Supervised Fine-Tuning

Semi-supervised federated learning (SSFL) trains models on clients' unlabeled data using a teacher to generate pseudo-labels, with a small labeled seed dataset on the server. Automatic Speech Recognition (ASR) is particularly fragile here: pseudo-label errors compound across the output sequence and across training rounds into divergence, leaving a large gap to fully-supervised FL. We show that closing this gap turns on two coupled design axes -- the teacher (which model generates the pseudo-labels) and the anchor (the server-side updates on labeled data that stabilize training). On the teacher axis, a per-client online teacher (each client's own evolving model) diverges on its own, but once stabilized it matches or beats the broadcast global teacher (one server model, fixed within a round) -- decisively in-domain and competitively under domain shift. As the seed grows stronger and the online teacher's advantage narrows, a transitioning teacher (global $\rightarrow$ online at round $r$) matches or beats both. On the anchor axis, the server must keep training on labeled data between rounds -- otherwise the online teacher drifts -- and this interleaving, more than the seed model, governs convergence. The two axes are inseparable: aggressive teacher choices pay off only once the anchor stabilizes training, which is highly sensitive to data augmentation and batch size -- the settings that govern how much input and gradient noise the server injects. How much stabilization is needed is domain-dependent, governed by the dispersion of the seed data and its overlap with client data. These findings yield guidelines for SSFL in ASR training, improving over the strongest prior method on 9 of 11 pairs, by $20.8\%$ on average in-domain and $10.0\%$ cross-domain, narrowing the gap to fully-supervised FL.

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.

Data attribution methods using gradient similarity are widely used to analyze and select training data for large language models, but what gradient similarity actually measures is debated. Some interpret it as identifying task-relevant skills, while other work reports that surface form is the main factor. We resolve this debate for supervised fine-tuning examples by varying task and answer format independently. Specifically, we render benchmarks in different answer formats, such that datasets can share a task without a format or a format without a task. We find that gradient alignment follows the answer format, as benchmark pairs sharing an answer format align strongly (disattenuated cosine near 0.4), while same benchmarks rendered with different answer format classes show no alignment (near 0.0). We demonstrate that this ordering holds from the earliest pretraining checkpoints through post-training, and across model scales and families. We then analyze the released selections of LESS, a gradient-based data selection method for instruction tuning, and find that each target's selections over-represent the target's own answer format. Hence, we demonstrate that gradient-based attribution methods track format similarity more than task semantics, meaning that such methods, as well as the semantic interpretation of the gradient, should be tested on data where answer format and task vary independently for greater robustness and reliability.

Generative query suggestion aims to enhance user engagement by anticipating user intents and recommending relevant follow-up queries. A central challenge is to generate slates whose individual queries are useful while the slate covers distinct intents. We propose an Intent-Driven Query Suggestion Framework with dual-stage optimization. First, intent-aware diversity modeling constructs intent-aligned supervised fine-tuning (SFT) data and uses an Intent-Aware Diversity Reward to optimize intent coverage. Second, query-level credit assignment routes individual quality signals to the corresponding query tokens while sharing a slate-level diversity signal across the slate. Experiments on a large-scale production dataset, including online A/B testing and offline evaluation, show improvements in click-through rate, query quality, and intent coverage.

Limited controlled evidence exists on how training data, adaptation method, and model scale jointly affect tool-calling performance in language-model agents. We evaluate supervised fine-tuning (SFT) with LoRA, reinforcement learning (RL) via Group Relative Policy Optimization (GRPO), and SFT followed by GRPO across six Qwen3 models from 0.6B to 32B parameters, covering both in-distribution performance and cross-dataset transfer. SFT with LoRA is the strongest in-distribution method throughout the 0.6B-32B range and best in 15 out of 18 experimental settings. On cross-dataset transfer, the methods are closer: GRPO wins 29 out of 54 settings where training and test datasets differ, but its margin over SFT averages under one point, and SFT->GRPO is rarely strongest in either comparison. Dataset mixing gives consistently strong transfer while staying close to specialized in-distribution training, regardless of method. Additional analysis further confirms that LoRA outperforms full-parameter fine-tuning, demonstrating that LoRA better preserves pretrained agentic behavior.

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.

Supervised fine-tuning (SFT) applies a uniform cross-entropy loss to all target tokens, even though different tokens provide unequal learning signals for mathematical reasoning. This uniform treatment can over-sharpen already mastered tokens while amplifying learning pressure on uncertain, low-confidence tokens, leading to suboptimal training dynamics. We propose Trimmed Logit-Gap SFT (TrimSFT), a simple token-level reweighting method that scales the SFT loss according to the logit gap between the gold token and its strongest competitor. TrimSFT trims supervision away from both extremes: tokens already mastered (large logit gap) and tokens weakly supported by the current model (small or negative logit gap), concentrating learning within an intermediate logit-gap region between them. We instantiate this principle with a Gaussian weight centered at margin m with bandwidth τ, requiring no reference model or additional forward pass. We evaluate TrimSFT on six base models from the Llama, Qwen, and DeepMath families across five mathematical reasoning benchmarks. TrimSFT consistently improves over standard SFT, achieving the best average performance on five out of six models, with gains of up to +26.9 points over SFT on MATH500. Further analyses show that the bandwidth τ matters more than the exact margin location, and that half-trim variants that remove supervision pressure from only one side yield inferior trade-offs. A token-level logit-gap distribution analysis suggests that TrimSFT reshapes model confidence in a more balanced way than uniform SFT or monotonic reweighting methods. These results suggest that reasoning SFT can benefit from trimming both extremes rather than treating all tokens uniformly.

How to divide a fixed annotation budget between supervised fine-tuning (SFT) and reinforcement learning (RL) during LLM post-training remains an open problem. Existing work characterizes only broad trends (e.g., SFT dominates in low-data regimes), lacks a principled allocation framework, and does not examine whether the optimal ratio transfers across model sizes. We frame this problem in terms of near-optimality: rather than seeking a single optimal SFT-RL ratio, we characterize the near-optimal region, the set of allocations within a specified tolerance of peak performance. Empirically, this region is wide even for small tolerances (2-10%), widens with model scale, and transfers reliably from small proxy models to large target models. This yields a practical strategy: small proxy-model experiments suffice to identify a transferable near-optimal region, eliminating the need for exhaustive large-scale search. Our results hold consistently across tasks, model families, and both preference-based off-policy and reward-supervision on-policy RL methods. We further analyze how the asymmetry in annotation costs between SFT and RL data shifts the near-optimal region.

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.

Goal-directed optimization is essential for steering molecular generators to propose candidates with desired properties. However, it is often implemented with policy-gradient reinforcement learning, which requires a generation-trajectory log-probability whose form depends on the model architecture and generation procedure. This makes an optimizer difficult to reuse across architectures and conditional generative designs. Supervised fine-tuning needs none of that machinery, but its update is driven by a fixed dataset, so the reward never enters the update. We introduce Elite-Weighted Supervised Fine-tuning (EW-SFT), which uses reward to guide elite selection of high-scoring molecules, and updates the model by its own pretraining loss on that set. Ablations show that reward information is passed primarily through elite selection, rather than through continuous weighting within the selected set. Because the update consumes only scored molecules and the model's native loss, the same rule applies across autoregressive, masked-diffusion, and discrete-flow generators, and across de novo, motif-extension, and linker-design tasks. Under a fixed budget of 3D shape alignment oracle calls on two kinase reference compounds, EW-SFT consistently outperforms the corresponding native optimizers. It further improves goal-directed optimization under a 2D similarity oracle on four held-out references and achieves comparable performance on a sample-efficiency benchmark without a trajectory-level RL formulation. These results demonstrate that EW-SFT is a unified and effective optimizer across molecular generators, design constraints, references, and oracles.

