Post-Training

Unlearning ensures LLM compliance by removing the influence of private or copyrighted training data. However, since LLM models typically undergo post-training compression, like quantization, in practical deployment, it has been observed that the unlearning effect can be substantially weakened, with the forgetting behavior degrading more severely than that of model utility. This paper proposes a quantization-robust unlearning framework that makes forgetting robust to quantization while maintaining overall model utility. We analyze this gap through the lens of loss landscape. Specifically, our analysis reveals a curvature-based criteria that pinpoints sensitive weights in the unlearned model that leads to both non-robust forgetting and reduced utility. We therefore propose sensitivity-guided noisy regularization, which is applied on the sensitive parameters to steer the model convergence towards a smoother minima of uniformly low forget and retain losses. Balancing unlearning and utility, we further propose forget-critical optimization, which updates only forget-critical layers, preserving most of the network to retain useful knowledge. Extensive experiments on the MUSE and TOFU benchmarks across multiple LLM unlearning algorithms show that our approach achieves substantially more quantization-resilient forgetting while maintaining utility.

Reward fine-tuning aims to update a pre-trained flow-based generative model to improve the downstream reward of its generated samples. Existing methods typically formulate this problem as sampling from a reward-tilted distribution, the solution to a KL-regularized reward-maximization problem. Here, we introduce an optimal transport regularizer built directly from the pre-trained drift. Unlike KL reward tilting, the resulting objective transports individual samples toward higher reward rather than reweighting the base distribution. We show that the resulting problem is equivalent to a deterministic optimal control problem on the flow. Given a pre-trained flow map, this equivalence yields a simulation-free reinforcement learning algorithm for fine-tuning generative flows. We call the resulting framework Wasserstein-Tilted Flow Maps (WTF), the first end-to-end fine-tuning recipe native to flow maps. The output is a fine-tuned flow map that retains strong reward-aligned performance at few-step inference budgets without post-hoc distillation. Experiments on ImageNet-256 and text-to-image show that WTF achieves higher reward with comparable or higher diversity than baselines, while requiring up to $280\times$ less training compute. More broadly, we argue that accelerated samplers such as flow maps are essential infrastructure for efficient post-training, and that the dominant KL-regularized formulation is only one of many choices worth revisiting.

Post-training quantization (PTQ) enables efficient deployment of large language models, and PTQ methods are usually optimized and evaluated with generic reconstruction, perplexity, or answer accuracy. But in explanation-critical domains, preserving only the final answer may be insufficient, since users may also inspect generated rationales to judge whether a prediction is trustworthy. We study this issue in medical multiple-choice question answering, where rationales should provide evidence that supports the selected answer. We propose an explanation-aware objective for transformation-based PTQ. Our method builds an offline faithfulness cache from full-precision teacher rationales and uses it during optimization to preserve answer-supporting evidence tokens and evidence-conditioned answer behavior. We instantiate it on OSTQuant under W4A4KV4 quantization and evaluate four 7B--8B medical and instruction-tuned LLMs on MedExQA, MedExpQA, and ChallengeClinicalQA. While a same-calibration OSTQuant baseline preserves task accuracy, it can substantially weaken answer-supporting rationales. Our objective is to preserve the full-precision model's answer-supporting behavior rather than improve gold-label accuracy, and our method better preserves the full-precision model's answer behavior and rationale-to-answer support. These results suggest that PTQ for explanation-critical settings should evaluate preservation of answer-supporting evidence, not only answer accuracy. Code and evaluation scripts are available at https://github.com/dut0817/EAQuant.

Post-training often improves task performance but can degrade confidence calibration, leaving post-trained language models (PoLMs) more overconfident than their corresponding pretrained language models (PLMs). Because task-specific labeled calibration data can be costly or unavailable, the corresponding pretrained PLM provides a natural label-free reference for post-hoc calibration. Prior agreement-gated PLM-referenced calibration fits a scalar temperature using only examples on which the PoLM and its PLM reference agree, excluding disagreement examples because direct alignment can drive the fitted temperature excessively high and induce under-confidence. We revisit this binary treatment. A controlled reintroduction diagnostic reveals a non-monotonic aggregate effect: admitting a moderate fraction of disagreement examples can improve calibration, whereas the benefit diminishes as unit-weight inclusion approaches the full disagreement set. We introduce SupportCal, a label-free post-hoc method that retains agreement examples at unit weight and assigns disagreement examples continuous weights based on the own-base PLM's relative support and corroboration from pretrained references selected from a size-compatible candidate pool. We further characterize when the resulting weighted objective admits a finite optimal temperature. Across MedMCQA and MathQA, SupportCal yields lower ECE than the agreement-only baseline for nearly all evaluated target-model configurations; supplementary TweetEval Sentiment results show the same pattern on a fixed-label classification task.

As scaling pre-training data alone yields diminishing returns, post-training is becoming increasingly important across physical AI domains such as autonomous driving. End-to-end driving policies are pre-trained in open loop with behavior cloning on human demonstrations. However, compounding errors during closed-loop deployment can take the vehicle outside the training data distribution, increasing the risk of safety-critical incidents. Closed-loop post-training can mitigate this risk but requires costly simulation for sensor-based policies. We propose OPTED (on-policy fine-tuning for end-to-end driving) which decouples reinforcement learning from the post-training of the end-to-end policy: a privileged teacher is trained using RL on vectorized inputs (HD-map and bounding boxes). This teacher then provides supervision to the pre-trained student during closed-loop post-training. We apply OPTED to two camera-based models, TransFuser and VaVAM, and fine-tune them in AlpaSim, using neural reconstructions (3DGS) of real driving logs. Driving scores increase by factors of 1.6$\times$ and 9.5$\times$, respectively. In controlled experiments OPTED matches closed-loop performance with approximately three orders of magnitude fewer simulator interactions than direct RL post-training, while staying closer to the human prior. Project page: https://01dami23.github.io/opted/

Large-scale vision-language-action (VLA) models provide powerful priors for robot manipulation, yet adapting them to a specific deployment remains challenging. Supervised fine-tuning (SFT) on task-specific demonstrations provides a step toward deployment, but faces two persistent limitations: static data provide limited coverage of out-of-distribution states, and standard imitation objectives do not distinguish progressing behavior from less useful data. Interactive post-training can address these limitations, but typically requires repeated policy execution and human intervention on a physical robot. We introduce HIL-UMI, a policy-guided Universal Manipulation Interface (UMI) framework for robot-free human-in-the-loop VLA post-training. During handheld UMI demonstrations, HIL-UMI queries the current policy on the same observation stream without executing its predictions. The Energy Score compares the human action trajectory with policy inference and triggers collection when their discrepancy indicates an out-of-distribution region. In a separate feedback loop, low online advantage predictions identify essential segments for refining a progress-based advantage estimator. The updated estimator then guides advantage-conditioned behavioral cloning using a balanced mixture of base demonstrations and new policy data. This design preserves the iterative and policy-aware nature of human-in-the-loop learning while decoupling data collection from robot deployment. Experiments on four real-world tasks spanning long-horizon and precise manipulation show that HIL-UMI achieves consistent improvement over SFT and benefits from both targeted collection and advantage refinement. Moreover, HIL-UMI outperforms HG-DAgger on Clean Up Table with lower per-frame collection time, suggesting a scalable path for VLA post-training across operators and locations.

When aligning frontier models through post-training techniques, it is not possible to directly demonstrate all of the behaviours we want a model to exhibit in all possible deployment environments; our model must generalise outside of the post-training distribution. One proposed solution is alignment midtraining (AMT), which continues pretraining on large volumes of alignment-relevant documents to encourage generalisation in later stages of training. Despite the prominence of AMT as an alignment approach, there is limited public evidence for its effectiveness. To resolve this, we identify several assumptions around midtraining and evaluate them across scale: up to 110 billion-parameter models and 1 billion midtraining tokens. For instance, we study a scenario where post-training data is ambiguous between two possible motivations. We find that midtraining can steer the model's motivation in simple versions of this setting. However, the presence of a tiny fraction of finetuning data which suggests a competing motivation erases the effects of AMT. We also study scenarios in which we want an AI to follow a number of rules, but only demonstrate a subset of them. We find that demonstrations must be present either in midtraining or post-training datasets for these rules to be robustly learned. Based on these and other findings, we do not believe that there is sufficient public evidence for us to confidently state that midtraining can address the core difficulties inherent in aligning powerful AI systems.

