interactive ai developers

Abstract: Training interactive language agents from terminal outcomes poses three coupled problems: which task and state continuations to sample, how to assign delayed feedback to decisions, and how to aggregate token losses for multi-token actions. We present TSBPO, a tree-structured sampling-to-update paradigm that addresses these problems within one training loop. A likelihood-fitted Root/Branch allocator uses retained task and state evidence to target continuations likely to produce contrasting terminal outcomes under fixed rollout budgets. On the resulting replayed physical trees, trajectory advantages compare immediate sibling segments, while observation-group step advantages compare outcomes at matched task/observation states. The actor then sums clipped token surrogates over each complete response and normalizes across active turns, avoiding response-length averaging in the policy-gradient term. With a Qwen2.5-1.5B actor, fixed checkpoints obtain 97.79% seen and 96.64% unseen stochastic success on ALFWorld and an 81.07% per-episode success rate over six stochastic episodes for each of the 500 WebShop test goals. A WebShop step-term ablation yields 75.93% at the same update-100 evaluation; an ALFWorld aggregation comparison shows higher observed command validity with complete-turn weighting. These results support the evaluated training configuration on two interactive benchmarks.

Abstract: Do preferences Large Language Models (LLMs) express in standard elicitation tasks predict their strategic behaviour? LLM agents interacting with humans and other AI systems face potentially conflicting objectives, such as individual welfare, group welfare, and environmental concerns. Prior work characterizes LLM preferences through behavioral-economic and psychological elicitation or strategic games, but rarely tests consistency across settings and prediction in other strategic environments. This requires modeling how competing objectives jointly shape behavior, whereas existing approaches typically measure preference dimensions in isolation. Drawing on utility theory, we introduce the Preference-Behavior Model (PBM), which combines a preference representation with mechanisms capturing imperfect utility maximization. We formulate estimation as an inverse-game problem, jointly inferring preferences and behavioral parameters from observed choices. We evaluate PBM in three social dilemmas that expose conflicting motives: a Public Goods Game, a Collective-Risk Dilemma, and a Common-Pool Resource dilemma. We find that while preferences estimated from established behavioral economic elicitation instruments do not consistently predict behavior in these dilemmas, models fitted in one dilemma retain predictive value in others, suggesting that inferred preferences capture meaningful aspects of LLM decision-making that elicitation instruments may miss. Finally, we assess alignment with persona instructions, finding that inferred preferences broadly reflect intended priorities with specific divergences across dimensions, LLMs, and games being identified by our approach. Our methodology provides a framework for studying how consistently LLM preferences explain behavior, assessing alignment with intended priorities, and revealing trade-offs and behavioral patterns that standard elicitation may miss.

Abstract: On-policy distillation aligns teacher feedback with a student's current behavior, yet acquiring useful supervision for multi-turn agents requires identifying both where guidance is needed and whether the student can sustain the interaction afterward. We introduce Frontier-OPD, which addresses this problem through the student's learning frontier: decision situations that remain challenging but admit useful autonomous continuation after brief teacher guidance. To operationalize this frontier, Frontier-OPD organizes on-policy experience into a shared-state trajectory graph, where trajectory-balanced aggregation of teacher-student disagreement identifies distinct intervention candidates while preserving the histories needed for replay. Short teacher bridges are then assessed by comparing agreement over the bridge and the ensuing student continuation with an equally long segment of the original trajectory, providing a behavioral proxy for guidance that the current student can build on. Accepted teacher responses augment ordinary distillation data, and fresh rollouts renew the frontier estimate as the policy changes. Experiments on ALFWorld, WebShop, and ScienceWorld report higher mean autonomous performance than the compared distillation baselines across the evaluated student scales and teacher configurations, with ablations indicating complementary effects of candidate selection and guidance filtering. In one ALFWorld timing setting, the additional collection cost is offset by a shorter time to a target success rate.

Abstract: Vision-Language-Action (VLA) policies increasingly reason before acting, but existing reasoning models compress this intermediate decision into a single representation. When several plans are valid, this commitment hides alternatives that could otherwise be selected or revised. We introduce the Generative Plan Latent (GPL), which represents the reasoning state as a low-dimensional conditional distribution over plans, such as which object, grasp, or route to use. This makes the reasoning state addressable: without retraining the VLA, plans can be sampled (coverage), selected on request (steering), and reweighted by deployment outcomes (outcome adaptation). We test all three on SimplerEnv manipulation with $π_0$, and extend the analysis to RoboCasa, LIBERO, and a maze domain. Steering the plan raises the fraction of executions that follow the request from 0.48 to 0.94, while the same guidance on the native action distribution shows no reliable gain. Finally, we show that an explicit plan is most useful when it carries decision information not already determined by the observation.