Supervised fine-tuning (SFT) trains a base language model to imitate target responses, and these targets may require knowledge the base model has not robustly internalized. We study this as a source of hallucinations and frame a group of mitigation methods as \emph{knowledge-aligned SFT}: constraining SFT training targets to the base model's parametric knowledge. Under a unified setup, we compare existing generation-based and estimation-based knowledge-alignment methods and introduce two new variants: Evidence Rewrite, which verifies base-model generations using external evidence, and Recall Rewrite, which retains claims only when they can be consistently recalled by the base model. Experiments with Qwen 3 4B and OLMo 3 7B show that knowledge-aligned SFT can reduce factual hallucinations on WildHalu and Biography while largely preserving general capabilities. Recall Rewrite yields the strongest factuality gains and improves refusal behavior on UnknownBench. It thereby confirms that SFT targets beyond the base model's knowledge drive hallucination behavior.

Current dialogue systems struggle with dynamic information retrieval, often leading to hallucinations and lower response accuracy. We address this by adapting the ReAct framework for Task-Oriented Dialogue, enabling Large Language Models (LLMs) to access external knowledge and produce factual responses. Mainly, we propose an unsupervised fine-tuning pipeline that harvests reasoning trajectories via in-context learning inference. High-quality samples are filtered using an LLM-based judge to construct a robust training set. This is enhanced by a unsupervised self-improvement loop, where improved checkpoints generate increasingly better trajectories for subsequent fine-tuning iterations. Experiments on the SIMMC dataset demonstrate that ReAct-based systems outperform baselines due to superior reasoning and tool use. Notably, our fine-tuned 8B model surpasses a 70B in-context system. Finally, we present an error analysis, impact of scene complexity, and cross-domain generalization.

Supervised fine-tuning (SFT) is the de facto standard for adapting large language models (LLMs) to target domains, but it often degrades the model's general capabilities, a phenomenon known as catastrophic forgetting. Existing approaches typically modify the SFT loss to mitigate forgetting, but they inevitably operate along a domain-generality trade-off. In this work, we step outside this trade-off by decoupling the two capabilities at the model level: we keep the original base model for general capability, and selectively invoke the SFT expert only when domain-specific knowledge is required. Specifically, we propose CPR (Critical-Point Routing), a token-level routing framework between a base model and its expert derivative, based on critical tokens where the base model fails but the expert succeeds. We train a lightweight hierarchical router that estimates the expert-call probability per token, and pair it with a tailored inference procedure that combines momentum smoothing and threshold gating. Across diverse model-domain configurations, CPR achieves state-of-the-art across all settings, surpassing SFT expert by 1.4-5.5% in domain performance while recovering its general-capability drop from 3.4-14.5% to at most 0.5%, with minimal overhead from invoking the expert on only one-third of tokens.

The reliability of Large Language Models (LLMs) for API calling degrades in multilingual settings. A common failure occurs when a model selects the correct tool but generates argument values in an inconsistent language, which we term Argument Language Mismatch (ALM). Although semantically correct, such outputs are operationally invalid and not captured by standard API-calling metrics. We revisit post-training strategies for mitigating ALM and find that, in our benchmark, supervised fine-tuning (SFT) provides a strong baseline, substantially improving argument language consistency and end-to-end function call accuracy. Under consistent model selection, SFT achieves performance comparable to, and sometimes exceeding more complex reinforcement learning (RL) approaches. We further examine whether RL with structured, argument-aware rewards offers additional benefits. While methods such as Group Relative Policy Optimization (GRPO) can improve language consistency and better preserve general reasoning ability, these gains are incremental and most pronounced in generalization and multi-objective trade-offs. Overall, our results suggest that much of the performance in multilingual API grounding can be achieved through careful supervised training, with RL providing targeted rather than fundamental improvements.

Supervised fine-tuning (SFT) is a standard approach for adapting LLMs to a target distribution, but in settings such as personalization, where each author requires separate weight access, optimization, storage, and retraining, its costs become prohibitive. We propose Weightless Fine-Tuning (WFT), a training-free decoding-time method that approximates the distributional effect of SFT without weight updates. WFT computes supervised residuals on an author's training sequence and transports them to the current prompt through a cross-prefix transport operator estimated from dropout-induced cross-covariance. The operator captures how a perturbation at one context propagates to predictions at another, replacing gradient-based parameter updates with logit-space corrections. On three LaMP personalization benchmarks, WFT achieves the best average performance across datasets, matches or exceeds SFT on individual tasks, and outperforms other lightweight baselines on average. In a budget-controlled comparison, WFT approaches SFT performance using less than 7% of the effective computation. Logit-level analysis shows a cosine similarity of 0.875 between the logit shifts induced by WFT and SFT over 95% of the next-token probability mass, suggesting that WFT captures the distributional effect of supervised adaptation without modifying model weights.

We present Stockmark-Nemotron-3-Nano-Omni-JapanDocReader, a Japanese document understanding model built from Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16. The central goal of this work is structured document parsing via capability injection and forgetting control: we inject Japanese structured document parsing capability into a reasoning-oriented multimodal model while preserving its document VQA capability as much as possible. We study parsing-centric SFT, which uses only structured document parsing data; mixed SFT, which combines structured document parsing and VQA data; and parsing-centric RL, which optimizes structured parsing with a task-level reward. Our experiments show that parsing-centric SFT substantially improves structured document parsing performance but causes measurable VQA forgetting. Mixed SFT mitigates this forgetting while preserving nearly the same structured parsing performance. Applying DAPO-based parsing-centric RL on top of the mixed SFT checkpoint further improves structured document parsing beyond the SFT ceiling, producing the final released model. The training data is constructed with a data engine consisting of two complementary synthetic streams: a Japanese Document VQA Stream and a programmatic structured document parsing stream. We also discuss reward design and variance-based prompt filtering for continuous structured document parsing rewards, highlighting their importance for making RL effective in long-reasoning structured document parsing tasks.

Multi-task supervised fine-tuning (SFT) often casts a heterogeneous data mixture as a single optimization problem, even though different tasks may reach their best generalization at different times. msft exposes this mismatch through task-wise roll-out, exclusion, and rollback, but its original formulation materializes the scheduler state as full-model checkpoints, making stage transitions costly to store, restore, and deploy. This paper introduces AuroSFT, a parameter-efficient framework that recasts the carried state of overfitting-aware multi-task SFT as a compact, mergeable adapter state. AuroSFT freezes the pretrained backbone, trains only injected adapters, rolls back adapter checkpoints at task-wise peaks, and continues on the remaining active mixture. At the layer level, each adapter applies an AuroRA-inspired adaptive nonlinear layer to a low-rank weight factor rather than to the sample representation. The resulting update remains linear in the input, rank-bounded, and exactly mergeable into the frozen projection. Under the retained-backbone comparison protocol, AuroSFT achieves 61.36% average accuracy, compared with 59.85% for the corresponding msft reference row, and obtains higher accuracy on all five backbones. Our code is available at the anonymous repository: https://anonymous.4open.science/r/AuroSFT-80D1.

Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) exhibit fundamentally different behaviors in enhancing multi-task reasoning for large language models (LLMs). Our preliminary experiments revealed a phenomenon: SFT suffers from severe task conflicts under multi-stage training, whereas RL enables stable coexistence across diverse tasks. Empirically, we trace this to the parameter level, observing that RL induces sparse and approximately orthogonal updates across tasks. We provide a theoretical explanation for this mechanism by analyzing multi-task gradient interference. Our results reveal a distinction: interference in SFT is norm-limited, scaling with the absolute gradient magnitude, whereas interference in RL is variance-limited, bounded by the gradient variance induced by advantage normalization and on-policy optimization. This small variance bound yields near-orthogonal optimization directions across tasks. Leveraging this insight, we propose Parallel-RL, a paradigm that decouples multi-task training, significantly improving efficiency and flexibility.

Supervised fine-tuning (SFT) can equip large language models (LLMs) with domain knowledge for high-performance computing (HPC) tasks such as data race detection and benchmark question answering. However, knowledge alone does not guarantee task-appropriate behavior: the same SFT model that correctly classifies 88.65% of C/C++ data race samples produces verbose, imprecise answers to factual queries, with 65.9% of MLPerf responses exceeding 40 characters. Reinforcement learning (RL) post-training addresses this gap by optimizing for task-specific rewards rather than token-level imitation. Yet HPC tasks exhibit extreme heterogeneity, with binary classification, factual QA, and semantic generation differing by 58x in answer length, spanning three distinct reward distributions, and showing widely varying SFT accuracy. This makes uniform-weight RL methods such as GRPO suboptimal. We propose HARGO, Heterogeneity-Aware Reward-Guided Optimization, which introduces per-response importance weighting via confidence-modulated advantage: computing a discrimination signal from group-level reward contrast and a confidence signal from reference model log-probabilities, then modulating the advantage before computing per-response weights, without requiring task-type labels. Across four HPC tasks and nine methods, HARGO achieves the best performance on all three primary metrics: WinRate 54.62%, Data Race F1 91.30%, and PLP Similarity 0.8558. Ablation confirms complementary contributions from both signals. HARGO establishes the best overall alignment quality among compared methods for heterogeneous HPC tasks.

Multi-frame medical VQA appears to reward increasingly complex adaptation: controller-style inference, localization-aware reranking, static hard-negative mixing, and staged continuation all appear plausible from first principles. We test a simpler competing hypothesis on MedFrameQA: methods that remain tightly aligned with the benchmark's final answer objective should be the strongest \emph{robust} adaptation family once evaluation is controlled across fixed splits, matched budgets, repeated seeds, and calibration. We compare controller-based methods, scaffold evolution, static mixed supervision, continuation-heavy variants, and direct answer-only supervised fine-tuning (SFT). The strongest robust family is direct decoder-only answer SFT on MedGemma-1.5-4B. Empirically, this family yields substantial improvements in held-out report accuracy over frozen baselines while remaining remarkably stable across repeated seeds and matched controls, ensuring our claims reflect true family-level robustness rather than an isolated hyperparameter peak. Furthermore, post-hoc calibration effectively repairs confidence estimation without compromising accuracy, and the core approach transfers consistently to secondary backbones like Qwen2.5-VL-3B. The main result is therefore not that a complex auxiliary mechanism wins, but that objective-aligned direct answer SFT is the strongest robust adaptation family we found for MedFrameQA. By establishing this strong, minimalist baseline, we hope to redirect community focus toward fundamentally robust optimization rather than architectural complexity.

Alignment training, model organisms, and toy models are usually treated as separate research areas. But projects in all three frequently use supervised fine-tuning (SFT) to pursue the same underlying goals. When projects share a goal, we should test whether lessons learned from one area transfer to the other areas. We study three such transfers, each taking a lesson developed in one SFT setting and testing it in another. First, we port a lesson about behavior generalization from alignment training into toy models. Training on the reason for a behavior, as in Teaching Claude Why, can make the behavior generalize better than training on examples of the behavior alone. Second, we port a lesson about capability preservation from model organisms into the Model-Spec Midtraining alignment setting. SFT on outputs written by a model other than the student (off-model outputs) can damage capabilities when trained on. Mixing in benign on-model (and on-policy) data into our training can prevent most of this damage while still embedding the target behavior. Third, we port a lesson about robustness from model organisms into the same alignment setting. We find that follow-up benign SFT can erase the alignment behavior while preserving capabilities, showing that capability preservation alone does not ensure robustness to subsequent training. Our work illustrates how porting SFT lessons between different research fields can uplift them all, suggesting more researchers should borrow techniques from outside their own areas.

Large reasoning models trained via reinforcement learning (RL) have been increasingly shown to outperform their supervised fine-tuned (SFT) counterparts on mathematical reasoning tasks; Yet the mechanistic basis for this advantage remains unclear. We therefore ask, what internal representational differences enable RL models' superior performance? Our work presents two converging lines of evidence: First, linear probes trained on layer-wise hidden states reveal that RL models tend to achieve higher accuracy in predicting answer correctness compared to SFT models, indicating more linearly separable and structured representations. Second, mean ablation studies show that RL models develop a hierarchical architecture where deeper layers become progressively more critical, whereas SFT models distribute importance uniformly across layers. Together, these findings demonstrate that RL training fundamentally restructures how models represent and process reasoning problems. Finally, we analyze token-count variability under repeated sampling across problems to assess adaptive compute allocation. While we observe higher variability in some RL-tuned models than in their SFT counterparts, we see strong consistency in others, suggesting that token allocation may depend more on the overall training pipeline than on RL versus SFT alone. We believe this token-allocation variability reveals the spread of plausible on-policy reasoning, highlighting which models exhibit stable policies versus those that are under-determined, potentially non-identifiable solution behaviour.

Developers judge a model checkpoint by how it behaves. After supervised fine-tuning (SFT), two checkpoints that perform about the same across relevant benchmarks are treated as interchangeable, equally ready for the next alignment stage, typically preference optimization. We ask whether this judgment misses a pretraining imprint: a difference that no post-SFT benchmark reveals, yet that decides how each checkpoint responds to further training. To find out, we run a controlled experiment on the final window of pretraining, the last data trained on before instruction tuning. Six branches fork from one partially pretrained checkpoint and differ only in this window: 500 million tokens, 0.1% to 1% of the tokens that precede it. Each branch trains its window on a single data source: generic web text, filtered web text, normative discourse, safety text, mathematical text, or synthetic educational text. SFT and post-training are then identical. After SFT the branches behave near-identically, within about one point on instruction following, refusal, and capability, yet the same post-training carries them to very different endpoints, under both a direct preference optimization update and a reinforcement learning update with a verifiable reward. We measure this deviation through refusal of harmful requests: when post-training begins the safety text branch refuses no more than the web text branch, yet by the end it has lost far less of its refusal. The other four branches gain little or no protection, so the effect is selective to what the window contained. The protection requires the safety text to arrive last rather than earlier in pretraining, and it reproduces on a second model family. What a model is pretrained on last shapes how it reacts to alignment. Therefore, a checkpoint should not be evaluated by its post-SFT behavior alone, and what it was trained on last should be reported with it.

Autonomous driving under adverse weather remains a critical challenge, yet existing vision-language benchmarks mainly evaluate under standard conditions, synthetic corruptions, or single modality. As a result, it remains unclear how vision-language models behave under real-world adverse weather with multi-modal inputs. We argue that a key difficulty lies in degraded environmental observability: under fog, rain, snow, and low illumination, multi-modal observations become unreliable and cross-modally inconsistent, posing challenges to scene understanding, and subsequent decision-making. To study this, we introduce \textbf{ObsDriveBench}, a real-world multi-modal benchmark for adverse-weather autonomous driving. Our benchmark is designed with three capability dimensions: \textbf{observability awareness}, \textbf{spatial reliability}, and \textbf{risk-aware decision-making}, enabling fine-grained diagnosis of model behavior under degraded observations. We construct the benchmark through observability meta-annotation, scene description, and capability oriented multiple-choice tasks over synchronized camera, LiDAR, and radar inputs, forming a benchmark with over 14k training and 13k test questions. Experiments reveal consistent performance degradation of existing vision-language models. We further introduce \textbf{ObsDrive} model with normal-weather supervised fine-tuning and adverse-weather reinforcement learning, improving robustness across all three capabilities. The dataset and evaluation code will be released at \href{https://github.com/russellyq/ObsDriveBench}{\texttt{ObsDriveBench}}.