Predicting cellular responses to perturbations supports the study of gene function, disease mechanisms, and therapeutic strategies. Despite advances in single-cell perturbation modeling, existing models typically optimize surrogate losses that do not directly reflect the biological criteria used for evaluation, so better data fitting need not yield better biological predictions. To address this mismatch, we introduce \textbf{CellRFT}, a reinforcement fine-tuning framework that uses biological evaluation as direct training feedback. CellRFT uses policy-gradient optimization to learn from non-differentiable evaluations of generated cell populations and integrates multiple biological rewards through hierarchical reward aggregation. Comprehensive experiments demonstrate CellRFT's applicability across different pretrained models and effectiveness in improving perturbation prediction, reveal that optimizing one biological criterion can help or hinder others, and show that complementary rewards can improve criteria beyond those directly optimized, offering a way to probe how biological metrics shape model behavior, with the potential to inform evaluation design. Code will be made available.

Imitation-learned vision--language--action (VLA) foundation models acquire broad manipulation capabilities by scaling robot data across tasks and embodiments, but reliable deployment on a specific downstream task and hardware platform still requires post-training. Dexterous hands make this adaptation particularly difficult: their broad behavioural repertoire and high degree of freedom create a large and structured action space. Three obstacles are central: open-source VLAs do not natively provide an action interface for high-DoF hands; gesture mismatch during human-gated DAgger takeover creates command discontinuities and contaminates corrective trajectories; and reinforcement learning in the raw joint space is sample-inefficient. We present a unified four-step post-training pipeline comprising a learned temporal hand-action codec, supervised fine-tuning, DAgger, and real-world residual reinforcement learning. The codec adapts a pretrained VLA to absolute dexterous-hand commands. Buffered rollback, pose alignment, and smooth command blending enable continuous, task-relevant DAgger corrections, while latent residual RL confines exploration to coordinated hand motions captured by the codec. We evaluate the pipeline on five diverse real-world tasks spanning bimanual transfer, in-hand reorientation, and tool use. Within the reported post-training budgets, the resulting policies achieve 100% success on every evaluated task over 20 trials per task. These results provide a practical path for adapting VLA foundation models to reliable real-world dexterous manipulation.

Compositional reasoning is critical for real-world problem solving: since training data is necessarily limited, models must generalize by composing learned skills in new ways. While post-training methods such as reinforcement learning (RL) have substantially improved the reasoning abilities of language models (LMs), their effects on compositional reasoning remain less well understood. We propose a dependency-graph framework to formalize compositional reasoning, yielding three levels of compositionality with increasing complexity. Empirically, we instantiate this framework with data-structure tasks, which provide deterministic reward computation and clear compositional structure. We find a consistent decomposed-to-composed asymmetry: decomposed-skill training does not reliably transfer to composed tasks, whereas composed-task training transfers more readily back to decomposed tasks. We provide theoretical explanation for this asymmetry, and further evaluate compositional generalization under length extrapolation, structural distribution shift, and transfer to tasks requiring unseen skills. Finally, we present a pilot study on real-world tool-calling benchmarks, showing preliminary evidence that the decomposed-to-composed asymmetry can extend to practical settings.

Direct preference alignment methods are widely used to align large language models (LLMs) with human preferences because of their computational and memory efficiency. However, likelihood displacement motivates alternative ways to extract information from preference pairs with small likelihood margins. In this paper, we propose and analyze Comparison-based Preference Optimization (ComPO), a zeroth-order alignment method based on comparison oracles. ComPO extracts directional information from these pairs without directly optimizing a differentiable preference loss on them. We establish a convergence guarantee for its basic offline scheme under smoothness, gradient sparsity, and compatibility between the oracle and a latent objective. We further introduce online ComPO, which retains the offline comparison mechanism and uses unlabeled policy generations for reverse-KL control relative to a reference policy. Following the coverage perspective of preference fine-tuning, we establish a performance guarantee for a basic constrained scheme under local coverage and in-distribution pairwise reward accuracy. Experiments on Mistral, Llama, Gemma-2, Qwen3, and Gemma-3 models demonstrate improvements over existing direct alignment methods, including length-controlled win rates, with pair-level diagnostics providing evidence consistent with mitigating likelihood displacement.

Reinforcement learning fine-tuning on top of large, pretrained Vision-Language-Action (VLA) models offers promise for highly reliable robot deployment. However, because of their scale, modern VLA models suffer from high inference latency, so the observation used to select an action is often stale by execution time, creating a distribution shift that can substantially degrade reliability and performance. Prior work has explored asynchronous policy execution to reduce the effect of latency, but these methods are mostly built on imitation learning and offer no mechanism for moving beyond the training distribution toward higher reliability. We close this gap by enabling RL fine-tuning that meets the real-time control requirements of dynamic real-world manipulation. Our approach builds on EXPO-FT, a framework for sample-efficient, reliable VLA fine-tuning with reinforcement learning, and decouples slow, expressive action generation from fast, reactive action edits: a large pretrained VLA proposes action chunks using its strong behavior prior, while a lightweight edit policy performs fast, reactive decision-making by editing actions in response to changes in state, conditioned on the latest observation. We instantiate this as Real-Time EXPO-FT, an RL framework for finetuning real-time VLA policies. On the Kinetix benchmark, Real-Time EXPO-FT enables a delayed policy to achieve the best performance among delayed and non-delayed methods in 10 out of 10 environments. On four dynamic real-world tasks, robot object passing, ball balancing, table soccer kicking, and dynamic object picking, with online robot data capped at 10 minutes, Real-Time EXPO-FT improves average policy performance from 42% to 97%, all without human intervention, demonstrating rapid, sample-efficient adaptation to challenging real-world dynamics. Website: https://pd-perry.github.io/real-time-expo-ft

State space models (SSMs), particularly Mamba, have emerged as efficient alternatives to attention-based architectures and have been extended to vision through ViM, VMamba, and Visual State Space Duality (VSSD). Yet the low-bit post-training quantization (PTQ) behavior of VSSD remains insufficiently understood. A weight-activation split on VSSD-Tiny identifies activation quantization as the dominant low-bit bottleneck, while representative inputs to selected VSSD-backbone linear layers exhibit strong channel-wise magnitude variation and token-localized extremes. We propose the Channel-wise Token-balanced Output-Aware Clipping (CTOAC) method, which learns per-input-channel clipping bounds by minimizing a token-balanced reconstruction loss on the corresponding linear outputs. Only the selected linear layers and their input activations are quantized; other backbone operations retain their original precision. Across VSSD-Tiny, VSSD-Small, and VSSD-Base, the proposed CTOAC method retains ImageNet-1K accuracy and remains substantially more robust than the evaluated baselines at more aggressive precision settings. Applying the same quantization scope to VSSD backbones on COCO and ADE20K preserves strong object detection, instance segmentation, and semantic segmentation performance. An optimized RTX 4090 deployment configuration achieves up to 1.42x end-to-end speedup over FP32.

Continual post-training of large multimodal models should add new capabilities while preserving those from pre-training, and the two goals pull in opposite directions. SFT gives explicit target supervision that learns a task from near-zero accuracy, but its off-policy targets move the model far enough to cause forgetting; on-policy methods such as RLVR and self-distillation preserve policy proximity yet supply little signal when the policy cannot yet solve the task. We introduce ReDraft (Reference-Driven Revision and Fine-Tuning), which obtains both from the model's own failures: using an expert response only as a reference, it has the model revise its own incorrect rollout, keeps the revision only if a verifier accepts it, and fine-tunes on what survives. Each retained target is therefore explicit, yet still close to the current policy. Across Counting, Clock Reading, and Jigsaw on Qwen2.5-VL-3B/7B, two of them with near-zero accuracy, ReDraft gains 56.9 points on the target task against SFT's 52.9 while cutting prior-task loss from 16.6 to 1.5 points (11.3x less forgetting), and improves on OPSD along both axes (19.3 gain, 6.2 loss). Data- and parameter-space analyses match the design: revised targets are more probable under the base model, and the updates they induce stay compact and follow SFT's direction more closely than OPSD's. Together, these results show that revising the model's own rollout rather than directly imitating an expert trajectory can reconcile cold-start acquisition with prior-capability retention.

