressing them into a small deployable portfolio, and assigning each query to the right workflow under a performance-cost trade-off. To this end, we present FlowBank, a three-stage framework for portfolio-based agentic workflow optimization. Diversifying proposes DiverseFlow to steer search toward under-covered queries and produce a high-coverage candidate pool. Curating proposes CuraFlow to compress this pool into a compact portfolio with minimal redundancy. Matching casts deployment as edge-value prediction on a query-workflow bipartite graph and routes each incoming query to the portfolio member with the best predicted utility. Across five benchmarks, FlowBank achieves the highest average score among the evaluated methods while remaining cost-competitive, improving over the strongest automated and handcrafted baselines by 4.26% and 14.92% relative, respectively.

The Hidden Bias of Process Reward Models:PRISM for Rewarding the Right Reasoning

Process Reward Models (PRMs) improve credit assignment for reasoning by providing step-level feedback. However, we identify a hidden bias in PRMs caused by severe imbalance in step-level training data. Standard cross-entropy training amplifies this bias, causing PRMs to overcredit plausible but incorrect steps and produce high false-positive rates. We show that these false positives have an asymmetric downstream effect: false negatives mainly slow exploration, whereas false positives actively steer Best-of-N selection, guided decoding, and policy optimization toward flawed reasoning. This suggests that PRM training should shift from pointwise label fitting to reliable relative comparisons. To address this, we propose PRISM (Precision Ranking for Improved Step Modeling), a policy-aware PRM training framework that learns from contrastive step-level comparisons and hard negatives generated by a temporal lookahead strategy, requiring no new human labels. We further use a difficulty-aware curriculum to optimize the contrastive step margin. Across PRMBench and ProcessBench, PRISM substantially reduces false positives (22% on PRMBench) and improves macro F1 over strong discriminative PRMs. When applied to policy optimization and search tasks, including guided decoding and Best-of-N selection, it consistently improves accuracy (up to 22% for guided decoding and 33% for Best-of-N) and robustness. More broadly, trustworthy process supervision is not just about assigning high rewards, but about rewarding the right reasoning for the right reasons.

RL with Learnable Textual Feedback: A Bilevel Approach

Reinforcement learning with verifiable rewards can improve LLM reasoning, but learning remains sample-inefficient when terminal rewards are sparse. This has motivated a growing line of work on RL with textual feedback, where a critic model generates natural language feedback to guide a reasoning model (the actor), augmenting scalar rewards with richer learning signals. However, existing methods typically treat feedback as fixed or auxiliary, which misses a key property: feedback should not merely be correct, but should improve the policy (actor model) when provided in context. This motivates a paradigm of learnable textual feedback for RL. Yet the learnability and usefulness of feedback depend on the policy's ability to learn from it, making RL with learnable feedback an inherently bilevel problem. We formalize this coupling as a Stackelberg bilevel program and derive Bilevel Natural Language Actor-Critic (Bi-NAC), which jointly trains a critic to generate reward-improving feedback and an actor to exploit it. Across MATH-500, MBPP, and GPQA, Bi-NAC improves sample and parameter efficiency over RL and fixed-critic baselines: our 2B model outperforms the 3B GRPO baseline, achieving 46.6% versus 41.4% on MATH-500, while our 6B model surpasses the 7B GRPO baseline, achieving 49.3% versus 43.6% on GPQA.

Agentic Critical Training

Training large language models (LLMs) as autonomous agents often begins with imitation learning, but it only teaches agents what to do without understanding why: agents never contrast successful actions against suboptimal alternatives and thus lack awareness of action quality. Recent approaches attempt to address this by introducing self-reflection supervision derived from contrasts between expert and alternative actions. However, the training paradigm fundamentally remains imitation learning: the model imitates pre-constructed reflection text rather than learning to reason autonomously. We propose Agentic Critical Training (ACT), a reinforcement learning paradigm that trains agents to identify the better action among alternatives. By rewarding whether the model's judgment is correct, ACT drives the model to autonomously develop reasoning about action quality, producing genuine self-reflection rather than imitating it. Across three challenging agent benchmarks, ACT consistently improves agent performance when combined with different post-training methods. It achieves an average improvement of 5.07 points over imitation learning and 4.62 points over reinforcement learning. Compared to approaches that inject reflection capability through knowledge distillation, ACT also demonstrates clear advantages, yielding an average improvement of 2.42 points. Moreover, ACT enables strong out-of-distribution generalization on agentic benchmarks and improves performance on general reasoning benchmarks without any reasoning-specific training data, highlighting the value of our method. These results suggest that ACT is a promising path toward developing more reflective and capable LLM agents.

Why Pass@k Optimization Can Degrade Pass@1: Prompt Interference in LLM Post-training

Pass@k is a widely used performance metric for verifiable large language model tasks, including mathematical reasoning, code generation, and short-answer reasoning. It defines success if any of $k$ independently sampled solutions passes a verifier. This multi-sample inference metric has motivated inference-aware fine-tuning methods that directly optimize pass@$k$. However, prior work reports a recurring trade-off: pass@k improves while pass@1 degrades under such methods. This trade-off is practically important because pass@1 often remains a hard operational constraint due to latency and cost budgets, imperfect verifier coverage, and the need for a reliable single-shot fallback. We study the origin of this trade-off and provide a theoretical characterization of when pass@k policy optimization can reduce pass@1 through gradient conflict induced by prompt interference. We show that pass@$k$ policy gradients can conflict with pass@1 gradients because pass@$k$ optimization implicitly reweights prompts toward low-success prompts; when these prompts are what we term negatively interfering, their upweighting can rotate the pass@k update direction away from the pass@1 direction. We illustrate our theoretical findings with large language model experiments on verifiable mathematical reasoning tasks.