Supervised fine-tuning (SFT) is widely used to adapt large language models to downstream tasks, but its effect on behavioral diversity in sequential decision-making remains under-explored. We study this question in a controlled suite of deterministic board games based on tic-tac-toe variants, where optimal actions are exactly computable and diversity can be measured directly. Across state-level evaluation, arena gameplay, and training trajectories, we find that reasoning-mode generation frequently suppresses action diversity without uniformly improving action accuracy. Furthermore, standard SFT improves accuracy but often induces premature diversity collapse, which exceeds what is minimally required by the accuracy-diversity tradeoff. We then show that action augmentation, which trains on all optimal actions per state rather than a single demonstrated action, would partially mitigates this effect. Our results identify narrow-support imitation as a source of policy collapse in LLM decision-making and suggest that preserving action support during SFT is important for maintaining exploratory behavior.

Procuring supervised fine-tuning (SFT) data forces a buyer to decide, before any downstream training, whether a candidate corpus is worth acquiring. We present \sys{}, a statistics-first gating architecture that treats procurement as a cost-aware routing problem over three intrinsic quality axes -- diversity, utility, and redundancy. Cheap blind measurements are summarised into per-axis estimates with confidence intervals; a gate accepts a decision only when intervals are tight, sample sizes are adequate, and the axes agree, otherwise it escalates the case to an adjudicative debate between a buy-advocate and a reject-advocate judge, resolved by a presiding verdict. On a controlled benchmark of 12 datasets ($2{\times}3{\times}2$ grid over the three axes) with 5 seeds, the gate reaches 0.90 accuracy and 0.83 $F_1$ at $0.017 per unit, sitting between an always-verify baseline (0.75) and an oracle upper bound (0.98) while spending less than always-escalate ($0.020). We further report honest negative diagnostics of the debate path: a con-side win rate of 0.80 ($p\approx3{\times}10^{-6}$) and a 52% position-flip rate under advocate swapping expose negativity and positional biases that a naive LLM-judge would hide. We frame the injected-knob evaluation explicitly as a controlled synthetic benchmark for measurement fidelity and routing calibration, and delimit external validity as future work.

Supervised fine-tuning (SFT) of open-weight LLMs on expert agent trajectories has emerged as a prominent approach to building capable code agents without reliance on proprietary models. A central yet underexplored question is how trajectory quality and quantity jointly shape model performance. We present a systematic empirical study of trajectory data filtering for LoRA fine-tuning of Qwen2.5-Coder-7B-Instruct on the SWE-trajectory dataset (67,074 trajectories, of which 32,161 are resolved). We propose a two-axis quality scoring framework -- Efficiency and Style -- and evaluate it through 16 controlled experiments spanning strategy, scale, and ablation analyses. Since 7B-scale models attain near-zero SWE-bench resolve rates, we adopt cross-entropy (CE) loss on held-out trajectories as the primary metric, validated via first-action generation: CE loss and ROUGE-L are perfectly rank-correlated (Spearman $ρ$ = -1.00), with limited-sample evidence supporting but not conclusively establishing this proxy. Our results reveal a scale-dependent quality-quantity trade-off: at small scales, doubling the dataset (500 to 1,000) yields ~12.7% CE-loss reduction whereas the TopQ-Random gap stays <1% (Mann-Whitney p > 0.10); at 2,000 trajectories this same gap widens to 3.6% (p = 0.016). Ablation further identifies error-retry rate as the dominant sub-dimension, performing comparably to the full composite ($Δ$ < 0.2%). Together, these findings establish trajectory-level quality scoring as a viable but scale-sensitive lever for code-agent SFT and offer a proxy-validated evaluation protocol for the regime where end-to-end resolve rate is statistically infeasible.

Large Language Models (LLMs) have revolutionized AI services, but a critical tension emerges: while personalization improves model performance, it consumes scarce computational resources that users must share. When should a user invest in expensive Supervised Fine-Tuning (SFT) versus lightweight In-Context Learning (ICL)? How does congestion from other users' personalization choices reshape these incentives? And what strategies should platforms adopt when offering multiple personalization algorithms? We develop a tractable framework for LLM serving that captures the statistical-economic trade-offs users face. Our analysis yields several surprising insights. First, we show that ICL and SFT dominate in different regimes, determined by an interplay between pretraining coverage and data signal-to-noise ratios, but congestion can flip these rankings. Second, equilibrium resource consumption exhibits pronounced non-monotonicity: improving pretraining precision reduces the congestion, while broader pretraining coverage and harder tasks sometimes increase it. Third, we prove that offering both personalization methods never hurts the platform's maximal profits, despite potentially increasing computational load. Experiments with GPT-2 on linear regression tasks validate our theoretical predictions about algorithm performance. Complementing these results, our review of documentation from 21 major AI platforms shows that the share offering both SFT and ICL increased from 9.5% in 2021 to 71.4% in 2025, consistent with our platform-design implications.

Supervised fine-tuning (SFT) is widely used to adapt self-supervised speech representations to downstream classification tasks. Small gains observed under a single pretrained checkpoint are often interpreted as method-level improvements, i.e., a higher attainable performance ceiling. We show that such conclusions are not always reliable because SFT outcomes depend strongly on the specific pretrained instance. We conduct a systematic study on 3 SUPERB classification tasks, evaluating 8 SFT variants across 9 pretrained checkpoints from wav2vec~2.0, HuBERT, and WavLM, with multi-seed repetitions on representative base-scale models. We find that the identity of the statistically indistinguishable top-group SFT recipe is often checkpoint-dependent, with limited transferability across pretrained instances. These findings suggest that many reported downstream gains reflect instance and seed dependent elicitation match, rather than universally improving the attainable performance ceiling.

Large language models have made strong reasoning gains through supervised fine-tuning, reinforcement learning, and on-policy distillation, yet these post-training methods are usually evaluated only by final-answer accuracy. We study how they reshape confidence during reasoning. We introduce a three-stage calibration framework that evaluates confidence before, during, and after chain-of-thought generation, corresponding to difficulty estimation, early termination, and answer aggregation. Through a controlled comparison on mathematical reasoning benchmarks, we find that OPD provides the most useful pre-reasoning confidence, SFT gives the strongest online signal for early stopping, and RL produces the most reliable trace-level signal for aggregation. We further show that confidence reliability is position-dependent: RL confidence becomes informative after a path-commitment phase, while OPD confidence is useful early but can become inversely calibrated later. Based on this observation, we propose PosConf, a position-aware confidence strategy that uses confidence only from reliable relative-position intervals. PosConf improves RL answer aggregation by 6.1 points over majority voting and consistently improves OPD early stopping under tight token budgets, with gains up to 4.3 points by avoiding its later inverse-calibration region, showing that \emph{confidence in reasoning models should be used both stage-wise and position-awarely}. Our code is available at https://github.com/EIT-NLP/Post-Training-Calibration.

