Trajectory data is getting more vital for training large language models for boosting the agentic abilities. Unlike the verifiable domains such as coding or mathematics, scaling trajectory data for open-ended tasks is much more difficult because these tasks lack singular ground truth and are costly to annotate or verify. In this paper, we propose RetroGen, a self-improving framework of retrospective process supervision. Our key observation is that although expert trajectories are scarce, high-quality final artifacts such as literature reviews, analyst reports and legal judgments, are abundant in pre-training data and can be viewed as compressed traces of the evidence-seeking processes that produced them. RetroGen reconstructs candidate latent trajectories from expert artifacts, verifies them against both the artifact and supporting evidence, and trains models on their own successful reconstruction data, without requiring trajectory data from stronger models. Experiments show that RetroGen improves grounding, faithful synthesis, and long-form evidence-seeking agent tasks.
Kaustubh D. Dhole, Charles L. A. Clarke, Eugene Y. Agichteincs.AI cs.CL cs.IR
Rubrics aim to make language-model evaluation transparent by decomposing response quality into interpretable criteria. However, natural-language rubrics are often ambiguous, require LLM judges, and typically assume criteria aggregated through linear weighted sums, limiting their ability to capture dependencies, alternatives, penalties, and override conditions. We propose ExecRubrics, a framework for representing rubrics as compact executable programs. ExecRubrics encodes evaluation logic as verifiable Python scoring functions, giving natural-language rubric intent an operational semantics: a fixed decision procedure that can be inspected, executed, and edited. On three long-form response benchmarks -- HealthBench, HelpSteer, and ArgQuality -- we show that ExecRubrics can recover substantial preference signal without an LLM judge at evaluation time. On ArgQuality and HelpSteer, the strongest executable variants are within 1.1 and 4 percentage points, respectively, of the direct GPT-5.5 agentic baseline. Executable rubrics are also considerably faster, achieving a 192x average speedup. We show that incorporating external logic and resources from text processing libraries such as NLTK and spaCy can further improve preference accuracy. Our results suggest a novel way of approaching automated evaluation, by offering a faster, more explainable, and less ambiguous alternative to black-box rubric evals, particularly in high-stakes domains such as healthcare and banking where precision and auditability are critical.
Cross-model latent guidance lets a frozen large mentor encode an input once and a frozen small student generate from the resulting signal. Existing methods keep this signal fixed, assuming it stays useful as the output grows; we show this fails in long-form generation. On multi-turn instruction following, static guidance pushes a 4B student's constraint satisfaction 2.5 points below its no-guidance baseline; a training-free refresh every 16 tokens changes only the memory content and restores a 2.0-point gain over that baseline. We propose MentorPulse to keep guidance fresh at practical cost: it compresses mentor states into a capped slot memory, incrementally processes newly generated tokens, and updates the memory that the student reads through gated cross-attention without resetting the student's KV cache. Windowed Refresh Training exposes the bridge to prefix-conditioned memory. Across thirteen datasets, MentorPulse closes 52.2% of the mentor-student gap on macro average, outperforming C2C, T2T, and equal-budget LoRA, with the largest gains on long outputs. It performs best on all eleven mentor-student pairs from three model families, with margins that narrow as the capability gap grows, and a lightweight read-pattern check predicts the gain before deployment. Measured costs identify refresh intervals that dominate text guidance on long outputs.
Where does the novelty a base language model produces with no task come from, and what can an LLM judge of a long stream actually see? We dismantle a cognitively inspired generation loop over 24 conditions on three base models. Most of its effect lives in one operation: a new subject injected every few hundred tokens (an interruption) into a stream whose literal repetition is damped (habituation). We judge windows of generated text only, with the premise as the unit (n=10) and a judge measured for repeatability, against a second judge family and against human readers. Under that protocol the interruption raises judged surprise by 1.2 to 1.4 points and connection by 0.8 over habituation alone. A connective that asks for continuity hurts; a bare paragraph break adds nothing detectable on fresh text; a reset context does at least as well as a kept one; and a pre-registered replication on new premises confirms the primary contrast. Three things the window judge could not see changed the first version of this study, and we think they are of general use. The judge scores the experimenter's injected sentence as the model's own. A fixed rotation of injected sentences makes the model replay its earlier segments from beyond the judge's horizon, and the judge scores the replay as surprise and connection (65-80% of post-interruption windows at periods 150-300). And the local gains do not compose: no arm produces an integrated document. The salience monitor, the in-loop judge, memory across interruptions and a judge-gated Review run with a gate that opens add nothing. On a problem with a verifier (online bin packing), the interruption multiplies valid, distinct candidate heuristics three- to fourfold without raising the quality of the best. We report an evaluation protocol for long generation and a controlled characterization of a simple intervention, not a mechanism of creativity.