Abstract: Experience-driven self-evolution offers a promising path for large language model (LLM) agents to continually improve through interaction. Yet existing approaches often rely on successful trajectories or external evaluation signals, leaving unclear whether agents can evolve when such positive anchors are unavailable. We study this question through Cold-Start Failure-Only Evolution (CFOE), where agents start without accumulated experience and learn only from failed interactions without privileged guidance. Across three self-evolving paradigms, multiple LLM backbones, and four interactive environments, we uncover a consistent repair-generalization gap: failure-derived experience can help repair previously failed tasks, but rarely yields sustained improvements on unseen tasks. Further analysis shows that agents often misinterpret failures, while even verified-correct experience remains difficult to transfer. Relaxing CFOE reveals that verifier feedback and additional interaction provide limited gains, whereas successful trajectories substantially improve transfer. These findings suggest that current self-evolving agents are better at repairing observed failures than building generalizable capabilities from failure.

Abstract: Real-time interaction requires agents to perceive evolving environments and act under temporal constraints, yet existing evaluations of real-time streaming models focus on conversational quality and streaming perception, leaving effective action under strict temporal constraints untested. We introduce Realtime-Bench, a suite of carefully curated dynamic tasks where delayed actions directly affect success. Under Wall Clock, where environments advance according to actual inference latency, the best success rate reaches only 44% and the best Pass^3 is only 36%, demonstrating that real-time interaction remains challenging. To separate serving latency from models' real-time streaming capability, we compare Wall Clock with Paused Clock, which removes inference delays as an idealized reference, and introduce Token Clock, which advances simulated time according to generated token counts, testing temporal awareness independently of physical serving speed. We find that temporal costs limit success in evolving environments and that higher reasoning effort has mixed effects under time pressure, reflecting a trade-off between decision quality and timely execution. These findings suggest directions for improving real-time multimodal agents, including adaptive reasoning with temporal awareness and continuous perception and action.

Abstract: Training language agents is bottlenecked by the high cost of long-horizon rollouts, yet standard agentic reinforcement learning (RL) learns only from a terminal scalar reward, leaving the supervision in intermediate steps largely unused. Prior work exploits this supervision through value targets that estimate per-step success probabilities or observation targets that fit environment response. Both are hard to construct well: (1) value targets come from the same single scalar signal for each step in the trajectory, and (2) observation targets, such as web pages and terminal logs, are long and noisy and reach only one step ahead. We propose \name, which adds a third kind of multi-step action targets. On successful on-policy trajectories, \name trains the policy backbone to generate the action sequences that follows each history. Each action is also supervised when forecast from the preceding histories, forming a temporal ensemble that the shared backbone absorbs during training, and this enables acting and forecasting to co-evolve. The policy still acts one step at a time at inference, with no overhead. Across five benchmarks (ALFWorld, ScienceWorld, WebShop, $τ^2$-bench, and AppWorld), \name (1) achieves stronger aggregate performance than GRPO and RL baselines with predictive auxiliary targets across simulated environments, while improving overall performance on both realistic applications; (2) delivers up to $1.9\times$ and $2.0\times$ higher sample efficiency than ECHO and GRPO, respectively, on ScienceWorld; and (3) achieves these gains with a single policy backbone, without additional models, separate pretraining, or inference overhead. The codebase is available at https://github.com/ulab-uiuc/CoPE.

Abstract: Multimodal Large Language Models have made significant progress in referring expression comprehension (REC), yet most REC benchmarks still focus on final-target localization. Complex queries often involve intermediate visual references with different dependencies on prior grounding, but final-only evaluation leaves this process unexamined. To address this gap, we formulate trajectory-level REC, in which models receive an ordered sequence of grounding steps constructed from a single complex query and localize the target regions associated with each step. We instantiate this setting in ReferFlow-Bench, which characterizes structured region transitions through four trajectory types, namely Refinement, Shift, Reset, and Mixture. Experiments show that current models struggle to consistently ground intermediate visual references across successive steps. We further propose ReferFlow, which decouples interleaved text-latent trajectory generation from coordinate decoding. Specifically, Stateful Region Latent Reasoning represents each target region as an object-indexed state composed of continuous latents, with explicit visual and spatial supervision to preserve region information for subsequent grounding. Trajectory-Aware Parallel Decoding then recovers all coordinate blocks in parallel through four geometry-constrained refinement rounds, with each block conditioned on its corresponding trajectory prefix, thereby reducing within-box sequential dependencies and eliminating cross-box autoregression. We further apply Grounding-Aware GRPO to promote consistent grounding across steps and learn adaptive trajectory generation from complex REC queries. ReferFlow achieves leading performance on general REC benchmarks and ReferFlow-Bench.