Supervised fine-tuning (SFT) is the standard approach for adapting pretrained language models to downstream domains, yet it often improves target-domain behavior at the cost of degrading pre-existing capabilities. Standard cross-entropy fine-tuning promotes only the observed label token and leaves unconstrained how probability mass is redistributed over other plausible alternatives, potentially distorting the rich local preference structure learned during pretraining. We first analyze next-token predictions using Shannon and Renyi entropies, revealing that pretrained models exhibit a regular multimodal entropy structure. These entropy peaks correspond to varying numbers of plausible alternatives, indicating that the base model intrinsically encodes rich distributional knowledge beyond the single supervised token. Motivated by this observation, we propose LP-SFT, a Local-Preserving Supervised Fine-Tuning objective designed to explicitly protect this inherent entropy structure. At each step, LP-SFT constructs an adaptive support of alternative tokens and applies a locally normalized preservation loss to maintain the base model's relative structure among them, while standard cross-entropy independently optimizes the supervised token. Across mixed-domain and single-domain fine-tuning experiments, LP-SFT improves overall performance over vanilla SFT and recent SFT-enhancement baselines, achieving the best balance between pass@1 accuracy and pass@k performance. These results suggest that local preservation helps mitigate capability degradation without collapsing sampling-accessible diversity.

Finetuning a language model on documents that are explicitly annotated as fictional results in a model that still actually believes the documents' core claims, an effect known as Negation Neglect. In our evaluations, models trained on documents prefixed and suffixed with such annotations correctly identify the relevant claims as fictional only about 9% of the time. To address this, we introduce Goggles, a learned module that intervenes on the finetuning gradient rather than the data. During supervised finetuning, a Goggles module edits the gradients an LLM LoRA receives, imparting a chosen epistemic frame (the stance the model takes toward the nature of what it reads) to whatever the documents teach. A Goggles instance is trained once for a given base model, frame, and LoRA configuration, then applied frozen to documents it was never trained on. Trained through Goggles on those same documents, now carrying no fictional annotation, the model flags the content as fictional roughly 91% of the time, while preserving capability (GPQA and TruthfulQA match or exceed baseline). The same architecture supports other frames: a Goggles instance can be trained to treat documents as "part of an AI safety evaluation by Redwood Research" rather than simply as fiction. The imparted frame persists under continued finetuning that pushes back toward the claim, where prior interventions revert. Goggles suggests a path toward training language models on known-misaligned data without absorbing the behaviors that data demonstrates.

While Large Language Model (LLM) agents demonstrate proficiency in static benchmarks, their deployment in real-world scenarios is hindered by the dynamic nature of user queries, tool sets, and interaction dynamics. To address this generalization gap, we formalize OpenAgent (Tool-Use Agent in Open-World), a problem setting characterized by distributional shifts across query, action, observation, and domain dimensions. To systematically diagnose its impact, we construct a controlled sandbox environment where we define fine-grained environmental shifts across a four-tier hierarchy, Perception, Interaction, Reasoning, and Internalization, and conduct a comprehensive series of experiments. Our analysis yields a series of key insights, demonstrating that agents trained via both Supervised Fine-Tuning(SFT) and Reinforcement Learning suffer from varying degrees of performance degradation when confronting open environmental shifts. Building on these insights, we propose Perturbation-Augmented Fine-Tuning, a disturbance-based intervention strategy for SFT that lays the foundation for enhancing agent robustness and utility in realistic environments. Our code will be released at: https://github. com/LAMDA-NeSy/OpenAgent.

Existing defenses are effective when harmful content is explicitly mixed into downstream fine-tuning data, but crafted samples can instead hide harmful supervision inside benign tasks. We propose Embedded Attack, where harmful QA pairs are embedded within benign training samples, and show that representative guardrails often fail to detect them at the example level. To address this, we propose Dual-Reference SFT (DR-SFT), which adapts DPO-style contrastive objective design to SFT through token-level regularization, mitigating harmful fine-tuning beyond coarse data filtering.

Sound event detection (SED) is a core module for acoustic environmental analysis, yet its performance is often limited by scarce labeled data. Recent systems leverage large pretrained audio foundation models, but effective fine-tuning remains challenging because labeled data are limited while unlabeled data are abundant. A previous work, ATST-SED, addressed this problem with a pseudo-label based semi-supervised fine-tuning framework. In this work, we further improve the framework by adopting an embedding-level self-supervised contrastive loss inspired by ATST-Frame pretraining. This contrastive objective better exploits unlabeled data during fine-tuning. One challenge is that mixup serves different roles in the two objectives: pseudo-label learning uses composition mixup, while contrastive learning treats mixup as a perturbation. To resolve this mismatch, we propose conditional mixup, which combines composition mixup and perturbation mixup in one semi-supervised framework and defines the corresponding embedding-level contrastive losses. The resulting model achieves 0.645 PSDS1 and 0.822 PSDS2 on the DESED validation set, establishing a new state of the art.

Supervised fine-tuning (SFT) on a small, high-quality set of long reasoning traces is an effective approach for eliciting strong reasoning capabilities in Large Language Models (LLMs). However, existing methods for curating high-quality SFT data rely heavily on strong reasoning models to filter examples based on diversity and difficulty, making the curation process costly while often yielding suboptimal data quality. In this work, we show that diverse and challenging reasoning examples can be identified using only the initial reasoning tokens. Specifically, we demonstrate that difficult problems can be reliably detected based on the loss of the first 100 reasoning tokens evaluated at a randomly perturbed checkpoint of the pretrained model. We further show that examples exhibiting similar loss patterns over their first 1k reasoning tokens across a small number of perturbed checkpoints extrapolating along the fine-tuning trajectory provably induce similar gradients. We validate our approach through extensive experiments on fine-tuning Qwen2.5-7B and Llama3.1-8B models on the M23K medical reasoning and OpenThoughts-Math datasets. Our method outperforms existing baselines by up to 1.7% while being 91% more token efficient.

Recent advances in large language models (LLMs) have demonstrated that reinforcement fine-tuning of pretrained base models can lead to significant gains in reasoning performance at inference time. In this work, we theoretically analyze why reinforcement fine-tuning induces better reasoning ability than purely supervised fine-tuning (SFT) methods. We model chain-of-thought (CoT) reasoning as a pathfinding problem on graphs and compare the popular method of reinforcement learning with verifiable rewards (RLVR) against traditional SFT. We prove that SFT, when trained on golden shortest paths without negative examples, fails to learn how to efficiently backtrack. In contrast, an RLVR-trained model can learn how to efficiently backtrack from dead ends using only outcome reward. This leads to an exponential separation in inference-time compute between the two methods, and demonstrates that RLVR leads the model to learn the location of difficult decisions in a reasoning chain, ultimately allowing for better allocation of inference-time compute. Finally, we show that the reasoning traces of an RLVR model can be distilled to train a base model to backtrack efficiently as well.

Collecting computer use data from human demonstrations is expensive and slow, motivating the need for scalable generation strategies. This requires two key ingredients: environments in which agents can act and verifiers that can judge whether their demonstrations succeeded. We introduce FaraGen1.5, a scalable data pipeline for computer use agents composed of three modular components: environments, solvers, and verifiers. FaraGen1.5 uses both live websites and synthetic environments that faithfully simulate domains gated by authentication or that require irreversible actions. It employs a solver harness that can be powered by multiple models, including strong frontier models such as GPT-5.4, and also incorporates a user simulator to enable multi-turn rollouts. Finally, FaraGen1.5 scores the resulting trajectories with three complementary verifiers covering task correctness, efficiency, and critical-point adherence. Using data produced by this pipeline, we train Fara1.5, a family of native computer use agents (CUAs) at three scales built on Qwen3.5 (4B, 9B, and 27B). To train these models, we employ a supervised finetuning (SFT) recipe that carefully balances data from FaraGen1.5 for broad coverage, specific high-value tasks, and target model deficiencies in an iterative approach. Each model sets a new state of the art for its size class on browser-use benchmarks: Fara1.5-9B reaches 63.4% on Online-Mind2Web and 86.6% on WebVoyager, while Fara1.5-27B achieves 72.3% on Online-Mind2Web, which is competitive with much larger proprietary systems.