Long-form generation exposes fundamental limitations of large language models. Even 70B-parameter models exhibit length collapse at 16k-token outputs, and multi-chapter stories frequently trigger the attribute drift characteristic of the ``lost-in-the-middle'' effect. The ``outline-first, write-later'' paradigm has gained wide adoption, yet existing research evaluates the final writing rather than the outline itself, conflating two evaluation objects that should be decoupled. We construct a unified head-to-head benchmark covering 7 representative long-form generation frameworks across 3 generation granularities -- single-chapter, multi-chapter, and whole-book -- and propose an anchor-based LLM-as-a-judge protocol that directly assesses outlines against the source text on a 5-point anchored scale. Across 21 framework-granularity cells, no single framework dominates; performance depends on the match between a framework's intrinsic output form and the target granularity. SuperWriter ranks first in the length-constrained single-chapter mode, but this advantage degrades in whole-book mode. The outline-side ranking correlates only moderately with the writing-side ranking, supporting the outline--writing decoupling principle. Compute constraints limit the writing-side evaluation to a subset of cases; follow-up experiments will expand the sample size and add cross-model evaluators to enable stronger statistical inference.
Generating long-form content from extensive internal reports remains challenging for organizations operating under strict privacy and security constraints, where proprietary cloud-based LLM APIs are often not viable. While locally deployed open-weight models offer a privacy-preserving alternative, existing retrieval-augmented generation (RAG) approaches on smaller models frequently lack effective global planning and accumulate factual inconsistencies over long outputs. To address these limitations, we present RH-RAG, a multi-agent framework for secure and trustworthy long form generation using local language models. RH-RAG decomposes generation into three coordinated stages: a Planner Agent that constructs a global document outline from high-level semantic summaries, a Writer Agent that incrementally generates coherent section-wise content using bounded coherence memory, and a Checker Agent that mitigates hallucinations through natural language inference-based factual verification and an attestation-driven revision loop. The framework further employs a dual-level retrieval index that supports efficient planning and fine-grained contextual generation on consumer-grade hardware. Evaluations across literary, financial, and legal domains demonstrate that RH-RAG consistently improves factual grounding, semantic coherence, and document-level alignment compared to standard and hierarchical RAG baselines, while achieving reliability competitive with proprietary cloud-based systems without compromising data privacy.
Large language models (LLMs) are increasingly used to generate long-form conversational content such as podcasts from textual sources. While these systems produce fluent and engaging narratives, they often introduce ungrounded information. In this work, we present the first systematic study of faithfulness in document-grounded podcast generation, where grounding must be maintained across conversational turns in long-form, multi-speaker transcripts. We construct a dataset of over 1500 documents spanning five domains and generate podcast transcripts using multiple LLMs. We introduce a turn-level LLM-as-a-judge framework for evaluating whether conversational turns are supported by the source document, and validate its reliability through human studies. Our analysis shows that even state-of-the-art models, including GPT-4o, frequently generate ungrounded content. To mitigate this issue, we propose catch-n-repair, a model-agnostic framework that detects and rewrites unfaithful conversational turns while preserving conversational flow. Experiments demonstrate consistent improvements in faithfulness across both in-domain and out-of-domain settings.
Long-form legal research reports increasingly rely on LLMs and agentic research systems, but their reliability depends not only on answering the task, but also on whether cited legal authorities are trustworthy. A citation can be risky even when it points to a real source: the report may omit limiting conditions, misdescribe the authority, or use it to support a stronger claim than the source allows. We introduce LegalCiteTrust, a benchmark for evaluating citation trustworthiness in Chinese long-form legal research reports. It contains 72 densely annotated report-level tasks and evaluates reports along three dimensions: Coverage, Support, and Citation Trustworthiness. Citation Trustworthiness is operationalized through citation-level Existence, Fidelity, and Applicability (E/F/A). Experiments on general-purpose LLMs, deep-research systems, and legal-specific systems show that task completion, evidence richness, citation density, and citation reliability expose different system behaviors. Retrieval tools can improve evidence support without reliably improving the Trust score, while E/F/A-based revision improves Trust and Final score more clearly than existence-only filtering. These results suggest that trustworthy legal research generation requires citation-aware evidence governance after retrieval: systems must not only retrieve legal authorities, but also select, describe, and apply them reliably.