Safety Recovery in Reasoning Models Is Only a Few Early Steering Steps Away

Reinforcement learning (RL) based post-training for explicit chain-of-thought (e.g., GRPO) improves the reasoning ability of multimodal large-scale reasoning models (MLRMs). But recent evidence shows that it can simultaneously degrade safety alignment and increase jailbreak success rates. We propose SafeThink, a lightweight inference-time defense that treats safety recovery as a satisficing constraint rather than a maximization objective. SafeThink monitors the evolving reasoning trace with a safety reward model and conditionally injects an optimized short corrective prefix ("Wait, think safely") only when the safety threshold is violated. In our evaluations across six open-source MLRMs and four jailbreak benchmarks (JailbreakV-28K, Hades, FigStep, and MM-SafetyBench), SafeThink reduces attack success rates by 30-60% (e.g., LlamaV-o1: 63.33% to 5.74% on JailbreakV-28K, R1-Onevision: 69.07% to 5.65% on Hades) while preserving reasoning performance (MathVista accuracy: 65.20% to 65.00%). A key empirical finding from our experiments is that safety recovery is often only a few steering steps away: intervening in the first 1-3 reasoning steps typically suffices to redirect the full generation toward safe completions.

Test-Time Scaling in Diffusion LLMs via Hidden Semi-Autoregressive Experts

Diffusion-based large language models (dLLMs) are trained flexibly to model extreme dependence in the data distribution; however, how to best utilize this information at inference time remains an open problem. In this work, we uncover an interesting property of these models: dLLMs trained on textual data implicitly learn a mixture of semi-autoregressive experts, where different generation orders reveal different specialized behaviors. We show that committing to any single, fixed inference time schedule, a common practice, collapses performance by failing to leverage this latent ensemble. To address this, we introduce HEX (Hidden semiautoregressive EXperts for test-time scaling), a training-free inference method that ensembles across heterogeneous block schedules. By doing a majority vote over diverse block-sized generation paths, HEX robustly avoids failure modes associated with any single fixed schedule. On reasoning benchmarks such as GSM8K, it boosts accuracy by up to 3.56X (from 24.72% to 88.10%), outperforming top-K margin inference and specialized fine-tuned methods like GRPO, without additional training. HEX even yields significant gains on MATH benchmark from 16.40% to 40.00%, scientific reasoning on ARC-C from 54.18% to 87.80%, and TruthfulQA from 28.36% to 57.46%. Our results establish a new paradigm for test-time scaling in diffusion-based LLMs (dLLMs), revealing that the sequence in which masking is performed plays a critical role in determining performance during inference.

MIRA: Towards Mitigating Reward Hacking in Inference-Time Alignment of T2I Diffusion Models

Diffusion models excel at generating images conditioned on text prompts, but the resulting images often do not satisfy user-specific criteria measured by scalar rewards such as Aesthetic Scores. This alignment typically requires fine-tuning, which is computationally demanding. Recently, inference-time alignment via noise optimization has emerged as an efficient alternative, modifying initial input noise to steer the diffusion denoising process towards generating high-reward images. However, this approach suffers from reward hacking, where the model produces images that score highly, yet deviate significantly from the original prompt. We show that noise-space regularization is insufficient and that preventing reward hacking requires an explicit image-space constraint. To this end, we propose MIRA (MItigating Reward hAcking), a training-free, inference-time alignment method. MIRA introduces an image-space, score-based KL surrogate that regularizes the sampling trajectory with a frozen backbone, constraining the output distribution so reward can increase without off-distribution drift (reward hacking). We derive a tractable approximation to KL using diffusion scores. Across SDv1.5 and SDXL, multiple rewards (Aesthetic, HPSv2, PickScore), and public datasets (e.g., Animal-Animal, HPDv2), MIRA achieves >60\% win rate vs. strong baselines while preserving prompt adherence; mechanism plots show reward gains with near-zero drift, whereas DNO drifts as compute increases. We further introduce MIRA-DPO, mapping preference optimization to inference time with a frozen backbone, extending MIRA to non-differentiable rewards without fine-tuning.

Enhancing Diversity in Large Language Models via Determinantal Point Processes

Supervised fine-tuning and reinforcement learning are two popular methods for post-training large language models (LLMs). While improving the model's performance on downstream tasks, they often reduce the model's output diversity, leading to narrow, canonical responses. Existing methods to enhance diversity are limited, either by operating at inference time or by focusing on lexical differences. We propose a novel training method named DQO based on determinantal point processes (DPPs) to jointly optimize LLMs for quality and semantic diversity. Our approach samples and embeds a group of responses for each prompt, then uses the determinant of a kernel-based similarity matrix to measure diversity as the volume spanned by the embeddings of these responses. Experiments across instruction-following, summarization, story generation, and reasoning tasks demonstrate that our method substantially improves semantic diversity without sacrificing model quality.