Recent advances in instruction-based image editing have enabled models to perform complex visual edits from natural language instructions. However, in product-centric scenarios where preserving product features, branding, and textual elements are critical, current open and closed source models often struggle to maintain this fine-grained object identity. This issue is further compounded by the lack of datasets for instruction-based product image editing with text fidelity constraints, leaving it largely treated as an implicit capability of instruction-based image editing models. In this work, we introduce the ProductConsistency dataset which is designed to improve product-centric image editing. Our approach includes a supervised fine-tuning (SFT) dataset of 87k samples for product editing, a reinforcement learning (RL) dataset with 869 unique product images, and a new benchmark dataset, the ProductConsistency Benchmark, to allow rigorous and standardized evaluation of editing models. To guide RL training, we propose a Cyclic Consistency reward that enforces semantic preservation of product identity by using caption similarity between the original product description and captions generated from the edited image. We fine-tune both Qwen-Image-Edit-2511 and Flux.1-Kontext-dev using our dataset and demonstrate consistent improvements over baseline models in OCR and Perceptual metrics, and MLLM-based evaluations as well, indicating stronger product consistency, text rendering, and overall visual quality; with the Qwen-Image-Edit-2511 model achieving a 5x reduction in the character error rate. The code and pipeline is available at https://anonymous.4open.science/r/ProductConsistency-6FCC/README.md

The standard heuristic of selecting the SFT checkpoint with the highest pass@1 for GRPO can fail when SFT compresses the rollout distribution. For binary rewards, the expected within group advantage variance is $p(1{-}p)(g{-}1)/g$; when early GRPO drives $p$ below $p^*(g)$, most groups have identical rewards and provide no group relative signal. We study SFT depth ladders for Qwen2.5-Coder-3B and DeepSeek-Coder-6.7B. We test Qwen2.5-Coder-3B across five depths and three seeds, and DeepSeek-Coder-6.7B across four matched depths and three seeds. On Qwen, pre RL pass@1 rises with SFT depth, but peak GRPO pass@10 falls from $0.806$ to $0.481$ (3 seed mean, $n{=}20$); pre RL entropy is positively associated with the GRPO outcome ($ρ{=}{+}0.69$). On DeepSeek, pass@1 remains far above $p^*(8){=}0.083$, and GRPO outcomes compress rather than invert. A two stage diagnostic, combining pre RL entropy triage with an early GRPO entropy monitor, flags high risk checkpoints and can stop failing runs early. Simple KL to reference regularisation and label smoothing variants do not rescue the collapsed Qwen checkpoint in our setting, suggesting the failure is not a trivial GRPO hyperparameter artefact.

Recent World-Action (WA) models demonstrate strong generalization ability and data efficiency, but they typically rely on expert trajectories for training. This reliance limits their ability to acquire fine-grained manipulation skills beyond the demonstration distribution and prevents them from continuously improving through real-world interaction. To address these limitations, we propose WAM-RL, a reinforcement learning framework that enables joint optimization of the world model and the action model through online interaction with the environment. By allowing the two components to co-evolve, our approach enhances fine-grained control and adaptability. Specifically, a WA model consists of a world model and an actor. We design a tailored reinforcement learning method with hierarchical optimization to coordinate their improvement. On the methodological side, we systematically investigate the effects of applying reinforcement learning to the action model, as well as online training of the world model within an RL setting. Our experiments reveal a key insight: optimizing only the actor yields improvements on short-horizon tasks, but fails to provide significant gains on long-horizon tasks. In contrast, jointly optimizing both the world model and the actor is critical for achieving strong performance in long-horizon settings. Our work is the first to introduce reinforcement learning into the World-Action paradigm, and provides insights into how online optimization of both the action head and the world model impacts overall performance.

Optimizing the training data distribution for Supervised Fine-Tuning (SFT) dictates the capability of Large Language Models (LLMs). While existing data curation methods excel at accelerating training under constrained budgets, they are less suited to elevating the capability upper bound. The challenge here is no longer to identify a smaller subset that preserves performance, but to refine the data distribution toward instances most capable of improving the final model. To address this problem, we explore instance-level data attribution using Influence Functions (IF). We identify that standard IF formulations struggle in this setting due to two structural limitations: a proximity gap caused by off-policy validation targets, and a severe bias towards gradient norm. We propose DRIFT (Data Refinement via On-Policy Influence Functions for Supervised Fine-Tuning). Instead of relying on external reference data, DRIFT utilizes the model's on-policy rollouts as validation targets, which empirically minimizes the parameter proximity gap and better aligns with the local neighborhood assumption of IF. It further applies signed weighting based on trajectory correctness and debiases influence scores against the gradient hacking issue, allowing a small set of validation queries to act as reliable anchors for attributing the full dataset. Experiments on 7B-parameter instruction and reasoning models show that DRIFT consistently raises the performance ceiling on both, outperforming existing data curation baselines.

In this report, we present Hy-Embodied-0.5-VLA, abbreviated as HyVLA-0.5, an end-to-end system that spans the full robot learning stack: data collection, model design, continued pre-training and supervised fine-tuning, RL post-training, and real-world deployment. Each component serves a distinct role in this stack.

Supervised fine-tuning (SFT) typically maximizes the likelihood of every token in a demonstrated trajectory. However, an observed token can be non-unique, noisy, or misaligned with the model prior. Strictly fitting toward this one-hot target may be suboptimal, especially when the pretrained model encodes a rich knowledge prior. In this work, we reinterpret SFT as target distribution design: instead of studying only the loss objective, we analyze the token-level target that the loss drives the model to match. We introduce the Q-target framework, which decomposes SFT supervision into two explicit choices: (1) how strongly to rely on the observed token, and (2) how to allocate the remaining probability mass over alternatives. This perspective unifies many existing SFT variants as implicit choices of the target distribution Q. Building on this view, we propose Target-SFT which constructs the training objective directly from the desired target distribution. This method consistently outperforms across the ten reasoning dataset-model settings evaluated, showing the effectiveness of this target-based approach. Overall, our formulation reveals a more fundamental design principle for SFT training and opens a broader search space for SFT objectives.

Supervised fine-tuning with synthetic rationale data is widely assumed to improve language model performance on clinical prediction tasks by teaching models not just what to predict but why. We test this assumption on five-year Alzheimer's disease and related dementias (ADRD) prediction from longitudinal health histories. Across a large-scale controlled experiment of 504 configurations, we find that rationale-based SFT consistently and substantially hurts prediction performance relative to label-only fine-tuning. The degradation persists across model families and data scales, and is not resolved by using a reasoning-oriented base model. Crucially, the failure is not explained by poor rationale quality: human expert annotation confirms that the generated rationales are medically accurate and faithfully grounded in patient-specific evidence, and few-shot experiments show that the same rationales improve performance when used as inference-time demonstrations rather than training targets. We identify the root cause as a structural conflict between narrative plausibility and discriminative optimization. We hope our work paves the path toward a more precise understanding of when and how rationale-based supervision helps and when it does not, guiding the responsible development of language models for high-stakes clinical prediction.