Xilun Chen, Zhaleh Feizollahi, Ross Goodwin +5cs.CL
Evaluating the factuality of long-form generations has focused predominantly on precision, measuring whether the claims a model makes are correct. The dominant decompose-search-verify pipeline catches incorrect claims well but says little about whether a response contains all the information it should. Measuring factual completeness, the missing half of factuality, is harder: it requires enumerating the full set of facts a complete answer should contain, and these facts rarely form a flat list. They often involve open-ended sets where coverage is what matters, ordered processes, and relationships among facts that a list of independent boolean checks fails to capture. We introduce a two-level meta-rubric framework for evaluating open-ended generation, and instantiate it as Gamut (Grounded Assessment of Multimodal Factuality), a benchmark for factual completeness in long-form generation. The framework rests on a two-level rubric representation: a structured meta-rubric captures the organization and importance of the required content, which is then mechanically compiled into a flat checklist of binary, machine-gradable rubrics that an LLM judge scores reliably. We construct 1,813 questions grounded in real wearable imagery across 10 diverse domains, each paired with an evidence-backed rubric verified by expert human annotators. Because the framework is modality-agnostic, we also release a text-only variant. Evaluating 14 frontier and open-weight models, we find the benchmark genuinely challenging (best score 58.7% from Gemini 3.1 Pro), highly discriminative, and robust to the choice of judge.
Long-form article generation remains difficult for large language models because it combines long context, long instructions, and long outputs. Existing multi-agent pipelines such as STORM improve information coverage by simulating role-specialized agents, but their capabilities are often entangled in prompts and fixed procedures, making them hard to inspect, reuse, or iteratively improve. This paper presents GEIS (Generation-Evaluation-Improvement loop of agent Skills), a loop of named and declarative skills for Wikipedia-style long-form article generation. Implemented and evaluated in Tasi Harness, GEIS composes skills for article writing, browser-based evidence and image collection, diagram rendering, PDF-aware pairwise evaluation, and rule-level skill improvement. Its core writing skill follows Request, Plan, Draft, Audit, Refine, and Deliver; the pairwise evaluation skill produces structured quality reports; and the improvement skill maps recurrent findings into permanent patches to the writing skill in our 20-topic experiment. We evaluate GEIS on 20 Wikipedia Featured Article topics. Under the same generation backend, GEIS improves over the Tasi Harness default writer by 8.0 points on a 100-point PDF quality rubric and outperforms STORM on the two comparable writing dimensions, structural quality and content quality. In the 20-topic improvement experiment, the patched writing skill raises the average score from 82.90 to 86.95, with 17 out of 20 topics improved and the gain mainly coming from content quality. These results show that long-form generation can be reframed from a fixed workflow into an inspectable, modular, and evaluation-guided improvement loop.
Ido Amit, Ido Galil, Ran El-Yanivcs.AI cs.CL cs.LG
As LLMs generate increasingly long outputs, effective uncertainty estimation must identify errors at fine-grained levels rather than discard entire responses. While such methods exist, evaluating uncertainty at any resolution (token to an entire generation) is challenging and highly sensitive to label imperfections, making zero-noise benchmarks essential; yet, long-form generation benchmarks tend to rely on fallible labels rather than deterministic ground truth. We introduce Single-answer Atomic Long-form Target (SALT), a benchmark of six procedurally generated tasks with single deterministic long textual ground truths, enabling unit-level evaluation of correctness, calibration, and ranking without external judges. Equipped with SALT, our analysis of 50+ LLMs reveals key insights: We identify which confidence functions dominate each uncertainty aspect and show that confidence ranking largely breaks at atomic resolution, even when clearer separability emerges at coarser line-level units. SALT further enables controlled atom-level interventions throughout generation, revealing two separable drivers of future errors: propagation from corrupted prefixes, dominated by global context correctness, and bounded degradation from increasing answer-context length. Finally, we demonstrate that reasoning, via Chain-of-Thought prompting or internalized through training, introduces a trade-off, improving accuracy while degrading confidence ranking. These findings directly impact risk-critical applications requiring reliable error identification and mitigation.