Continued pretraining (CPT) with data augmentation such as paraphrasing can store inside a large language model (LLM) the knowledge of a small source corpus. The stored knowledge, however, is not always retrieved correctly. We study the eliciting side rather than the storing side: we use preference optimization, which learns from pairs of a preferred (chosen) and a dispreferred (rejected) response, so that the model elicits its stored knowledge more accurately. One proposed approach takes the model's own erroneous response as rejected and the gold answer as chosen, so as to suppress the error. When the target knowledge is partially known, however, most of these rejected responses are factually correct. Using direct preference optimization (DPO) then pushes down rejected responses that contain correct knowledge and differ from the chosen answer only in style, such as length and wording. We propose style-debiased DPO (SD-DPO), which scores whether the rejected response of each pair is factually correct, inverts the preference of such pairs, and weights them so that the learning signal due to differences in style cancels out as a whole. We first test whether, on top of EntiGraph, a representative storing-side method that runs CPT on text synthesized from the corpus, our method adds accuracy efficiently. On QuALITY, the reading-comprehension QA benchmark on which EntiGraph was evaluated, SD-DPO exceeds a baseline we CPT on EntiGraph's synthetic data from the same base model and evaluate with the same procedure. The training tokens this requires are a few dozen times fewer than the additional CPT needed for the same gain. For knowledge updating, the main goal of this work, we use AToKE, a knowledge-editing benchmark for facts that change over time. There, SD-DPO reaches an overall accuracy of 0.982 and answers with the new or the old fact according to the queried period.

Weight-only post-training quantization is the cheapest way to shrink a retrieval embedder, and the received advice for applying it -- protect the embedding table, allocate bits by module sensitivity, prefer a ranking-aware objective over weight reconstruction -- was carried into LLM quantization largely intact. We test that advice on retrieval embedders directly, quantizing five checkpoints from four architecture families across a grid of bit widths and group sizes, and isolating the embedding, attention and feed-forward blocks at each width. Every heuristic fails to transfer as stated. The embedding table never emerges as the dominant isolated protection priority in any family, despite being the largest tensor in several of them. Module sensitivity does not survive as a transferable ordering: at INT4/g16 the spread between modules is too small to allocate against, at INT3 the ordering becomes family-dependent and joint damage stops being the sum of its parts, and at INT2 comparable reconstruction error accompanies retention ranging from 1.3 to 65.9 percent of full precision. A cheap reconstruction proxy is useful for screening uniform bit widths but substantially less reliable for choosing which tensors to protect; its apparent strength across the whole grid is a range-extension artifact. A distilled 109M student at INT3 holds 78.04 NDCG@10 in 68.4 MB and dominates the extreme-PTQ arm of its own 0.6B teacher, 297.9 MB at 64.46, on both size and quality -- but only inside the task it was distilled for. Sizes are byte counts of files that exist rather than arithmetic estimates, and the measurement repository carries the byte provenance for every one of them.

Machine translation systems are periodically upgraded to stronger models, but the available preference signal is human post-edits of an older system's outputs, which the newer model may already surpass. Moreover, collecting fresh post-edits for every new model is prohibitively expensive. We call this the Stale Preference problem. Standard DPO can fail in this setting: it may increase the likelihood of inferior post-edits, erode the model's existing quality, and fail to provide the per-token control needed to correct localized errors. We introduce StalePO, an objective derived from three requirements this regime imposes. Likelihood movement must be downward on both responses, the policy must be anchored to its own base response, and the KL constraint must apply at the token level. These requirements are jointly necessary. In ablations, each mechanism in isolation leaves the model's performance indistinguishable from the base model, and only their combination converts stale feedback into gains. On English-to-Hindi and English-to-Turkish localization data, StalePO improves the fraction of segments passing all LLM-as-judge MQM quality checks by 14.9 and 4.6 percentage points, respectively, with gains concentrated on style and fluency. A human evaluation under the same framework confirms these gains on English-to-Hindi, raising the fraction of segments passing all seven human checks by 13.8 percentage points.

Large language models (LLMs) often learn both desirable and undesirable properties during post-training. We study whether midtraining, an earlier training stage, can shape which of these properties later generalise. We introduce Inoculation Midtraining, a technique that teaches a base model that unsafe behaviour belongs to a designated <quarantine_token> context, as indicated by the <quarantine_token> neologism (a new token) introduced during midtraining, and then post-trains the model on unsafe data within that context. We then evaluate the model outside the context, with the <quarantine_token> neologism excluded from the system prompt. Across supervised fine-tuning and reinforcement learning post-training regimes, we find that Inoculation Midtraining can reduce misalignment while preserving the transfer of benign data properties (e.g., speaking in German or Shakespearean prose). However, our approach does not outperform standard Inoculation Prompting, is sensitive to training configuration, and produces a leaky boundary that nearby contextual cues can reactivate. These results show that inoculation with a learned association introduced via midtraining can shape selective generalisation. Still, more work is needed before this approach can become a load-bearing component in a developer's safety framework.

Post-training with reinforcement learning (RL) is a critical phase in the development of code-generating large language models (LLMs), as it ensures adherence to instructions and the production of functionally correct code. This process typically requires computationally intensive code sample generation from Transformer-based LLMs and substantial GPU-CPU communication for sequence verification. To address these computational challenges, this work examines whether RL-based post-training can be performed entirely offline by leveraging existing datasets rather than generating new samples. The findings indicate that, with only a few hours of training, zero-shot code generation performance of LLMs can be substantially improved without online sampling. Additionally, offline RL produces performance gains across models ranging from 0.5B to 7B parameters, although the extent of improvement varies among model families.

Electronic health record (EHR) foundation models trained on longitudinal patient trajectories have demonstrated strong performance across diverse clinical prediction tasks. However, their clinical reasoning capabilities remain constrained by next-token prediction on limited and incomplete EHR data. To address this, we propose a reinforcement learning (RL) fine-tuning framework that treats EHR foundation models as generative policies over patient trajectories. We formulate common clinical prediction problems (e.g., hospital readmission) as event-conditioned, time-windowed reasoning tasks. We then design time-aware, rollout-sensitive rewards to account for finite rollout lengths and temporally inconclusive outcomes. We find that RL fine-tuning consistently improves over pre-trained backbones and strong baselines. Notably, it enables smaller models to surpass larger pre-trained models in data-limited regimes and induces positive transfer across tasks. Further analysis shows that RL fine-tuned models generate trajectories with stronger structural and semantic alignment to ground truth and greater downstream utility.

Dexterous manipulation requires coordinated multi-finger control and effective tactile feedback, yet learning these capabilities remains challenging due to the lack of large-scale real-world data and the difficulty of extracting effective representations from sparse tactile signals. We build a robot platform and teleoperation system to collect a 200-hour bimanual dexterous manipulation dataset with synchronized visual, tactile, and language annotations, comprising 10,576 trajectories across 65 tasks, 69.5% of which involve dexterous multi-finger manipulation. We further propose STAR, an integrated training recipe for vision-tactile-language-action (VTLA) models that addresses the spatial, temporal, and informational sparsity of tactile signals through visual-tactile joint pre-training, sparse-global tactile token representation, and sparse future tactile prediction. Trained on this dataset, STAR achieves a 61% average success rate across four real-world tasks with 100 post-training trajectories per task, demonstrating dexterous performance under task-specific post-training.

A notable byproduct of LLM alignment training is mode collapse: the progressive loss of output diversity that narrows a model's expressivity at inference time. This degradation is especially limiting for applications requiring open-ended exploration and pluralistic perspectives, such as scientific ideation and creative writing. We present MoDA (Mode-conditioned Diversity Alignment), an online post-training RL algorithm that jointly optimizes generation quality and diversity, inspired by the coordination perspective in multi-agent reinforcement learning (MARL). MoDA trains a single shared LLM policy conditioned on abstract numbered roles, where each role acts as an agent competing to produce outputs distinct from the others. This formulation encourages mode-conditioned agents to explore complementary regions of the high-quality output space without requiring hand-crafted personas or architectural modifications. MoDA employs a prompt-adaptive quality gating mechanism that calibrates a reference quality threshold and grants diversity rewards only to responses that meet the threshold, preventing reward-hacking behaviors that compromise response quality. To study quality-diversity tradeoffs, we evaluate MoDA on a comprehensive suite of benchmarks spanning seven general capability tasks and four domain-specific diversity tasks in scientific ideation and creative writing. MoDA improves SBERT diversity by 265% on the Infinite-Chat held-out prompts, while increasing average general capability pass@1 by 10.3% over the Qwen3-8B baseline. Compared with the strongest DivPO baseline, MoDA improves SBERT diversity from 0.274 to 0.482 (+75.9%) and E-Vendi from 2.86 to 4.4 (+53.8%), while improving average general capability pass@1 by 7.0%. Overall, MoDA provides a drop-in alternative to standard post-training methods that preserves and expands the model's expressive output space while improving quality.