Supervised fine-tuning (SFT) is an efficient approach for downstream task adaptation and often serves as the initialization stage for reinforcement learning (RL), but it can show weaker generalization than RL. A key limitation is its off-policy objective: SFT fits fixed demonstrations token by token, including targets poorly aligned with the model's pretrained distribution, which can lead to overfitting. A recent line of work addresses this issue by assigning larger training weights to tokens better aligned with the current model's predictive distribution, with the intuition that fitting these tokens are less distortive to the model's pretrained knowledge and representations. However, computing the token weights from the model that is currently fine-tuned entangles token weights with the optimization trajectory, inducing a self-reinforcing dynamics as the distribution rapidly departs from the pretrained model. To address this, we propose PriFT (Prior-support guided Fine-Tuning), which derives token weights from a frozen pretrained reference to obtain a stable reweighting signal unaffected by fine-tuning. This signal estimates prior support: the extent to which each target token is supported by the pretrained distribution. Across multiple existing token-reweighting rules, replacing the reweighting signal from the online model to pretrained model consistently improves performance. We introduce two instantiations: PriFT-prob uses pretrained token probability, while PriFT-mass selects tokens by cumulative probability mass under the pretrained distribution. Extensive experiments on mathematical reasoning, code generation, and medical question answering show that PriFT achieves state-of-the-art results among SFT baselines and provides a better initialization for subsequent RL training.

Supervised Fine-Tuning (SFT) followed by Reinforcement Learning (RL) has become a standard pipeline for Large Language Model (LLM) post-training. SFT is expected to provide a useful behavioral prior for RL to further enhance model capabilities. However, checkpoints with excessive SFT often show limited improvement during RL. We attribute this failure to the loss of model plasticity: the reduced ability of an SFT-initialized policy to be effectively reshaped by subsequent RL. To better understand this phenomenon, we conduct detailed analysis from multiple perspectives, including parameter changes, output spaces, and RL optimization dynamics. Our results show that models from excessive SFT tend to produce over-confident token distributions and exhibit sharp parameter landscapes, which make them harder to optimize in the RL stage. To enable a more robust SFT-to-RL handoff, we propose \texttt{Rejuvenation}, a simple yet effective method that restores plasticity while preserving useful SFT-acquired priors. Rejuvenation leverages base-anchored model fusion to reduce excessive SFT-induced drift with targeted neuron reset to mitigate model rigidity. Experimental results on both math reasoning tasks and agentic tasks demonstrate that our approach consistently improves RL performance on over-trained SFT models, while also enhancing generalization to out-of-distribution tasks.

Supervised fine-tuning (SFT) is a prevailing method for adapting large language models to reasoning tasks by imitating offline expert demonstrations, often treating a single expert trajectory as the target behavior. However, reasoning is not simple path imitation: rigidly following one demonstrated solution may overfit to surface forms and suppress the model's own reasoning distribution. We propose Rollout-Adaptive Supervised Fine-Tuning (RASFT), a policy-aware SFT framework that calibrates expert supervision according to problem-level solvability estimated from verified on-policy rollouts. For each problem, RASFT strengthens expert guidance when the current policy struggles, while relaxing rigid imitation and incorporating correct self-generated trajectories when the model already exhibits reliable reasoning behavior. To preserve useful reasoning priors, RASFT further introduces a clipped inverse ratio between the frozen reference model and the current policy to constrain excessive policy drift. Experiments across multiple models on six mathematical reasoning benchmarks and two code reasoning benchmarks show that RASFT achieves better overall performance than SFT, SFT variants, and representative RL methods. The code is available at https://github.com/zjd1sq/RASFT.

Supervised fine-tuning (SFT) improves end-to-end classical planning in large language models (LLMs), but do these models also learn to represent and reason about the planning problems they are solving? Due to the relative complexity of classical planning problems and the challenge that end-to-end plan generation poses for LLMs, it has been difficult to explore this question. In our work, we devise and perform a series of interpretability experiments that holistically interrogate world model recovery by examining both internal representations and generative capabilities of fine-tuned LLMs. We find that: a) Supervised fine-tuning on valid action sequences enables LLMs to linearly encode action validity and some state predicates. b) Models that struggle to use output probabilities for classifying action validity may still learn internal representations that separate valid from invalid actions. c) Broader state space coverage during fine-tuning, such as from random walk data, yields more accurate recovery of the underlying world model. In summary, this work contributes a recipe for applying interpretability techniques to planning LLMs and generates insights that shed light on open questions about how knowledge is represented in LLMs.

Supervised fine-tuning is the predominant approach for adapting autoregressive vision-language models to downstream tasks. Recent work has shown that this paradigm is highly vulnerable to backdoor attacks, and that existing defenses are ineffective in open-ended generation settings. In response, we propose BYORn, a backdoor-robust fine-tuning framework motivated by the observation that poisoned target responses are often semantically implausible given the corresponding image-text inputs and a pretrained model. BYORn identifies such misaligned responses and dynamically replaces them with alternative responses generated by the model, thereby breaking the correlation between triggers and target outputs. The resulting objective gradient corresponds to the gradient of the empirical estimate of the population risk upper bound over the clean data distribution. Empirically, BYORn consistently improves robustness to backdoor attacks while preserving clean-task performance, establishing a new trade-off frontier between generalization and attack success rate. Finally, we demonstrate that BYORn remains effective against adaptive attacks specifically designed to circumvent the proposed defense.

Large language models are increasingly deployed in multi-turn interactive settings where users or environments can iteratively provide lightweight feedback. Unfortunately, optimizing such behavior presents a sharp dilemma in practice: online reinforcement learning is able to effectively address multi-turn dynamics but is prohibitively expensive due to the cost of generating full correction trajectories at every update, whereas offline supervised fine-tuning (SFT) is efficient but suffers from distribution shift and behavioral collapse. To this end, we novelly propose DRIFT (Decoupled Rollouts and Importance-Weighted Fine-Tuning), a framework that operationalizes the theoretical insight that the KL-regularized RL objective is equivalent to importance-weighted supervised learning. DRIFT decouples rollout from optimization by sampling offline interaction trajectories from a fixed reference policy, deriving return-based importance weights, and optimizing the policy via weighted SFT on the resulting dataset. Empirically, we demonstrate that DRIFT matches or exceeds the performance of multi-turn reinforcement learning baselines while maintaining the training efficiency and simplicity of standard supervised fine-tuning. Code is available at https://github.com/2020-qqtcg/DRIFT.

Emergent misalignment (EM) is the surprising tendency of language models to become broadly misaligned after fine-tuning on narrowly misaligned examples. While EM has been extensively studied in the supervised fine-tuning (SFT) setting, evidence that it also arises from reinforcement learning (RL) is limited to large, closed-source models, leaving the phenomenon expensive to study and difficult to reproduce. We characterize EM from RL in small, off-the-shelf open-weight models along three axes. First, we show that rewarding narrow, overtly misaligned behavior produces substantially higher general-domain misalignment than sample-matched SFT. Second, we show that EM from RL can be induced by reward signals that could plausibly arise naturally, such as unpopular aesthetic preferences or poor rhetorical appeals. Third, we evaluate in-training mitigations developed for SFT-induced EM and find that they broadly transfer, with preventive steering with persona vectors, interleaving safety data and inoculation prompting all performing well.