Large Language Models (LLMs) generate fluent long-form text, however, often add unsupported factual claims. Existing verification techniques improve factuality by grounding generation in external evidence. However, the same verification policy usually applies to all claims despite being differences in hallucination risks. We propose \textit{FACTOR} (\textit{FACTuality-Oriented Risk-aware Verification}), an inference-time model that adapts verification criteria according to claim-level uncertainty. FACTOR combines uncertainty estimation, adaptive language inference verification, and candidate re-ranking to allocate verification effort where it is most needed. We evaluate \textit{FACTOR} on FactScore benchmark showing that adaptive verification improves factuality while reducing verification cost simultaneously. We further perform different ablation studies to identify the primary driver of these gains. Our results show the effective and model-agnostic performance of \textit{FACTOR} for improving factuality in long-form generation.
Open-ended deep research (OEDR) requires systems to acquire knowledge through multi-round retrieval and generate coherent long-form reports. The outline plays a central role as a structural scaffold that coordinates retrieval, evidence organization, and generation. However, existing methods either fix the outline before writing or refine it with local heuristics, leading to scaffold drift under continuous information accumulation and delayed feedback for evaluating outline modifications. We propose ScaffoldAgent, a utility-guided dynamic outline optimization framework for OEDR. ScaffoldAgent models outline evolution as a structured decision process with three operations: Expansion, Contraction, and Revision, enabling controlled updates to the report scaffold. It further introduces a utility-guided feedback mechanism that estimates the downstream value of each outline operation from retrieval gain, structural coherence, and trial-generation quality. The resulting utility signal guides node selection, operation scheduling, and termination during inference. Experiments on DeepResearch Bench and DeepResearch Gym show that ScaffoldAgent consistently improves long-form report generation and factual grounding over existing deep research agents.
Autoregressive video diffusion models provide a natural formulation for streaming and variable-length video generation by conditioning newly generated frames on previously generated content. However, extending these models to minute-level generation remains challenging: the limited KV-cache budget prevents the model from retaining the full history, while repeatedly conditioning on self-generated frames induces a context distribution shift that accumulates over time, leading to visual artifacts, quality degradation, and temporal drift. In this paper, we propose TetherCache, a training-free and plug-and-play cache management strategy for drift-resistant long video generation. TetherCache organizes the cache into sink, memory, and recent regions, and introduces two complementary mechanisms. First, GRAB (Gated Recall with Attention-Diversity Balancing) selects long-range memory frames using a gated score that combines attention-based relevance with temporal diversity, preserving informative yet diverse historical context under a fixed cache budget. Second, TAME (Trusted Alignment via Memory Editing) lightly edits newly recalled memory tokens by aligning their statistics to a trusted context distribution, reducing the pollution caused by drifted historical features. Built on Self-Forcing, TetherCache consistently improves long-video generation quality on VBench-Long across 30s, 60s, and 240s settings. In particular, for 240s generation, it substantially improves overall and semantic scores while reducing quality drift from 7.84 to 1.33, demonstrating its effectiveness for stable long-horizon autoregressive video diffusion.
Generating coherent and controllable long-form content remains a persistent challenge for Large Language Models (LLMs). While reasoning-enhanced models have demonstrated success in logic-intensive domains, our evaluation reveals that they suffer from a severe length collapse in open-ended writing, where performance degrades sharply as target lengths exceed 2,000 words. We attribute this failure to the limitation of static hierarchical planning, which struggles to provide dynamic guidance over extended contexts. To bridge this gap, we introduce the Interleaved Structural Chain-of-Thought (IS-CoT) framework. Unlike external agentic workflows, IS-CoT embeds a dynamic Plan-Write-Reflect cycle into the generation process, enabling continuous strategy adaptation and global alignment without additional assistance. Based on this framework, we construct a high-quality dataset of interleaved reasoning traces via a multi-teacher pipeline and train IS-Writer-8B. Experiments demonstrate that IS-Writer-8B achieves state-of-the-art performance on challenging long-form benchmarks (e.g., +3.08 vs. DeepSeek-V3.2 on LongBench-Write), exhibiting robust length compliance and coherence competitive with significantly larger proprietary models.