A core goal of efficient reasoning is to improve the accuracy-efficiency frontier. However, jointly improving reasoning accuracy and inference efficiency can be challenging, as the two objectives can favor different reasoning behaviors. Independently post-trained models already offer distinct strengths in accuracy and efficiency. We introduce Lightning Weave, a post-training framework that extracts and composes these independently learned capabilities in a single student through on-policy distillation. Each acquired capability is represented by the policy shift from the model before post-training to the resulting specialist. Lightning Weave combines aligned log-ratio shifts at shared student token states and uses Tilted-Target DOPD to convert the cached signals into a stable learning target. Each anchor pair scores the cached trajectories once, enabling subsequent student training without serving multiple live anchor models concurrently. Across diverse student models and benchmarks in mathematics and code, Lightning Weave substantially improves upon the base students and achieves a state-of-the-art accuracy-efficiency frontier. On Qwen3.5-4B, it raises HMMT 2025 accuracy from 59.2% to 64.0% with 10.7% fewer response tokens, and LiveCodeBench v5 accuracy from 41.7% to 54.2% with 9.6% fewer response tokens. Adjusting the relative strengths of the anchor signals yields a strong empirical accuracy-efficiency Pareto frontier. These results establish Lightning Weave as a new practical route to efficient reasoning through capability composition. Code is released at https://github.com/jet-ai-projects/Lightning-Weave.

Large language model serving costs scale directly with output sequence length, yet standard preference alignment often inflates response verbosity without improving utility. We study whether the parameterization of post-training updates affects generation length: low-rank subspaces alter sequence length without modifying the alignment loss. We present LOCUS, a method that selects a task-aware low-rank adaptation subspace to minimize output-token cost subject to a utility constraint. Within this subspace, post-training retains the native preference objective with a frozen backbone. Across Anthropic HH-RLHF dialogue preferences, we evaluate two $\sim$3B decoder backbones, Pythia-2.8B and Qwen2.5-3B, against protocol-matched full-parameter DPO and DrDPO branches and the released SamPO checkpoint. LOCUS reduces continuation length by up to 39.84% on Pythia-2.8B and by 14.87--17.58% on Qwen2.5-3B while updating only 0.24--0.28% of model parameters, with no material change in the internal preference diagnostic.

Audiobook narration, conversational agents, and audiovisual dubbing require speech that conveys changing emotions and adapts its pacing within a single utterance. But most existing TTS systems typically rely on utterance-level style conditioning, making such fine-grained control difficult to achieve. In light of this, and inspired by the success of post-training in large language models, we propose a unified post-training framework that equips pretrained text-to-speech models with natural-language control over segment-level emotion and duration. Supervised fine-tuning establishes instruction-conditioned speech generation, while reinforcement learning with group relative policy optimization refines control accuracy using emotion and duration rewards alongside content and speaker preservation objectives. By reusing the pretrained architecture, our approach avoids additional inference-time control modules. Experiments demonstrate significantly improved fine-grained controllability while maintaining speech intelligibility and speaker identity, highlighting post-training as a practical approach to extending existing speech synthesis models.

Ultra-low-bit language models promise reductions in storage and memory traffic, but a nominal "1.58-bit" label does not specify the deployed representation or its execution cost. We study a scale-up of an aggressive post-training conversion pipeline from Qwen3-4B to Qwen3-8B. The conversion uses KOTMS rotation, E2M-ATQ adaptive ternarisation, and GPTQ-style error compensation in a weight-only A16 configuration. We do not claim these algorithms as new. Our contribution is the end-to-end scale-up characterisation: an external reproduction gate, matched 4B/8B capability analysis, cross-corpus perplexity, effective-bit accounting, lossless lattice-aware packing, and direct packed execution. The 8B model reaches a three-corpus perplexity ratio of 1.361x, with WikiText-2, C4, and PTB ratios of 1.318x, 1.393x, and 1.371x. On eight zero-shot tasks at n = 500, mean accuracy is 64.6% versus 72.4% for FP16, corresponding to 78.5% chance-corrected retention and a 7.8-point absolute cost. The matched 4B run retains 69.6%, yielding an 8.9-point 8B advantage. The packed checkpoint is 8.24 GiB and preserves the recorded perplexity to measurement precision. Direct packed execution reaches 15.52 tokens/s in 7.35 GiB, while a preliminary packed GEMV remains slower than FP16 cuBLAS. The result is a validated scale-up baseline: model size improves robustness to aggressive post-training discretisation, actual serialisation is solved for the measured artefact, and direct execution is feasible, while broader seeds, calibration distributions, and kernel optimisation remain open.

Financial text, textbooks, and question-answer pairs are abundant, but only a small fraction is directly usable for reasoning-focused post-training. Existing QA pairs often lack explicit reasoning, sufficient context, or reliably verifiable answers, while textbooks must first be transformed into synthetic training examples. We present a data-centric pipeline that constructs complementary corpora by mining open-source reasoning traces, distilling financial instruction data, and generating knowledge-graph-guided question-answer pairs from financial educational material. After semantic deduplication, three lightweight sequence classifiers select finance-relevant examples, reject under-specified questions, and identify tasks suitable for reinforcement learning with compact rule-based verifiers. For model adaptation, we study supervised fine-tuning and reinforcement learning, while self-distilled fine-tuning and post-training model merging are used to prevent the loss of financial capabilities already present in the starting model. We evaluate the adapted language models using FINESSE-Bench, reporting aggregate performance and changes relative to their starting checkpoints. Across the selected comparisons, ordinary SFT reduces FINESSE-Bench accuracy by 3.2-4.0 percentage points, whereas self-distilled SFT improves over the corresponding starting models by 1.0-2.8 points. Equal-weight merging recovers 3.0 points over its SFT parent and finishes 0.9 points above the original model; GRPO on hard tasks adds 0.4 points after self-distilled SFT or 3.0 points when applied directly to verifiable tasks. These results show that retention-aware adaptation can improve financial reasoning without the regressions observed after ordinary SFT.

LLM agents for sequential decision tasks are often post-trained with trajectory-level outcome labels, but such labels provide little supervision for preserving multiple successful branches from the same decision state. We study this problem as successful strategy coverage: how broadly a model realizes distinct successful strategies under a fixed rollout budget. We present Direct Diversity Optimization (DDO), an offline post-training method that combines Divergence-Tree Collection (DTC) with the Reference-Relative Target-Odds Objective (RTO). DTC constructs state-aligned branch sets rooted at shared decision states, and RTO trains the model to match reference-relative targets over successful alternatives. DDO achieves the strongest task success and successful strategy coverage among the compared post-training methods across BabyAI, BabaIsAI, and WebShop. It also achieves the highest recovery rate after local action replacement and higher task success and coverage than successful-only imitation and decoding-time diversification controls.

Task vectors enable post-training model editing by identifying semantically meaningful directions in weight space, typically computed as the difference between a fine-tuned model and its pretrained initialization. However, this reliance on fine-tuning makes discovering such directions costly and limits the practicality of post-training model editing. To address this limitation, we introduce Training-Free Task Vectors (TFTVs), a novel method to compute task-vector-like directions without requiring fine-tuning. Our method maps activation steering vectors to rank-one weight-space edits using only forward-pass statistics, while satisfying arithmetic properties that directly support learning via addition, forgetting via subtraction, and the composition of multiple edits. Empirically, we evaluate TFTVs on large language model behavioral control tasks and show that they consistently amplify, suppress, and compose target behaviors while preserving general knowledge and problem-solving skills. We also validate our method against other editing and steering baselines, experimentally demonstrating that TFTVs achieve stronger trait control with better or competitive utility preservation. We hope our work opens new directions for the community in post-training model editing and broader training-free model control. Code is available on the project website: tftv-llm.github.io.

