ai application developers

Abstract: Reward-specialized post-training produces strong experts for flow-based generative models, while multi-teacher on-policy distillation (OPD) consolidates their capabilities into a single student. Existing methods, however, route each prompt to a single teacher according to its semantic category, implicitly binding the desired capability to prompt content. This coupling makes capability invocation vulnerable to prompt perturbations and prevents users from explicitly adjusting the strength of the desired capability at inference time. In this work, we introduce CapField-OPD, an OPD framework that integrates multiple teachers into a continuous capability field through explicit capability coordinates. We use teacher models as anchors to construct this field, with the coordinates determining how their outputs are combined. Each capability configuration thus receives a unique supervision target, and capability control no longer depends on prompt semantics. Since the training anchors may not be optimal at inference time, we further profile the learned field on a small calibration set. The coordinate with the highest mean reward serves as the recommended default, while coordinates that are frequently optimal offer a promising candidate set for test-time scaling. Extensive experiments on compositional generation, text rendering, and visual aesthetics demonstrate that CapField-OPD consolidates multiple specialized teachers into a single student while preserving or surpassing their performance, reliably invokes the desired capabilities under semantics-preserving prompt variations, and supports continuous capability control and coordinate-based test-time scaling.

Abstract: In autonomous deep research, a user provides a task and relevant background, then leaves the agent to conduct an extended investigation without further human intervention. We study how this initial user information is reflected in intermediate actions and how these actions relate to final recommendations. We introduce DRaligned, a counterfactual behavioral evaluation framework built on PDR-Bench. By varying one task-relevant user factor while keeping the remaining context fixed, we compare acquisition requests, working drafts, and final reports. Source-grounded extraction, blinded local judgments, and deterministic aggregation yield coarse directional measurements while leaving ambiguous cases unresolved. Our experiments show that strong user-specific delivery can emerge from a largely shared research process: agents investigate similar broad questions but allocate requests differently, and final recommendations distinguish user conditions more clearly than explicit requests do. Reports can also integrate user factors that were not jointly visible during acquisition. In readable draft-to-report comparisons, recommendations often retain their coarse user-specific direction despite substantial rewriting. Final directional differences recur across tested agent models, execution harnesses, and evaluator models, even as execution paths vary. These findings describe how initial user information shapes autonomous research and clarify the relationship between the process an agent follows and the recommendations it delivers.

Abstract: Reinforcement learning with binary correctness rewards trains correctness, not calibrated confidence. The confidence that reasoning models verbalize is systematically overconfident, and the problem is not merely one of scale: verbalized confidence tracks how willing a model is to commit to an answer, not how likely the answer is to be right. Post-hoc rescaling therefore fits one distribution but rarely transfers. We look inside the model instead. On factual question answering, a linear probe on the hidden state between the chain of thought and the answer is substantially better calibrated: its expected calibration error is 5 to 38 times lower than that of the verbalized score across four benchmarks and two model families. However, when used to pick among N sampled answers, that same probe nearly ties majority voting yet falls far short of the oracle. Internal states answer "how certain am I" well and "which answer is right" poorly, so the signal should be reported as a confidence rather than used to select answers. As a result, we introduce probe-guided self-distillation (Probe-SD): score a model's own sampled traces with the probe, overwrite the confidence each trace states, and finetune the base checkpoint of the same family, so nothing but the model itself remains at test time. On Qwen3-14B, Probe-SD cuts ECE from 0.178 to 0.024 in-domain and from 0.542 to 0.113 out-of-domain, where it also beats post-hoc recalibration and self-consistency distillation. The resulting confidence is well-calibrated and useful for weighted voting, behaviors previously attributed to online RL, here obtained with supervised finetuning alone.

Abstract: Large language models can solve increasingly diverse reasoning tasks, yet their performance remains highly sensitive to task prompts, intermediate instructions, and the way reusable problem-solving knowledge is incorporated. Existing optimization methods usually focus on only one part of this design space: they either optimize a monolithic prompt, or separately induce and refine skills from model traces. As a result, they lack a principled mechanism for deciding which component should be updated when failures occur, and they rarely optimize prompts, skills, and skill-use policies in a unified framework. We propose SPARO (Skill, Prompt, And Routing Optimization), a framework that jointly optimizes task instructions, reusable skill blocks, and routing rules. It performs controlled counterfactual evaluations, converts examples' effects into a probabilistic responsibility distribution over prompt, skill, and routing components, samples one component from that distribution, and applies the corresponding targeted mutation. This design moves language-program optimization beyond global prompt rewriting toward structured, reusable, and selectively activated task knowledge. Across five benchmarks and five worker models, SPARO consistently outperforms both prompt-centered and skill-centered optimization baselines. These results suggest that effective language-program optimization depends not only on discovering useful task knowledge, but also on deciding where that knowledge should be stored and when it should be activated.

Abstract: Large language model agents increasingly rely on natural-language skills to solve complex tool-use tasks. However, such tasks often admit multiple valid solution paths, making it inappropriate to improve skills by forcing failed trajectories to match a fixed successful trajectory. Moreover, failed trajectories are rarely entirely wrong: an agent may first collect useful evidence and make meaningful progress, but later deviate into an erroneous suffix. We therefore argue that skill self-evolution should identify where productive problem solving begins to break down, rather than reflect coarsely over the entire failure. Based on this insight, we propose SkillPivot, a deviation-point-guided framework for skill self-evolution. SkillPivot detects the transition from a useful prefix to an erroneous suffix using execution validity, goal progress, and action diversity. A stronger teacher then continues from the same prefix and produces a successful alternative under the same interaction history. By contrasting the student's failed suffix with the teacher's successful suffix, SkillPivot generates localized skill updates while preserving already effective guidance. Experiments on ToolQA, LogicBench, and WildClawBench show that SkillPivot consistently outperforms competing skill-evolution methods, improves multiple agent models, and produces compact, transferable skill updates.

