NeurIPS 2026 Papers

Terminal observations are not ordinary long-context text: they are heterogeneous, low-information-density execution traces in which sparse but exact evidence (e.g., error messages and file paths) is interleaved with large amounts of repetitive terminal output. For long-horizon CLI agents, retaining raw observations rapidly increases context cost and can dilute critical signals, while LLM-based summarization or fixed heuristics often fail to adapt across heterogeneous terminal tasks and may discard precise task-relevant evidence.We propose TACO, the first self-evolving Terminal Agent Compression framework, which treats compression rules as reusable, preservation-aware knowledge acquired from interaction trajectories. Rather than relying on manually designed rules, static pruning, or task-specific compressor training, TACO autonomously discovers, refines, and reuses structured compression rules from agent interaction trajectories. A Global Rule Pool accumulates effective rules across tasks, while task-time rule evolution adapts them online to the current workflow. This allows TACO to filter redundant terminal observations while conservatively preserving exact task-relevant evidence.Across six benchmarks, including TB 1.0, TB 2.0, SWE-Bench Lite, CompileBench, DevEval, and CRUST-Bench, TACO consistently maintains or improves task success across models and agent scaffolds. On TerminalBench, TACO yields 1%--4% absolute accuracy gains under standard evaluation and improves accuracy by 2%--3% under matched token budgets. On downstream benchmarks, it reduces total token consumption by 12%--27% while maintaining or improving task success. These results show that self-evolving observation compression can unlock latent capability in existing CLI agents by allocating context budget toward task-relevant evidence, without model fine-tuning or human-crafted compression rules.

Modern training pipelines for language model agents begin with a supervised fine-tuning stage in which a small student imitates a costly teacher. Recent work mitigates the covariate shift of pure imitation learning by collecting teacher feedback at states the student itself reaches, with a prevailing trend toward elaborate filters on the teacher's responses. We frame this design choice as a budget-allocation problem and compare three constructions of supervised training data: short unfiltered teacher continuations at learner-induced states; full teacher trajectories filtered for success (the rejection-sampling step in recent on-policy expert-correction work); or those further restricted to tasks the student cannot solve on its own. Across three agentic benchmarks (HotpotQA, ALFWorld, and Terminal-Bench-Dev), short unfiltered teacher continuations beat pure behavioral cloning at matched supervision budgets, and on HotpotQA also match or exceed the filtered alternatives. On Terminal-Bench-Dev, this construction at one-tenth of the corpus budget and with no reinforcement-learning stage matches the OpenThoughts-Agent baseline that uses the full corpus together with reinforcement learning; the optimal continuation length is task-dependent. The same supervised-learning checkpoints yield faster early-stage gains under subsequent reinforcement learning, with mixed evidence on whether this advantage persists later in training.

Agents built on large language models (LLMs) rely on a range of reliability techniques, including retry, majority voting, and self-consistency, that have been developed in parallel rather than within a common analytical framework. We observe that an LLM sampled at temperature $T$ is a discrete stochastic channel $p(y \mid x)$ in the sense of Shannon's coding theory, and use this identity as the entry point for such a framework grounded in communication theory. Each of these techniques is a special case of one of six classical reliability operators: diversity combining, hybrid retransmission, iterative generator-critic decoding, rateless sampling, structured redundant verification, and difficulty-adaptive routing. Within the framework we give two closed-form results: a noise-variance threshold above which uniform averaging beats quality-weighted averaging, and a contractivity criterion for generator-critic refinement, consistent with a contractive-to-divergent transition we observe between 3B- and 14B-parameter models. We further introduce a cost-aware semantic-nearest-neighbor router whose single Lagrangian knob traverses the quality-cost frontier without retraining. Across six channel configurations spanning local and cloud models on 69 hard tasks, no fixed model-technique-budget choice dominates, motivating per-task allocation. On a 300-item hard split of MMLU, GSM8K, and HumanEval, our router occupies the full empirical Pareto frontier: at matched quality, its normalized cost is ${\approx}56$\% lower than the strongest fixed technique; at matched normalized cost, it improves quality by ${\approx}7$\% ($26$\% over single-shot decoding). These results argue for consolidating these reliability techniques into a single tunable layer informed by channel coding.

Long-form omni-modal video understanding requires models to integrate vision, speech, and ambient audio with coherent long-context reasoning. Existing video benchmarks often trade off temporal scale, modality coverage, open-ended interaction, and interpretable scoring. To address this gap, we introduce $\textbf{LongShOTBench}$, a long video evaluation benchmark designed around three coupled goals: holistic omni-modal integration, intent-driven open-ended interaction, and rubric-level diagnosis. It constructs single- and multi-turn questions from practical viewing scenarios, involving systematical tasks for probing visual, speech, ambient-audio, temporal, and cross-modal reasoning. Each item includes a reference answer and a weighted criterion-level rubric, enabling evaluation to identify which perceptual facts, temporal links, modality-grounding requirements, and reasoning steps are satisfied or missed. All samples are manually verified and corrected to improve grounding, clarity, and rubric reliability. We also introduce $\textbf{LongShOTAgent}$, a training-free omni-modal evidence-seeking agent that couples full-video preprocessing with targeted retrieval, query-adaptive segment refinement, and explicit claim verification over visual, speech, and non-speech audio evidence. Its iterative search-refine-verify loop exposes intermediate evidence and lets modality-specific specialists re-analyze relevant moments before answering. We perform comprehensive evaluation of 105 video-capable models spanning open-source omni-modal models, vision-language systems, audio LLMs, agentic pipelines and closed-source APIs. Across this broad evaluation, current MLLMs remain far from saturating LongShOTBench, while our LongShOTAgent emerges as the strongest training-free system, reaching 66.64\% overall. By releasing the benchmark, leaderboard, and agentic method, our work provides the community with a shared, interpretable testbed for evaluating and advancing long-form omni-modal video reasoning.