Weak-to-strong generalization asks whether stronger models can learn from weaker supervisors and surpass them. This question is particularly important for successive model generations and multi-domain consolidation, where repeating frontier-scale post-training from scratch can be prohibitively expensive. Yet conventional distillation treats the weak teacher as an optimization target, potentially imposing its capacity ceiling on the student. We introduce On-Policy Reverse Distillation (OPRD), which evaluates the teacher's policy shift relative to its reference policy on student rollouts and amplifies the component of the student's verifier-driven policy gradient along that direction. By rescaling only verifier-supported updates, OPRD preserves the stationary points of policy optimization while accelerating learning beyond the teacher. In both successive model transfer and multi-teacher distillation, OPRD achieves higher performance with fewer student updates than existing RL and distillation approaches. Response-style analysis shows that OPRD students remain closer to models trained with verifier-based RL alone than to their weak teachers, suggesting that teacher guidance accelerates rather than redirects the student's own optimization. Results in conventional strong-to-weak distillation further demonstrate that OPRD effectively combines verifier-driven policy optimization with teacher guidance regardless of capacity ordering.

We present Miles v0.1, a full-stack, production-ready system for frontier post-training. Building upon the clean design of slime, Miles designs each stage of the reinforcement-learning (RL) training loop around a single principle: components should be verified, clean, and customizable. With accuracy, efficiency, reliability, and scalability as first-class goals, Miles aims to make frontier-scale RL accessible to researchers and enterprises alike. This report walks through the system end to end: rollout engines built on SGLang, a trainer with a choice of two backends (NVIDIA Megatron-LM and PyTorch FSDP), and three weight-synchronization transports for different deployment topologies. Beyond full-parameter RL, Miles also supports LoRA RL, on-policy distillation, supervised fine-tuning, and true-on-policy rollout-training alignment, and extends the same architecture to diffusion models. We close with an end-to-end case study: fully asynchronous agentic RL on a GLM-5.2 744B-A40B model over terminal-use coding tasks, running on 64 NVIDIA GB300 GPUs with a median step time of 263 seconds over the first 30 measured steps. Miles is open-sourced at https://github.com/radixark/miles, with the project website at https://miles.radixark.com.

Recursive self-improvement (RSI) requires a concrete mechanism through which an AI system observes its capabilities and converts that evidence into the next round of learning. We present NeoHorse-1, a family of agent-native models developed to explore this path through agentic post-training. Our system combines a heterogeneous model pool with intelligent routing, recording the predicted capability demand, selected service tier, and subsequent interaction for each user turn. These records are converted into training examples that preserve interleaved reasoning, tool calls, and harness context, and are admitted through structural validation, six-dimensional semantic evaluation, and subscene-level labeling. Routing signals organize supervised fine-tuning into a three-stage curriculum and extend to routing-guided on-policy distillation, where a teacher supervises student-generated responses under the same progression. Capability-guided allocation then converts evaluation feedback into the next training mixture, closing an evaluation-selection-update loop in which what the system learns to do shapes what it learns from next. Across eleven benchmarks covering harness-based agents, tool use, coding, and instruction following, post-training raises the macro-average from 58.94 to 64.87 at 4B and from 65.60 to 69.04 at 9B, substantially narrowing the aggregate gap between the post-trained 4B model and the 9B base model. NeoHorse-1 provides an initial prototype of this feedback-driven process and a path toward harness-mediated RSI across successive iterations.

Mixture-of-Experts (MoE) pretraining relies on an auxiliary load-balancing loss (LBL) to drive per-expert utilization toward uniformity. Post-training inherits a different situation: the base router already encodes non-uniform expert co-activation structure, which a re-imposed uniformity objective flattens away. We show that downstream performance depends instead on holding this inherited routing softly, a principle we term soft router anchoring, and instantiate it as Router Prior Bias (RPB), a training-time bias that pulls the router logits toward a prior read off the frozen base router while leaving the router itself trainable. On math post-training of Moonlight-16B-A3B, RPB attains 45.77 in-domain accuracy against 31.91 under re-applied LBL and 29.44 under unanchored fine-tuning, and retains more out-of-domain capability than either. The ordering against LBL reproduces on a second model family (Qwen3-30B-A3B-Base), and the advantage over LBL is resolvable on an independently sourced corpus. Anchors defined on the router weights, on its logits, or on its output distribution perform comparably with no consistent ordering, which places the effect in the softness of the constraint rather than in the particular prior RPB supplies. Retained community structure in the expert co-activation graph tracks these gains wherever the base router is non-uniform enough for communities to form, yet enforcing the same prior as a hard assignment preserves that structure while performance falls sharply. Community structure is therefore a footprint of soft anchoring rather than its source, and the practical lesson is that inherited routing should be held softly during post-training, since both flattening it toward uniformity and enforcing it absolutely carry a downstream cost. Our code will be released at https://github.com/naver-ai/rpb.

Large Language Models (LLMs) are increasingly deployed in settings where rare but severe harmful generations can have significant consequences. Existing Risk-Averse RLHF addresses this issue by optimizing Conditional Value-at-Risk (CVaR), but it trains policies for fixed risk levels and therefore cannot adjust the desired degree of risk aversion at inference time. In this paper, we propose risk-conditioned RLHF, a framework that trains a single policy that provides a continuous risk-control interface, enabling users to select different degrees of risk aversion without retraining or deploying multiple risk-specific models. Experiments across multiple benchmarks demonstrate that a single risk-conditioned policy can adapt to different risk levels at inference time, enabling more flexible and risk-aware LLM deployment.

Existing post-training pipelines for coding and terminal agents suffer severe token and control fidelity errors: simplified training environments mismatch production deployments, and offline token reconstruction from agent logs distorts original prompts and conflates policy calls with background model operations. We present a fidelity-aware training coupling framework that retains trainer-side sampling over original prompts, eliminates spurious model calls via a negotiated training protocol, and restricts loss computation to verifiable token spans with closed-failure guarantees. We further propose Certified Divergence Proximal Policy Optimization (C-DPPO), which establishes tight two-sided TV certification bounds, adaptive-K rules, budget-aware sequence guarantees, and error-robust policy masking atop standard DPPO. Evaluated on matched Baize5B and Baize10B models with identical training and test protocols on TMax-100, C-DPPO yields a consistent +3.0-point performance gain over standard DPPO across model scales. Certificate audits validate the reliability and full operational coverage of our certified training pipeline.

Large language models demonstrate increasingly strong reasoning capabilities through effective post-training. Yet, prevailing post-training methods optimize over massive numbers of tokens, implicitly assuming that effective learning must be token-intensive. We revisit this assumption in the on-policy distillation (OPD) setting, which naturally admits dense teacher supervision at every generated token. Using the Qwen3 family, we discover a counter-intuitive phenomenon: reasoning can be effectively incentivized by an extremely small fraction of generated tokens--as few as one or two tokens per reasoning trajectory, corresponding to only 0.05% of all tokens. Surprisingly, this sparse supervision in most cases matches or surpasses full-token training in improving reasoning ability, despite excluding the vast majority of generated tokens from the training objective. This phenomenon is consistently observed across nine teacher--student configurations spanning different model scales on mathematical reasoning tasks, and is further validated on coding reasoning, Llama models and Proximal Policy Optimization (PPO)-based reinforcement learning with verifiable reward (RLVR). Interestingly, such extremely sparse supervision may be closer to the natural learning process: rather than correcting every step word by word, one reflects on a few critical reasoning steps, updates prior understanding, and continues the trial-and-error, avoiding micro-level corrections while remaining remarkably effective. Overall, our results challenge the assumption that effective post-training must be token-intensive and point to a new direction for understanding and designing more efficient post-training algorithms.

How do the methods used to train language models to refuse harmful requests shape how that refusal actually works inside the model? We compare three post-training methods - supervised fine-tuning, reasoning-augmented fine-tuning (training on reasoning chains that justify a safety decision), and preference optimization (ORPO) - across three architecturally distinct models (Llama-3.1-8B, Gemma-2-9B, Qwen3-8B). We find that training method, not just data, reshapes how refusal is computed internally: reasoning-augmented training consistently produces a distinct kind of refusal computation, visible across all three models, while architecture independently shapes internal structure and how reliably refusal can be steered. Most importantly, no method we study achieves all three properties we would want from safe alignment at once: refusal that isn't concentrated in a few fragile components, safety gains that don't cost general capability, and safety behavior correctable through small, targeted edits. We caution against treating current post-training methods as a solved, reliable defense, especially for security-critical use. Code and models are available in https://github.com/hoangcuongnguyen2001/Beyond-Shallow-Alignment.