Abstract: Training LLM agents on long-horizon interactive tasks with reinforcement learning typically relies on human-annotated tasks and reliable verification, both expensive to construct for stateful environments. Self-play offers an alternative, but must generate achievable, compositional tasks and provide feedback beyond binary outcomes. We propose Explorer-Zero, a self-play framework that jointly generates tasks and fine-grained, verifiable rewards through environment interaction. A learned proposer explores the environment, constructs tasks, and records intermediate and final states as verification targets. These checkpoints reward partial completion, while alternating proposer--solver training adapts the generated curriculum. We show theoretically that subtask rewards reduce collapse among unsuccessful rollouts when they complete different numbers of subtasks. Without human-annotated training tasks, Explorer-Zero outperforms the evaluated self-play baselines across ALFWorld, ScienceWorld, and AppWorld; on AppWorld, our 30B agent reaches 59.67% task success, approaching 62.65% for RL on human-annotated tasks. Further analyses support task-relevant composition and benefits on complex tasks, while self-play initialization improves subsequent human-data RL across all three benchmarks.

Abstract: Long-horizon agentic RL is constrained by the futures visited by the current policy. When rollouts fail to reach useful behaviors, improved credit assignment cannot provide supervision for states and actions that were never encountered. On-policy distillation (OPD) introduces teacher guidance on policy-induced states, and recent multi-turn variants improve how this guidance is scheduled, replayed, refined, or validated. However, they do not directly use the oracle-controlled future itself as the training trajectory. We introduce \textbf{State-Rooted Executable Rollout Optimization (SERO)}, which augments on-policy agentic RL with executable oracle continuations. SERO restores a state actually reached by the current policy and lets a training-only oracle continue interacting with the real environment, making the downstream observations and oracle actions themselves available for learning. It then applies \textbf{policy-reachability weighting}, which weights each oracle step by the frozen behavior policy's compatibility with the preceding oracle actions. This allows deeper supervision to become effective as the policy becomes more compatible with the oracle branch. The oracle loss is optimized separately from the base RL estimator and no oracle is used at inference time. We evaluate SERO on two challenging interactive-agent benchmarks, ALFWorld and WebShop, using Qwen2.5-1.5B-Instruct and Qwen2.5-7B-Instruct. On ALFWorld, learning the full executable continuation is substantially more effective than supervising only the root oracle action, while policy-reachability weighting further improves out-of-distribution performance.

Abstract: In-context learning (ICL) allows large language models to use demonstrations without updating their parameters. For interactive agents, demonstrations are trajectories of reasoning, actions, and observations that illustrate how to complete a task. Reusing demonstrations in new situations is challenging: directly inputting full trajectories consumes limited prompt space and constrains the overall effectiveness, while natural-language summaries require the agent to translate general guidance into specific action choices. Therefore, we study whether expert action choices can directly supervise a reusable scoring rule. We propose CB-Agent, which extends concept-based ICL to sequential action selection. We extract environment-action decisions from expert trajectories and fit a concept vector by maximum likelihood to model the conditional distribution over candidate actions given the context. This model combines the concept with frozen language-model features and priors. During ReAct interaction, the agent uses this distribution to select environment actions without including demonstrations in the action-scoring context. Across TextWorld, ALFWorld, and ScienceWorld, CB-Agent achieves the highest cross-benchmark average among the evaluated methods at all three tested model sizes and ranks first on TextWorld and ScienceWorld at each size. ALFWorld experiments support the learned concept's contribution beyond prior-only scoring and show further gains when combined with Reflexion. Code is available at https://anonymous.4open.science/r/CB-Agent-DFF9/.

Abstract: Dense affordance supervision is an appealing auxiliary signal for vision-language-action (VLA) policies, yet naively co-training an affordance head can severely damage instruction following. We present a controlled study of how to wire such a head into a modern VLA on the LIBERO benchmark. Our recipe reads the backbone through a stop-gradient and re-injects an intermediate head feature into the action expert via a learned bridge. The stop-gradient is a precondition: letting affordance gradients reach the backbone drops the policy below the headless base (85.5% vs. 93.1%). With the backbone protected, a same-budget 2*2 ablation over injection topology (concatenation vs. residual) and bridge initialization (zero vs. random) shows initialization is the dominant lever. The best wiring, an actively initialized residual bridge, reaches 96.2%, matching the far more elaborate three-expert AffordanceVLA (95.8%) with under 1% extra parameters. Two probes explain the mechanism: ground-truth affordances fed as an input hurt, and inference-time zeroing shows a lazy bridge acts only as a training-time regularizer while an active bridge becomes load-bearing.

