HART alternates Transformer and RWKV blocks so prediction positions can retrieve history directly, beating pure RWKV at medium and long horizons on GIFT-Eval and TIME-Benchmark and matching Toto-2.0-1B's long-horizon MASE on GIFT with 36M parameters.
Workflow selection for LLM agents is recast as a contextual bandit over a reusable pool, reaching an 81.37% average across five benchmarks while cutting MATH training-phase token overhead by at least 40.55%.
WebLoop jointly optimizes generation, critique, and refinement as roles of one shared policy for functional Web generation: Generator and Refiner outputs receive direct browser-execution rewards, while the Critic never sees execution results and is trained from requirement-level execution labels (discriminability) plus the execution gain induced by subsequent refinements (helpfulness), using a staged reward that first establishes diagnosis and then adds consequence-aware credit. With Qwen3.5-9B, WebLoop reaches 41.5 Overall on WebRise and 38.9% accuracy on WebGen-Bench, improving the base model by 11.3 and 15.4 points.
The work presents LiteTrajEval, a lightweight architecture for budget-bounded trajectory evaluation that derives compact domain-specific rule profiles offline, then preprocesses each trajectory online, marks heuristic failure signals, serializes it under a fixed global budget, and invokes a single rubric-guided LLM judge to produce structured diagnostic reports; on public Magentic-One-style and tau-bench-style trajectory datasets it improves failure-localization alignment with human annotations by roughly 20–35 percentage points on Magentic-One and up to 23 percentage points on tau-retail compared with AgentRx, while reducing cost by about 6x and evaluation time by more than 8x, and it has also been deployed in the authors' enterprise agentic platform.
DAGO models each parent-workflow combination as a bandit arm represented by pretrained embeddings of its constituent workflows' code and prompts, uses a diagonal LinUCB policy to trade off predicted offspring quality against uncertainty-driven exploration, has an LLM generate a child through summary-guided fusion with the child's validation score as the reward, and across six benchmarks in mathematical reasoning, code generation, and question answering achieves the highest macro-average score among the evaluated baselines, improving over AFlow from 80.3 to 81.7 under matched validation-evaluation budgets while reducing aggregate search expenditure by 11.2%.
The work proposes CIPO: it first mines executable composite skills from successful tool-use trajectories via budget-constrained BPE-style merging, then augments GRPO by branching at the first eligible granularity decision of each base rollout and using the paired outcome difference as a bounded supplementary reward; across TOOLATHLON, TOUCAN, and TRAJECT-Bench with Qwen2.5-7B and Llama3.1-8B, CIPO attains the highest Tool F1 and TSR in all six settings with an average Comp./Raw of 0.845 versus 0.912 for GRPO + Skills, and the gain does not come from simply invoking skills more often.
Analyzing on-policy self-distillation (OPSD) through the advantage formulation of RLVR, this work finds that a fixed divergence objective under-penalizes and over-rewards incorrect trajectories and that teacher supervision reliability varies with trajectory outcome and cumulative average teacher entropy; it proposes OG-OPSD, which applies forward KL across correct trajectories and reverse KL only within a prefix of incorrect trajectories whose cutoff is set by the last crossing of the cumulative average teacher entropy curves, consistently outperforming vanilla OPSD and multiple strong baselines on mathematical reasoning, multimodal reasoning, and out-of-distribution tasks across Qwen3-1.7B/4B/8B and Qwen3-VL-2B.
The work proposes Test-Time Calibration Learning (TTCL), a label-free framework that derives self-supervision signals for correctness and calibration from multiple model-generated responses and jointly adapts reasoning accuracy and verbalized confidence on unlabeled target-task data, reporting consistent improvements in both accuracy and calibration on mathematical reasoning and factual question answering.
The work first audits existing latent communication interfaces: across five method-dataset pairs, replacing each message with one from an unrelated question changes accuracy by at most 0.60 points even when communication adds 15.44 points over the receiver alone, indicating the gain can come from interface adaptation rather than message content; it then introduces Draft-KV, which sends the key-value states formed while the sharer drafts an answer through linear projections into a side memory read by a gated attention branch, trained progressively from message reconstruction to answer supervision under a one-sided guard on harm from mismatched messages, with both models frozen and only 1.05M interface parameters trained (348x fewer than C2C); with a Qwen3-8B sharer and a frozen Qwen2.5-0.
This work proposes Top-Theta (Top-j) Attention, a training-free inference-time method that replaces top-k search with static per-head calibrated thresholds to sparsify attention elements, and introduces Softmax Denominator Compensation (SDC) and V-Mean Compensation (VMC) to preserve accuracy, achieving 3–10× reduction in V-cache rows and up to 10× fewer attention elements on LLaMA2/LLaMA3 Q&A, code generation, and long-context summarization tasks with no more than 1% accuracy degradation.