Abstract: Reinforcement learning (RL) improves reasoning, but its performance depends on the policy from which training begins. We study on-policy distillation (OPD) as a preparation stage for RL and ask whether its benefits extend beyond improvements in the distilled model's initial accuracy. Under shared RL settings, students initialized with OPD reach higher final performance than those trained with direct RL or supervised fine-tuning followed by RL. This advantage can emerge even when OPD produces little immediate improvement in accuracy. Pre-RL Pass@k does not fully explain the benefit: similar or even higher values do not necessarily lead to better performance after RL. Behavioral analyses point to alignment with the teacher's distribution beyond top-1 agreement as a possible explanation. Such alignment may favor higher-quality reasoning paths while retaining alternatives that RL can further refine using outcome feedback. We further examine how trajectory sources and divergence objectives affect the value of distillation for subsequent RL. Standard reverse-KL OPD performs better before RL, but forward-KL OPD overtakes it afterward; with teacher-generated distillation trajectories, reverse KL remains ahead at both stages. These findings suggest that the preferred distillation objective depends on both the trajectory source and the training that follows. Our results support evaluating OPD as preparation for RL and selecting distillation choices by the performance achieved after subsequent training.

Abstract: Test-time scaling with parallel branches is widely adopted to improve performance on challenging reasoning tasks. The predominant approach, repeated sampling, draws branches independently from a single policy, which can produce redundant attempts and thereby limit the gains from additional inference compute. To address this limitation, we propose Planned Test-Time Scaling (PTTS), which replaces independent sampling with a coordinated joint policy: a planner generates a solution outline for each branch, steering the branches toward distinct reasoning paths, and an executor produces a full solution conditioned on each outline. Formally, we show that PTTS strictly generalizes repeated sampling and, in a stylized setting, provably promotes coverage of complementary reasoning modes and yields better pass@k scaling. We instantiate PTTS on top of strong reasoning models, keeping them fixed as executors while replacing repeated sampling with PTTS inference to further enhance test-time scaling. Concretely, we develop two variants: PTTS-ZS prompts a model to jointly generate outlines for all branches in a single autoregressive pass, while PTTS-RL directly optimizes the planner against the pass@k reward using truncated execution rollouts for efficient training and a sharper reward signal. Across five mathematical reasoning benchmarks with Qwen3-1.7B and 4B, PTTS-ZS improves pass@64 over repeated sampling by up to 6.7 points, while PTTS-RL further increases the gain to up to 13.4 points. Further analysis indicates that broader coverage of distinct reasoning paths contributes to these gains. Overall, PTTS provides a general framework for improving test-time scaling by coordinating reasoning branches, with zero-shot and trainable instantiations that yield substantial performance gains.

Abstract: Vision-language models (VLMs) have achieved strong performance on tasks such as visual question answering, yet small image resizes can turn correct answers into errors. We investigate whether changes in visual configuration, such as image tiling and token arrangement, contribute to this instability. Across seven checkpoints and four benchmarks, equally small resizes cause more correctness flips when they switch configurations. Surprisingly, in over half of these cases, models answer the question incorrectly but can still read the correct answer when told what to read. Furthermore, attention interventions in LLaVA-NeXT suggest that configuration changes can weaken the use of readable information during answering. We therefore guide models using field cues and their own transcriptions. With annotation assistance, these forms of guidance together correct 97.2% of errors with readable information. These findings show that configuration changes can affect how models use information they can still read.

Abstract: External skills provide domain procedures without parameter updates, but existing methods often edit skills directly from failed rollouts without structured routing from an observed failure to an editable location; existing skill graphs also underuse semantic boundaries, object addresses, and topological dependencies for skill retrieval, targeted updating, and scoped validation. We introduce SkillAA (Skill Abductive Attribution), a structured skill-optimization framework for frozen language models. It represents skill applicability, execution, and composition in a unified graph, allowing the same structure to support skill selection, attribution-guided repair, and update validation. SkillAA contrasts successful and failed executions to route candidate repairs to specific graph objects, updates only the selected local structure, and uses Local and Big Gates to screen candidate changes before commitment. With gpt-5.6-sol, SkillAA reaches 81.5%, 66.7%, and 91.2% on SearchQA, LiveMath, and DocVQA, respectively, and attains the highest observed mean in every main setting. These results support the utility of attribution-guided graph editing and graph-scoped validation.

Abstract: Multimodal large language models (MLLMs) have made strong progress on visual question answering and image captioning, yet they still produce fluent claims about objects, attributes, or relations that are not grounded in the image. Many remedies either modify decoding at test time, which adds latency, or fine tune with preferences such as DPO variants, which teach which answer is preferred but not when the model's own answer is unreliable. We argue that calibrated self assessment is the missing signal. We introduce Savor, a training framework that (i) augments the output schema with token and answer confidence, (ii) optimises the policy with a Group Relative Policy Optimisation (GRPO) objective that penalises calibration error and poor abstention decisions, and (iii) uses the learned confidence at inference time to revisit visual evidence only when the model is uncertain. Experiments on POPE, HallusionBench, AMBER and MMHal-Bench across two recent backbones (InternVL3-8B and Qwen3-VL-8B) show that Savor reduces hallucination while preserving general capability on MME and MMBench, with lower Expected Calibration Error than DPO and decoding baselines.