Human evaluation plays a critical role in assessing the quality of generated text. However, the reliability and reproducibility of these evaluations depend on transparent and well-documented protocols -- details that are frequently missing in current practice. In this work, we conduct a large-scale analysis of human evaluation protocols for evaluating long-form generation tasks in *CL conference publications from 2023--2025, including a full manual review of 284 papers and LLM-assisted analysis for another 1.8k+ papers. We define a set of 20 reportable criteria related to reproducibility of human evaluation studies, and apply these criteria to systematically examine reporting norms and practices within the community. We find widespread under-reporting of important aspects of human evaluation study design, leading to ambiguity about what was measured and how, who contributed judgments, and how judgments should be interpreted. Based on these findings, we outline actionable recommendations to support more transparent and reproducible reporting in future research. Our analysis code and annotated dataset can be found at: https://github.com/larchlab/Illusions-of-the-Gold-Standard
Rewards that penalize unsupported claims can improve grounding in long-form generation, but they can also teach models to answer less. We study this refusal-to-richness trade-off in long-form hallucination RL. Instead of using global richness proxies such as length, claim count, detail, or pairwise relevance, we represent each question with a key-point rubric that specifies the required and optional information a useful answer should cover. These rubrics define coverage directly and are used both for evaluation and as reward signals. Across grounding-only, proxy-based, rubric-only, and combined rewards, we find a stable trade-off: strict grounding rewards improve support but suppress coverage, while unconstrained rubric rewards improve coverage but weaken grounding. A soft combination of grounding, rubric coverage, and relevance gives the best balance in our experiments, improving in-distribution support while transferring better to out-of-distribution checklist tasks than either grounding-only or rubric-only rewards.
Lin Li, Georgia Channing, Suhaas M Bhat +2cs.CL cs.AI cs.LG
Large language models (LLMs) have achieved remarkable progress in open-ended text generation, yet they remain prone to hallucinating incorrect or unsupported content, which undermines their reliability. This issue is exacerbated in long-form generation due to hallucination snowballing, a phenomenon where early errors propagate and compound into subsequent outputs. To address this challenge, we propose a novel inference-time hallucination mitigation framework, named Segment-wise HAllucination Rejection Sampling (SHARS), which uses an arbitrary hallucination detector to identify and reject hallucinated segments during generation and resample until faithful content is produced. By retaining only confident information and building subsequent generations upon it, the framework mitigates hallucination accumulation and enhances factual consistency. To instantiate this framework, we adopt semantic uncertainty as the detector and introduce several vital modifications to address its limitations and better adapt it to long-form text. Our method enables models to self-correct hallucinations without requiring external resources such as web search or knowledge bases, while remaining compatible with them for future extensions. Empirical evaluations on standardized hallucination benchmarks demonstrate that our method substantially reduces hallucinations in long-form generation while preserving or even improving the informativeness of generation. Code is available at: https://github.com/TreeLLi/hallucination-rejection-sampling.
Hyunwoong Kim, Seongeun Lee, Hannah Yun +2cs.CV cs.AI cs.LG
Group Relative Policy Optimization (GRPO) and its variants, originally developed for Large Language Models (LLMs), have recently been applied to Multimodal LLMs and produced strong results. However, their coarse-grained holistic credit assignment from a single scalar advantage underfits vision-language (VL) tasks, where outputs are often long-form responses grounded in semantically rich images. To address this limitation, we exploit a structured signal that single-scalar formulations discard: the natural segmentation of long-form VL outputs. Concretely, we propose Segment-Decomposed GRPO (SD-GRPO), which z-normalizes verifiable per-segment rewards across the rollout group, yielding a vector of per-segment advantages in place of a single scalar. We evaluate SD-GRPO across three settings spanning controlled and real-world long-form VL generation, organized by increasing semantic entanglement across segments. On a controlled multi-panel dense-captioning task constructed from DOCCI, where segments are semantically independent, SD-GRPO consistently outperforms the GRPO baseline, with larger gains at higher segment counts. Extending to a controlled multi-chart long-form VQA task constructed from MultiChartQA, we show both theoretically and empirically that rollout-level rewards suffer from cross-segment credit misattribution that scales with output length. On a real-world scientific figure captioning task on the MMSci dataset, where subfigure captions share context across the figure, blending holistic and per-segment rewards further improves on both, suggesting per-segment normalization alone is insufficient when segments are semantically entangled. Finally, by integrating SD-GRPO into Dr. GRPO, we confirm that it can be applied to any GRPO framework with minimal implementation overhead to enhance long-form VL generation.