Ultra-low-bit language models can reduce storage and memory bandwidth, but a nominal "1.58-bit" label does not fully describe the stored representation, retained capability, or runtime behavior. We study an end-to-end post-training conversion of Qwen, an instruction-tuned 4B-parameter model, using KOTMS rotation, E2M-ATQ ternarization, and GPTQ-style error compensation from TWLA. The experiment is weight-only: activations remain at 16-bit precision, so ILA-AMP is omitted. We evaluate effective bit accounting, task capability retention, perplexity, calibration sensitivity, checkpoint composition, and deployment behavior. The final conversion uses 1.641 effective bits per weight for quantized linear weights, with 81.62% of model parameters targeted. Across ten scored capability comparisons, accuracy falls from 64.5% to 54.7%. Degradation is uneven: BoolQ retains 84.6% chance-corrected teacher performance, while ARC-Challenge retains 43.8%. Perplexity rises from 13.639 to 18.748 on WikiText-2, 24.700 to 31.992 on PTB, and 19.831 to 28.966 on C4. A subsequent packing run preserves the ternary planes and scales, reducing reported model size from 8.29 GiB to 3.96 GiB with essentially unchanged perplexity. A separate third-party packing attempt was lossy and is excluded from the primary artifact claim. The packed artifact has not been benchmarked end-to-end for task accuracy or generation throughput. A preliminary Triton GEMV microbenchmark is 4.6x slower than FP16 cuBLAS on one tested shape. We therefore do not claim that compression alone yields faster inference.

Edge vision models are difficult to deploy on resource-constrained hardware, making low-bit post-training quantization (PTQ) attractive. In practice, standard FP32 training often produces heavy-tailed activation distributions whose outliers destabilize activation quantization: preserving the full range wastes quantization bins on rare extremes, while aggressive clipping causes information loss. Existing solutions typically rely on quantization-aware training (QAT), which adds training complexity and bit-width coupling, or advanced PTQ procedures that repair the model after training. We present SCULPT (Statistical Clipping and Uniform Loss for Post-Training), a training-time method that improves PTQ readiness during ordinary FP32 fine-tuning. SCULPT combines a topology-aware activation regularizer that suppresses quantization-hostile skewness and kurtosis with a stable percentile-based clipping mechanism that learns deployment-ready activation bounds. Unlike QAT, SCULPT does not simulate quantization during optimization; unlike post hoc outlier-repair PTQ methods, it does not require runtime activation transformations. The learned clipping bounds can be exported directly into a standard PTQ workflow for low-bit deployment, including INT8 and lower-bit settings such as W4A8.

Data-residency constraints force enterprises to self-host LLMs, but continuous adoption of newer models without decommissioning their predecessors expands the serving fleet, fragmenting a finite GPU pool. We consolidate traffic from over 200 internal applications onto a single model by closing quality gaps identified through production error analysis along three axes: instruction following, function-calling, and internal task distribution. Quality is tracked by offline benchmarks stratified to production traffic and scored by deterministic verifiers or calibrated LLM judges. Rather than optimising all objectives jointly, which introduces cross-domain reward interference, we train a separate GRPO expert per axis and merge them via two-stage SLERP. Each expert's reward exposes a distinct failure mode, namely semantic collapse, over-calling, and verbosity hacking, each requiring a domain-specific fix. In non-reasoning mode the recipe surpasses a ${\sim}7\times$ larger by total parameters baseline on the in-house Arena with 69.6 to 65.8, instruction following with 0.85 to 0.83, and function-calling with 0.79 to 0.77, while lifting general dialogue benchmarks. The model absorbs 50% of platform traffic, 116M requests per month, at a fraction of the serving cost.

Post-training large language models usually applies a single training recipe to all samples, even though the model's own rollouts reveal different sample-level learning states. We propose Self-Routing, a behavior-conditioned post-training framework that uses rollout correctness and confidence to decide how each sample should be optimized. Depending on its behavior state, a sample is routed to GRPO, on-policy self-distillation, regularization, or skipping, allowing training to adapt without external teachers, extra annotations, or additional sampling. Experiments on mathematical reasoning across Qwen3 and Qwen3.5 backbones show that Self-Routing consistently improves over uniform GRPO, uniform OPSD, fixed mixtures, and simpler routing baselines. Further analyses show that the routing distribution changes over training and reduces unnecessary updates on low-signal or already stable samples.

Every supervised fine-tuning run forces the same chain of decisions, such as learning rate, batch size, LoRA or full fine-tuning, how many epochs, which optimiser, and what data to feed the model. Each of these is typically rediscovered from scratch for every new model and dataset. Here we measure them under one instrument: a sweep that varies one lever at a time, and spans dense and mixture-of-experts models in two families (Qwen3 and Llama), on four real-world customer SFT datasets, for both LoRA and full fine-tuning. These datasets give a controlled testbed: each task carries an evaluation built with the customer, and its training data is produced by iterative supervised fine-tuning that refines model outputs until they pass that evaluation, so the supervised target is internally consistent and the task judge we report against is the criterion the data was built to satisfy. We ask how the optimal learning rate and batch size move with model scale, family, and data, and whether one selection rule transfers across them; what LoRA trades against full fine-tuning, and how its rank and alpha set what the adapter can learn; whether validation loss (or other metrics, such as loss landscape flatness) faithfully ranks downstream quality; whether post-training gains scale with model size and data volume, on a model ladder extended through mixtures-of-experts to 235B parameters; how many epochs to train before general instruction-following erodes; and whether a geometry-aware optimiser improves on AdamW. Each recommendation is paired with a measure of its uncertainty.

Language models can ignore prompt evidence when it conflicts with memorized knowledge. Post-training can make models follow such evidence more reliably, but it is unclear whether these gains require new machinery or strengthen machinery already present. We compare nine post-training arms spanning GRPO, SFT, and DPO from one starting checkpoint, with key comparisons extended across scales and families. We estimate a grounding direction from that checkpoint before training. Across five tested GRPO variants, grounding gains are small. For the two variants replicated across seeds, equivalence tests bound their effects below the conflict-SFT gain even as the rewarded metric improves. Conflict-SFT improves grounding moderately, while DPO drives grounding near ceiling on its matched distribution. Conflict-SFT and DPO largely use the same causal attention-head set as the starting model. Subtracting the starting-model direction suppresses both gains, while adding it to the starting model recovers 35% of DPO's gain at a dose passing all stated side-effect checks. After a supervised warm start makes the context answer appear in more rollouts, the same GRPO recipe adds essentially no further grounding gain. In our setting, grounding gains largely depend on machinery already present in the starting model.

Vision-and-Language Navigation (VLN) requires an agent to navigate through unseen 3D environments according to natural-language instructions. Explicit reasoning can improve instruction understanding and semantic grounding, but autoregressive generation at every step accumulates large decision-stage latency over multi-step navigation. We propose VerNav, a verifier-first framework for low-latency LLM-based VLN. The verifier reduces decision-stage latency by replacing per-step autoregressive generation with batched action verification, while an entropy-based adaptive generator is invoked only for uncertain decisions to produce compact state evidence. To further improve navigation performance with the verifier, we introduce a two-stage alignment scheme: (i) VPO improves local action-preference alignment in static verifier training, and (ii) step-level reinforcement fine-tuning provides dense progress rewards over multi-step navigation rollouts during dynamic task execution. Experiments on the Room-to-Room (R2R) benchmark show that the verifier-only decision path of VerNav achieves competitive navigation performance among representative LLM-based VLN agents while reducing average decision-stage LLM latency per step by more than $10\times$ compared with autoregressive methods.

Rubric-based reinforcement learning decomposes open-ended instructions into prompt-specific, flexible rubrics, making it better suited than reinforcement learning with verifiable rewards for post-training LLMs on open-ended tasks. However, static rubrics are inevitably hacked as the policy evolves, and existing dynamic approaches introduce new problems: undirected rubric extraction, unreliable hack detection, and unbounded rubric proliferation. We propose $\textbf{CARE}$ ($\textbf{C}$ontrastive $\textbf{A}$nchor-based $\textbf{R}$ubric $\textbf{E}$volution), which grounds every rubric evolution step in a high-quality anchor response generated by a frontier model conditioned on the prompt and its rubrics. At each training step, CARE contrasts the highest-scoring rollout against the anchor, enabling two complementary mechanisms: an Adaptive branch that reactively repairs reward misspecification; and a Chase branch that proactively converts frontier-level quality gaps into sharper rubrics. Together, the two branches $\textbf{maintain discriminative accuracy in the high-reward region}$---the precise region where reward over-optimization mostly originates. Experiments on WildChecklist-9K with Qwen2.5-7B-Base and Qwen2.5-7B-Instruct show that CARE achieves state-of-the-art performance on Arena-Hard-2.0, InfoBench, and FollowBench, and is the $\textbf{only}$ method whose win rate against GPT-4.1 anchor responses shows sustained improvement throughout 300 training steps; additional results on Llama-3.1-8B-Instruct and Qwen3-8B further indicate that CARE generalizes across model families.