Domain-specific supervised fine-tuning (SFT) often improves in-domain performance at the cost of degrading a model's general capabilities. We view this degradation through two practical gaps in domain SFT: a supervision-compatibility gap, where domain targets differ in style and reasoning format from the original model's natural responses, and a trajectory-preservation gap, where teacher-forced SFT optimizes fixed target tokens without constraining the model's behavior on its own generated prefixes. This process fails to preserve the model's original behavior. We propose RAFT (Data Refinement and Adaptive Distillation for Domain Fine-Tuning with Alleviated Forgetting), a two-stage framework that addresses both factors. First, RAFT constructs model-compatible supervision through self-conditioned rewriting, semantic filtering, and answer fusion. Second, RAFT performs Answer-Conditioned On-Policy Distillation, where the original instruction-tuned model provides soft targets on student-generated trajectories while being conditioned on the fused answer as helpful context. We further introduce top-K temperature distillation and EMA-based adaptive loss balancing to stabilize the domain-general trade-off. Across three instruction-tuned backbones and five domains, RAFT improves average domain accuracy by 23.2% over standard SFT, while recovering part of the SFT-induced degradation on MS-Bench and IFEval, with relative improvements of 18.2% and 10.2%, respectively. These results show that coupling data refinement with trajectory-level preservation provides an effective recipe for domain fine-tuning with alleviated forgetting.

Continual supervised fine-tuning (SFT) is the de facto recipe for adapting large language models (LLMs) to a stream of downstream tasks, but it suffers from catastrophic forgetting of earlier capabilities. Recent work shows that on-policy signals -- training on the model's own outputs -- reduce forgetting more reliably than off-policy supervision. Existing on-policy methods route this signal through a new training objective (e.g., self-distillation losses with a teacher copy), inheriting an extra forward pass, schedule sensitivity, and stylistic drift from the teacher.We instead route the on-policy signal through the training data source. Our method, On-Policy Replay (OPR), rolls out the most recent checkpoint on a small budget of historical prompts, filters the generations by a task reward, and replays the surviving (prompt, model response) pairs as ordinary SFT examples. There is no teacher, no auxiliary loss, and no on-the-fly distillation. Across three 7--8B instruction-tuned backbones (Qwen2.5-7B-Instruct, Qwen3-8B, Llama3.1-8B-Instruct) on the TRACE continual-learning benchmark, OPR consistently reduces forgetting; on the sharpest stress test (Qwen2.5-7B-Instruct, Sequential SFT BWT -13.93), OPR lifts BWT to -0.65 at a 10% replay budget and to -2.29 at a 1% budget -- a 46% reduction in |BWT| over a tuned Vanilla Replay baseline, with 42--46% reductions observed across all three backbones. We give a KL-shrinkage interpretation that places OPR and prior on-policy distillation methods on a single axis, and we present a counterintuitive finding that explains why Vanilla Replay is already a strong baseline: low-score replay is uniformly worse than Vanilla Replay, demonstrating that the active ingredient in OPR is the on-policy distribution, not the response quality alone.Our code is available at https://github.com/Yancey2024/OnPolicyReplay.

We benchmark three supervised fine-tuned models against frontier zero-shot baselines on a 661-row held-out slice of PiSAR (Persona, intent, Screen, Action, Rationale), a 12,929-tuple corpus of screen-anchored behavioural rationales curated from public app-store reviews, Pew American Trends Panel demographics, and the OPeRA shopper traces. Every model, frontier or fine-tuned, is evaluated on the same 661-row slice with the same scoring pipeline. Two findings. First, frontier zero-shot baselines (Claude Opus 4.7 and GPT-5.5) reach sem_sim 0.459 and 0.482 respectively; a fine-tuned Qwen3-VL-8B-Instruct reaches 0.783 and clears sem_sim >= 0.7 on 79% of rows, against 1-2% for either frontier baseline, a gap of 0.30 absolute on the same test set. Second, the same training data and recipe on Gemma-4-26B-A4B-IT scores only 0.441, in the same band as the frontier zero-shot baselines rather than the fine-tuned Qwen. We read this as a recipe-vs-model mismatch: the reasoning-tuned high-parameter model resists displacement and would likely need either more data or a stronger fine-tuning method.

Large language models (LLMs) have achieved remarkable progress, with post-training playing a crucial role in enhancing their reasoning capabilities. Among post-training paradigms, supervised fine-tuning (SFT) is widely used: it leverages external data to provide dense supervision and enables efficient training. However, directly fine-tuning on expert data can hurt generalization when the data distribution is mismatched with the target model's own distribution. In this work, we propose Data Adaptation for Reasoning Tuning (DART), which formulates the use of a fixed, potentially distributionally misaligned SFT dataset as an optimization problem over demonstration transformations. DART trains a mapper model with reinforcement learning to convert original SFT data into model-adapted supervision that better matches the target model's distribution and learning preferences. The transformed data are then used for SFT, allowing the target model to better exploit external supervision. Experiments across multiple models and datasets show that DART improves generalization, achieves higher training efficiency than direct RL, and helps models surpass standard SFT. Our code is available at https://anonymous.4open.science/r/DART525E50D.

Speech editing and zero-shot Text-to-Speech (TTS) share a similar generative foundation conditioned on speech prompts, yet speech editing demands far stricter local acoustic consistency with surrounding unedited content. While prior work has shown that Supervised Fine-Tuning (SFT) enables TTS models to acquire functional editing capability, this approach remains fundamentally bottlenecked by imperfect paired editing data and coarse-grained optimization signals. To address these limitations, we propose CosyEdit2, a speech editing model built on a two-stage post-training framework that progresses from supervised editing initialization to editing-oriented Group Relative Policy Optimization (GRPO) over target-speech-free data. Extensive experiments demonstrate that CosyEdit2 not only substantially advances speech editing performance, but also unlocks better zero-shot TTS capability, revealing a deeper mutual relationship between the two tasks. Audio samples are available at https://cjy1018.github.io/CosyEdit2.

Weak-to-strong (W2S) generalization is a promising framework for scalable oversight, yet existing evaluations often test students under matched train-test distributions. Therefore, we study W2S preference learning under zero-shot distribution shift and find that strong students trained on weak preference labels can appear successful in-distribution while failing to transfer across preference datasets. We provide evidence for a representational failure mode in which weak-supervised fine-tuning can pull the strong model toward source-domain features instead of maintaining broadly transferable preference representations. To mitigate this, we propose Representation Anchoring (Anchor), a simple yet effective regularizer that constrains excessive drift from the pretrained strong model's representation space during fine-tuning, while still allowing task-relevant adaptation. Across preference domains, datasets, and model families, Anchor consistently improves out-of-distribution transfer while maintaining competitive in-distribution performance. Together, our evaluation protocol, transfer-aware metrics, and method expose hidden brittleness in current W2S reward modeling and provide a practical path toward more robust preference transfer.

Fine-tuning large language models (LLMs) frequently induces catastrophic forgetting of prior capabilities. Recent work has shown that reinforcement learning (RL) retains prior capabilities more effectively than supervised fine-tuning (SFT), attributing this to policy-gradient updates remaining closer to the base policy \cite{shenfeld2025rl}. We extend this behavioral account to the mechanistic level and ask whether RL's advantage is mirrored by stronger preservation of internal computational circuits. We introduce differential circuit vulnerability, a head-level measure of how much a circuit degrades under fine-tuning, and use it to compare RL and SFT on Qwen2.5-3B-Instruct adapted to scientific question-answering. We find a clear mechanistic trade-off: SFT adapts more rapidly to the target task but produces substantially greater circuit disruption and forgetting of prior capabilities, whereas RL preserves a larger fraction of the base circuit at the cost of slower task adaptation. These findings suggest that circuit preservation may help explain why RL is more robust to catastrophic forgetting. We released our code here: https://github.com/rl-sft-circuit-research/differential-circuit-vulnerability.