Abstract: Hierarchical reinforcement learning improves long-horizon control by organizing primitive actions around persistent subgoals and assigning credit at multiple temporal scales. Recent hierarchical language agents bring these benefits to interactive tasks by explicitly separating subgoal planning from action execution. We observe, however, that an explicit hierarchy does not by itself determine how stable the resulting temporal abstraction is: the learned boundary policy may replace the subgoal almost every turn, making it effectively transient, or retain a subgoal after it has stopped being appropriate. We call this temporal abstraction instability. We propose Stable Temporal Abstraction via Constrained Optimization (STAC), a constrained boundary-policy optimization method that represents premature replanning and stale persistence as constraint costs. STAC applies the resulting Lagrangian costs only to the sampled boundary decision, leaving the underlying algorithm's rewards, critic targets, subgoal advantages, and primitive-action advantages unchanged. Across two backbones and two benchmarks, STAC improves success over a strong hierarchical baseline by $8.1$ and $7.9$ points on ALFWorld and WebShop with Qwen3-0.6B, and by $23.5$ and $15.8$ points with Llama-3.2-1B-Instruct.

Abstract: A multimodal large language model that correctly reports a spatial fact does not necessarily use that fact in subsequent reasoning. To study this distinction, we introduce \textsc{SpaceConflict}, a benchmark of 23{,}196 inputs for the construction and use of spatial state. Under a unified Supported/Contradictory/Unknown judgment interface, it covers local fact binding (L1), relational composition (L2), cross-observation consistency (L3), and state judgment under transformation (L4). Posing a direct-state query, a full-transformation query, and an explicit-initial-state query on the same world reveals an availability--utilization gap: models recover the initial state from visual evidence yet fail when that state must drive a transformation. For Qwen3.5-9B, 50 of 100 sequences with a correctly recovered initial state fail the full transformation, and supplying the state explicitly repairs all 50; the gap narrows with scale but does not close. We therefore propose Operational State Supervision (OSS), which supervises task-relevant spatial states and their transformation trajectories and aligns shared facts across contexts. OSS improves paired accuracy on matched judgments most on L3 and L4, the levels that depend on organizing and using state. Evaluating multimodal spatial reasoning thus requires asking not only whether a model can see and state a spatial fact, but whether that fact becomes a usable state in subsequent computation.

Abstract: Vision-language-action models benefit from the understanding and reasoning capabilities of pretrained vision-language models, but action-only supervision provides limited grounding in world dynamics. Conversely, world-action models inherit spatiotemporal priors from video generation models, yet remain limited in semantic understanding and reasoning under distribution shifts. We introduce UniWAM, a unified architecture that integrates a physical reasoner, a world generator, and an action predictor to jointly learn semantic understanding of the physical world, visual generation, and action prediction. To ensure the quality of the training data, we developed a rigorous data cleaning and annotation pipeline for both human egocentric data and robot data. To adapt the vision-language component to embodied tasks while preserving its inherited language capabilities, we represent low-level actions in natural language and introduce a pre-training recipe that assigns complementary supervision from visual question answering (VQA) data, human egocentric data, and robot demonstrations to the appropriate model components. During post-training, future visual noise augmentation reduces reliance on precise future predictions, while history-conditioned flow matching uses encoded action history to initialize action generation. Together, these designs significantly reduce denoising steps while maintaining performance. UniWAM achieves state-of-the-art (SOTA) performance across multiple evaluations, including in-distribution performance, robustness, generalization, instruction following, and long-horizon task execution. Furthermore, we uncover a log-linear scaling law of unified human-robot co-training, demonstrating the effectiveness of large-scale pre-training on a mixture of human and robot data.

Abstract: Language agents increasingly rely on reusable skills, but post-failure repair is often handled by opaque one-shot reflection: a model generates a skill patch without explicitly maintaining how failure explanations relate to candidate repairs or how unsuccessful retests should influence later edits. We introduce RESKILL, a structured repair framework that maintains an explicit repair state across repair rounds. Given a failed rollout, the framework links failure hypotheses to candidate skill patches, selects local repairs through coverage-based attribution, retests the edited skill set in the environment, and uses retest outcomes to guide subsequent repair updates. The language model supplies structured repair factors, while the repair procedure records them, compares local skill patches by how well they address active failure explanations, and carries unsuccessful retest outcomes into later repair rounds. We evaluate RESKILL on ALFWorld and TextCraft across three model sizes under fixed repair budgets. RESKILL obtains the strongest final success in all six benchmark-model settings, improving average final success by 3.7 percentage points over direct repair and 3.3 points over hypothesis-conditioned repair. These results suggest that explicit attribution alone is insufficient; durable improvement emerges when attribution is integrated with repair selection and persistent retest-conditioned update.