Industrial post-training is a brownfield regime. Teams inherit a deployed checkpoint and must land targeted improvements under fixed compute and mixture budgets without regressing the rest. The maintained artifact is increasingly dataware: behavior governed by a curated post-training mixture, updated via bounded mixture patches rather than clean-slate retraining. From an industrial code-generation improvement effort, we offer a maintainer's perspective on why this work is hard in practice, distilling three recurring challenges, zero-sum mixture design, yield as the binding metric, and end-to-end integration under uncertainty, and arguing that progress depends less on one-off recipes than on an engineering discipline for programming dataware. In our case study, interventions that raised the conversion of teacher distillation into usable training data increased accepted supervision by 2.84 times while using the same solution teacher and four solution attempts per candidate problem. In our primary evaluation, the yield-engineered patch improved CodeForces pass@1 by +2.59 points (+3.11 pass@3) and held-out LiveCodeBench v6 pass@1 by +6.11 (+8.05 pass@3), all statistically significant across 16 stochastic evaluations of each benchmark from one fixed checkpoint per condition, with internal AIME and MATH regression suites within tolerance.

Codec-based text-to-speech (TTS) models make language-model post-training applicable to speech generation, but it remains unclear when learned perceptual predictors can serve as reinforcement learning rewards without losing alignment with human listeners. We study this question with Group Relative Policy Optimization (GRPO) using learned rewards for anime-like speaking style, naturalness, likability, and arousal. To prevent perceptual rewards from being optimized through transcript drift, we introduce a character error rate (CER) zone constraint and compare policy optimization with Best-of-$N$ reranking under the same reward gate. Across single-reward runs, each reward primarily improves its own target metric, showing that subjective predictors are not interchangeable quality surrogates. Multi-rater A/B tests further show uneven human transfer, while a reward-gap analysis separates average transfer from within-axis calibration: signed reward gaps significantly predict listener choices in the pooled analysis, whereas residual CER gaps do not, but per-axis calibration remains heterogeneous. Best-of-8 is a strong human-level baseline and is not clearly worse than GRPO perceptually, suggesting that GRPO should be viewed as amortizing reward-selected behavior into the policy rather than uniformly outperforming reranking. These results support analyzing subjective speech rewards as predictor-axis-base tuples and provide practical diagnostics for selecting rewards before multi-reward speech post-training.

In large-scale e-commerce retrieval, dual-encoder retrievers are op- timized for contrastive similarity, whereas downstream rerankers capture finer-grained relevance preferences; this objective mis- match limits end-to-end retrieval quality. Reinforcement Learning offers a way to use reward-model feedback for retriever adaptation, but we observe that standard policy-gradient updates can degrade embedding geometry, especially when the document index must remain frozen due to industrial constraints. To address this, we propose PAO (Positive-Advantage-Only), a selective RL optimization method. Our analysis reveals that in- discriminate penalization of negative samples (pushing away) in a frozen high-dimensional space disrupts pre-trained semantic man- ifolds. PAO selectively applies gradient updates only to retrieved items with positive advantages, effectively pulling query embed- dings toward high-reward regions while preserving global topo- logical stability. Experiments on both a massive industrial dataset and public benchmarks demonstrate that PAO significantly outper- forms standard RL and distillation baselines.

Reinforcement learning (RL) is used to improve the reasoning abilities of LLMs, while training data span heterogeneous tasks. However, most RL post-training pipelines rely on fixed or manually designed task mixtures, even though task usefulness changes as training progresses. Online curriculum methods often define learnability by update magnitude, ignoring whether the update translates into reward gains, which can misallocate rollout budget toward tasks with large but ineffective updates. We propose PAC, a Progress-Augmented Advantage Curriculum for multi-task RL of LLMs that combines two task-level signals: advantage-derived learnability, which measures the magnitude of the policy update a task can induce, and recent reward gains, which show whether those updates have improved task performance. A Bayesian Thompson Sampling controller uses these signals to allocate rollouts across tasks during GRPO training. We evaluate PAC under two settings: a multi-level reasoning setting and a multi-domain reasoning setting. PAC improves sample efficiency and final performance: it reaches comparable validation scores with fewer rollout steps and achieves higher final averages than random sampling and advantage-based curriculum baselines in both settings. These results show that jointly tracking advantage signals and actual reward gains yields an effective online curriculum for LLM post-training.

Industrial anomaly detection is a critical component of modern manufacturing. Most traditional unsupervised methods rely on modelling normal feature distributions, inherently limiting generalization to unknown categories. To improve generalizability, some recent methods incorporate vision-language models (VLMs) for zero-shot detection via text prompts. However, we observe that reasoning-oriented post-training can cause anomaly discrimination to collapse, with some fine-tuned models performing worse than their base VLMs. Existing methods also provide only textual decisions or coarse boxes, without pixel-level segmentation. A more explicit detection principle comes from human inspection: anomalies are identified by comparing a query image with a defect-free reference. Inspired by this, we propose InspectorGPT, a VLM framework centered on comparative reasoning. Given a normal reference and a query image, InspectorGPT compares them to identify discrepancies and perform multiple inspection tasks with detailed reasoning. We internalize this capability through Chain-of-Thought (CoT) fine-tuning and Group Relative Policy Optimization (GRPO) with tailored, verifiable rewards. We further introduce InspectorGPT-Seg for pixel-level anomaly masks. Segmentation supervision improves anomaly discrimination but weakens semantic reasoning, while joint training fails to balance them. We therefore train the two branches separately and combine them through task-vector fusion. Extensive experiments demonstrate superior multi-dimensional performance and generalization to unseen benchmarks, validating comparative reasoning for comprehensive industrial inspection.

To mitigate the time complexity of generative models, one-step generative models have recently emerged through direct mapping from noise to data in a single forward pass. However, the reward-guided fine-tuning method of one-step generative models remains largely unexplored. To address this, we consider one-step generators from an optimal transport view, investigating Wasserstein Gradient Flow (WGF) for modeling smooth and controlled distributional evolution in probability space. We then propose a novel reward-guided fine-tuning of a one-step generative model via WGF. We derive a practical training method that requires no reward gradients, thereby handling both non-differentiable and differentiable rewards. Moreover, our method provides smooth and stable reward-guided distributional updates while mitigating reward hacking and mode collapse. Experiments on 2D synthetic data, CIFAR-10, and ImageNet 256$\times$256 with diverse rewards, including JPEG (in)compressibility, class probability, Black-and-White and CLIP alignment, show that our method achieves better reward alignment compared to baselines.

Current large language model development relies on massive, often non-permissible datasets, creating a high barrier for researchers committed to open-source and ethically sourced data. We introduce Mimir v1, a 1-billion-parameter language model based on the Hierarchical Reasoning Model (HRM) architecture, that is trained from scratch and delivers highly competitive performance for English and sets a new state of the art for Danish using only permissible post-training data. Trained on a mixture of 161 datasets, Mimir v1 outperforms the original HRM-Text 1B and competes with larger frontier models like Qwen 3.5 4B and Gemma 4 E2B, tested across 20 benchmarks for English, Math & Code and Danish. The model is available on the Hugging Face Hub: https://huggingface.co/danish-foundation-models/DFM-Mimir

Scientific discovery increasingly requires AI systems that can reason over scientific evidence of heterogeneous modalities, interact with scientific tools and environments, and sustain progress across long task horizons. We present Intern-S2-Preview, a series of scientific agentic foundation models designed to support multimodal scientific understanding, reasoning, generation, and long-horizon tasks. The training pipeline begins with scientific multimodal pre-training over rendered scientific documents, interleaved image-text data, and diverse scientific corpora. Starting from the pretrained checkpoint, we apply a unified post-training pipeline consisting of supervised fine-tuning, scalable multi-task reinforcement learning (RL), black- and white-box agentic RL, and on-policy distillation. This pipeline is supported by practical techniques that improve rollout and training stability and efficiency, including partial rollout with off-policy correction, adaptive length regularization, online speculative decoding, robust multi-task optimization, and trace-aware experience assembly for agentic tasks. At the architecture level, Intern-S2-Preview-397B extends time series modelling from efficient long-sequence understanding to numerical forecasting, while Memory Decoder is studied as a separate memory-augmented path for rapid scientific specialization without modifying the frozen 397B backbone. Evaluations across scientific, multimodal, agentic, and general-purpose benchmarks show that Intern-S2-Preview-397B achieves competitive or leading results in multiple settings. The time series modules improve scientific signal understanding and forecasting on SciTS, while the separate Intern-MemDec-4B extension improves the Biology-Instructions average score from 56.92 to 60.32 without modifying the frozen 397B backbone.