Procedural 3D modeling through code is emerging as a versatile paradigm, offering deterministic, engine-ready, and precisely editable assets that neural 3D generators inherently lack. Authoring such procedural content, however, demands deep expertise in 3D software APIs, parametric design, and code-level geometric reasoning. In this paper, we propose 3DCodeBench, a systematic benchmark for evaluating vision-language model (VLM) agents for procedural 3D generation in 3D modeling software. Specifically, 3DCodeBench evaluates how effectively 11 advanced VLMs can serve as procedural 3D modelers by transferring text and image references into procedural code for software. Recognizing that automated metrics may not fully capture the perceptual quality of 3D shapes, we build 3DCodeArena, a ranking platform based on human preference between generated pairs. From extensive evaluations and results, we observe that: (1) Failures mostly arise from API usage, while successful renders still suffer from disconnected or floating 3D geometric components. (2) Test-time scaling, such as higher thinking budgets and multi-turn refinement, improves performance overall. Our findings highlight a critical need for high-quality procedural coding data to advance commercial VLMs. Furthermore, effective procedural 3D modeling requires a robust execution environment that provides high-fidelity feedback for iterative refinement. To accelerate research in this domain, we release 3DCodeBench—including the curated data, benchmark, evaluation protocol, and qualitative results—as a foundational tool for developing VLM-based 3D modelers.

Predictions from ML models support human decision making in several fields, including high-stakes ones such as healthcare and the judiciary. Yet, we still lack a clear understanding of how decision makers learn from ML-based decision support (ML-DS). In this paper, we introduce a general computational framework, the 2-Step Agent, to capture this process. As a prediction from an ML model contains information about the training data, a prediction can also be used for inference. Our framework models (i) how a prediction for a new observation affects the beliefs of a rational Bayesian agent, and (ii) how this change in beliefs affects the estimation of causal effect, the downstream decision, and the subsequent outcome. In addition to the framework itself, we make three contributions. First, for the linear Gaussian setting, we derive a tractable solution for the challenging Bayesian inference problem we introduced, i.e. one in which the agent infers from an ML prediction. Second, we experimentally identify conditions under which ML-DS is beneficial. Third, we show that a single misaligned prior belief can be sufficient for ML-DS to lead to worse downstream outcomes compared to no decision support *even when the ML model is well-specified and the agent is perfectly rational*. Hence, even under ideal conditions, ML-DS can do more harm than good.

End-to-end automation of realistic healthcare operations stresses three capabilities underrepresented in current benchmarks: *policy density*, decisions must be grounded in a large library of medical, insurance, and operational rules; Multi-role *composition*: a single task requires the agent to play multiple roles with handoffs; and *multilateral interaction*: intermediate workflow steps are multi-turn dialogs, such as peer-to-peer review and patient outreach. We introduce $\chi$-Bench, a benchmark of long-horizon healthcare workflows across three domains: provider prior authorization, payer utilization management, and care management. Each task hands the agent a clinical case in a high-fidelity simulator of 20 healthcare apps exposed via 87 MCP tools, which it must drive to a terminal status through tool calls and writing the role's artifacts, guided by a 1,290+ document *managed-care operations handbook* skill. Across 30 agent harness/models configurations, the best agent resolves only **28.0%** of tasks, no agent clears **20%** on strict pass^3, and executing all tasks in a single session slumps the performance to **3.8%**. These results raise the hypothesis that similar gaps are likely to surface in other policy-dense, role-composed, irreversible enterprise domains.

Scaling on-policy distillation (OPD) for large language models (LLMs) confronts a fundamental tension: asynchronous execution is necessary for system efficiency but structurally deviates from the ideal on-policy objective. To address this challenge, we theoretically decompose the objective discrepancy into rollout drift and supervision drift, capturing staleness in student occupancy and teacher context, respectively. Building on this, we introduce a sample-level freshness score that quantifies the reliability of buffered sample with respect to the on-policy objective. Guided by this signal, we further propose $\boldsymbol{f}$**-OPD**, a novel framework that adaptively regulates stale-sample influence and constrains policy drift accumulated under asynchronous optimization. Across reasoning, tool-use, and coding-agent tasks of increasing interaction horizon, $f$-OPD consistently achieves task performance comparable to synchronous training while retaining the throughput advantages of asynchronous execution. Our results establish the first recipe that seeks to achieve a performance–efficiency trade-off in OPD, paving the way for long-horizon agentic post-training at scale.