Reinforcement learning against rubrics, lists of criteria graded by an LLM judge, has become a standard way to post-train language models on tasks with no deterministic answer. The rubric, however, is a fixed proxy for quality, never a complete description of it, and a policy trained against it long enough will learn to exploit the difference. We measure this directly. Training Qwen3-8B with Group Relative Policy Optimization (GRPO) on medical and science rubrics and grading out-of-distribution (OOD) benchmarks with both the training judge and a stronger gold judge, we find that the two scores diverge during training. The training judge's score keeps climbing while the gold judge's score peaks and then falls, by 3 points on HealthBench-Hard and by 22 points on ResearchQA. A judge with a fixed bias would shift the gold curve by a constant, not send it down while the training score rises, so the divergence is reward hacking, not judge noise. We propose Rubric Dropout, a one-line fix borrowed from neuron dropout. At every step, we randomly drop a subset of the rubric's criteria before computing the reward, so the policy never optimizes the same rubric twice. The dropped subset is shared across each rollout group, so GRPO's group-relative advantages stay comparable, and evaluation always uses the full rubric. Comparing no dropout against dropout at 30% and 50% on both benchmark pairs, dropout raises the OOD gold score at every matched checkpoint (+1 to +2 points on HealthBench-Hard, +6 to +7 points on ResearchQA), lowers the two hacking measures we track, and costs nothing in domain. Sweeping the dropout fraction shows a broad 30-50% sweet spot, while the natural alternative, reweighting criteria by how useful they are to training, performs worse than no intervention at all in our setting.

Reinforcement Fine-Tuning (RFT) has enabled medical Multimodal Large Language Models (MLLMs) to produce Chain-of-Thought (CoT) reasoning for visual question answering, yet these models suffer from $\textit{confidence miscalibration}$---a systematic gap between expressed certainty and actual diagnostic accuracy that undermines clinical trust. We propose $\textbf{CARE}$, a $\textbf{C}$onfidence-$\textbf{A}$ware medical $\textbf{RE}$asoning framework that jointly optimizes accuracy and calibration through a dual-stage pipeline. First, a scalable Medical-CoT synthesis provides structured cold-start data for Supervised Fine-Tuning. Second, Group Relative Policy Optimization (GRPO) with a novel $\textbf{Confidence-Aware Reward (CAR)}$ mechanism ties the model's confidence to diagnostic correctness within the reward signal. Across three Medical VQA benchmarks, $\textbf{CARE}$ achieves the highest diagnostic accuracy while obtaining the lowest Expected Calibration Error and Hallucination Rate, establishing a foundation for trustworthy clinical decision support. Our code is available at https://github.com/anotherbricki/CARE.

Reinforcement learning (RL) post-training of large language models (LLMs) is computationally intensive and involves complex system pipelines with substantial debugging overhead. In practice, factors such as framework adaptation, numerical precision, and operator implementation can cause failures, including gradient overflow and loss divergence. Reproducing such failures directly on large models requires considerable time and computational resources. This paper systematically analyzes failures encountered during large-scale RL training on the Huawei Ascend platform, summarizes representative failure types, and identifies three model-side factors relevant to fault reproduction. Based on these factors, we propose a proxy-model construction method for low-cost fault investigation and auxiliary diagnosis. It employs structure-preserving, clustering-based expert pruning to select representative experts while retaining the model's backbone architecture, routing mechanism, and basic task capabilities. Our experimental results show that the proxy models reduce accelerator requirements by 50%-87.5% and achieve up to a 33.3x reduction in per-step NPU-hour cost, while preserving major training dynamics and reproducing fault responses consistent with the original models. Overall, the proxy models can serve as low-cost surrogates for fault reproduction, targeted validation, and auxiliary diagnosis in RL post-training.

We study reference-free post-training for multilingual machine translation with open large language models. Starting from the supervised-finetuned MiLMMT-46-v0.1 models, we apply Group Relative Policy Optimization (GRPO) with a reward that averages two reference-free quality estimation models and is gated by language identification. We then linearly interpolate the supervised fine-tuning (SFT) and reinforcement learning (RL) model checkpoints to obtain MiLMMT-46-v1.0. Across 46 languages, the resulting models consistently improve translation quality over their SFT counterparts, outperform strong recent open baselines, including Seed-X, HY-MT2, and TranslateGemma, and achieve leading reference-free scores against evaluated proprietary systems such as Google Translate, Gemini 3 Pro, and GPT-5. We further investigate on-policy distillation (OPD) and find that it reaches, but does not surpass, the quality frontier achieved by RL with checkpoint interpolation. We release the models and code to facilitate future research.

Large language models (LLMs) are trained on massive and largely undisclosed corpora that may contain copyrighted or privacy-sensitive content. Data contamination detection (DCD) therefore aims to determine whether a given text is a member of the pre-training corpus of a target LLM. Recent state-of-the-art DCD methods follow a feature-based paradigm that derives membership features from the input text and the corresponding model output. However, most modern LLMs undergo post-training, such as instruction tuning, preference optimization, and reasoning-oriented training, which can alter model outputs and shift the corresponding membership features, thereby reducing the separability between members and non-members. To address this problem, we propose CalibDCD, a broadly applicable calibration framework for feature-based DCD methods, comprising (1) Multi-View Shift Detection, which identifies recurring feature shifts associated with post-training, and (2) Bounded Feature Correction, which selectively mitigates their influence on membership prediction. Specifically, Multi-View Shift Detection evaluates controlled prompt variants on known non-member texts and consolidates the most informative views to identify recurring feature shifts. Bounded Feature Correction selectively adjusts feature components aligned with the detected shifts and controls the correction extent to preserve useful detection information. Experiments show that CalibDCD consistently improves existing feature-based detectors, with gains of up to 7.0% in AUC and 15.0% in TPR@5%FPR.

Feedback signals used to train Large Language Models (LLMs) are the primary driver of their behavior and our main lever for instilling alignment with human values and objectives. However, a key limitation of current post-training methods is the inability of human annotators and automated reward functions to faithfully capture the feedback we would like to give. We introduce Evaluation-Conditioned Training (ECT), a post-training framework that uses natural language to condition each training sample on the fidelity of the feedback we provide and then elicits the desired behavior by conditioning the LLM on a high-fidelity monitor in deployment. ECT is aimed at improving performance under imperfect feedback and works as an add-on to existing algorithms such as SFT and PPO. We first provide a conceptual framework for ECT and discuss its potential to address persistent sources of reward mis-specification. Then we motivate ECT in the context of the eliciting latent knowledge (ELK) problem. Finally, we evaluate ECT on two proof-of-concept experiments: increasing even-handedness in news article generation and reducing sycophancy on an arithmetic task. In each setting, we utilize imperfect feedback, rewarding bias and agreement with the user, respectively. In both settings, ECT improves the targeted behavior relative to direct training.

Large language model evaluations typically focus on performance under nominal conditions, creating an illusion of capability where models comfortably walk a narrow, highly optimized generation corridor. In real-world deployments, however, complex system prompts, safety guardrails, and structural constraints continuously force models off this nominal path, driving a divergence between benchmark scores and deployment performance. To address this issue, we introduce Decoding-Level Taboo, a zero-prompt diagnostic stress test that intervenes directly in logit space at runtime, forcing models out of their nominal paths. By dynamically masking primary candidate tokens at word boundaries, Taboo forces machine circumlocution. Evaluating Taboo across several open-weight model families reveals that off-path robustness is heavily influenced by both parameter scale and post-training instruction alignment, with robustness generally improving with model size and alignment. Beyond the results presented in this paper, Taboo provides a novel primitive for generating diverse synthetic datasets, stress-testing runtime safety guardrails, and auditing model reliability prior to real-world deployment.