StaminaBench: Stress-Testing Coding Agents over 100 Interaction Turns

We introduce StaminaBench, a benchmark that measures the stamina of coding agents: how many consecutive interaction turns (change requests) they can handle before failing. Unlike the prevailing fraction-of-tasks-solved metric, this matches real vibe-coding where sessions run dozens or hundreds of turns. In StaminaBench, agents implement a REST API server and modify it across a tunable number of procedurally generated follow-up change requests - 100 in our experiments, resulting in codebases of up to 6,000 lines. Tests are generated fully programmatically without LLM involvement, ensuring reproducibility and reliability; change sequences are drawn from either a hardcoded or LLM-driven sampler, both constrained to a structured action space to ensure changes are valid. The agent and the server run in an isolated environment and communicate with the benchmark through HTTP, making testing fully black-box and language-agnostic. We evaluate six agent harnesses paired with seven open-source LLMs across 20 scenarios of 100 turns each and find that: (1) all the tested models fail within 5-6 turns, confirming that vibe-coding-style programming without thorough testing produces bugs; (2) passing test feedback back to the agent and allowing it to retry improves passed turn count by up to 12x; and (3) a good harness is required for strong performance: stronger models exhibit up to a 6x gap between their best and worst harness, while weaker models fail with any harness. We release the benchmark and the generated tasks to enable further research into multi-turn coding agent behavior. Benchmark code and data: github.com/amazon-science/StaminaBench.

LLMZero: Discovering Adaptive Training Strategies for RL Post-Training via LLM Agents

RL post-training strategies are dataset-dependent and reveal a recurring empirical pattern: capacity parameters accumulate monotonically across stages, while regularization parameters predominantly oscillate in response to shifting training dynamics. This distinction matters because fixed schedules commit all parameters to fixed trajectories and therefore cannot express the non-stationary exploration-exploitation tradeoffs that regularization must track; the principle provides actionable design rules for multi-stage training. We discover this through LLMZero, a system where LLM agents search over training trajectories via tree search, diagnosing pathologies at each checkpoint and proposing coordinated multi-parameter transitions. Across 4 diverse GRPO tasks, LLMZero discovers strategies that improve over the base model by 9% to 140% relative and over grid search by 6% to 15% relative, consistently outperforming random search and the skill-based agent. The structural principle transfers across tasks, providing an explanation for why discovered strategies take qualitatively different forms yet share similar parameter dynamics.

MotionWAM: Towards Foundation World Action Models for Real-Time Humanoid Loco-Manipulation

World Action Models (WAMs) couple a video dynamics prior to the policy and have shown encouraging results on tabletop manipulation, but iterative denoising over high-dimensional video-action latents leaves them too slow for real-time humanoid loco-manipulation. The problem is compounded by the dominant hierarchical paradigm, in which a high-level manipulation policy controls only the upper body while a low-level controller tracks coarse base commands -- placing upper and lower body in inconsistent action spaces and reducing the legs to balance-preserving locomotion. We present MotionWAM, a real-time WAM that drives autonomous humanoid loco-manipulation from a single egocentric camera by conditioning the policy on the intermediate denoising features of a video world model. MotionWAM replaces the upper-lower split with a unified motion latent and predicts whole-body motion tokens that jointly cover locomotion, torso motion, height regulation, foot interaction, and hand manipulation in a single action space. A three-stage learning framework progressively adapts the video world model to egocentric visual dynamics and to the target humanoid embodiment. On nine real-world Unitree G1 tasks, MotionWAM runs in real time, substantially outperforms Vision-Language-Action (VLA) baselines fine-tuned on the same demonstrations by over 30% in overall success rate, and executes task-driven foot interaction that decoupled upper-lower policies cannot reach. Our results suggest that video-pretrained WAMs can be lifted from tabletop manipulation to coordinated, human-like whole-body humanoid control.

Perceptive Behavior Foundation Model: Adapting Human Motion Priors to Robot-Centric Terrain

Humanoid behavior foundation models aim to acquire reusable whole-body control policies from broad human motion priors, enabling a single controller to produce diverse and expressive behaviors. However, existing motion-centric foundation policies largely assume that the reference motion is already physically compatible with the robot's surroundings. This assumption breaks when the demonstrator, operator, and robot inhabit different environments: a human motion may specify the intended behavior, but not the footholds, clearance, body height, or contact timing required by the robot's local terrain. We introduce \emph{Perceptive Behavior Foundation Model} (Perceptive BFM), a terrain-aware humanoid control framework that grounds human motion priors in robot-centric perception. The model preserves raw kinematic motion references as the behavioral interface, while using local terrain observations to adapt contacts, posture, and timing. To provide scalable terrain supervision, we develop \emph{terrain-conformal reference synthesis} (TCRS), which converts locomotion-oriented human motion clips into terrain-consistent references through contact-aware foothold construction, foot-geometry-aware swing optimization, support-aware root reconstruction, collision repair, and multi-point inverse kinematics. We then train a blind adapted-reference teacher and transfer its terrain-conformal behavior to a deployed raw-reference student through target-frame action alignment. The student is an identity-gated Transformer tracker whose terrain features enter through residual pathways initialized to preserve the motion-tracking prior and trained to produce local corrections only when needed.

DLLG: Dynamic Logit-Level Gating of LLM Experts

Leveraging multiple specialized LLMs can combine complementary strengths, but existing approaches trade adaptability for stability: routing commits prematurely, heuristic ensembling depends on fragile proxies, and parameter merging introduces interference. We propose DLLG (Dynamic Logit-Level Gating), a dynamic logit-level ensembling framework that learns token-level expert fusion from sparse response-level supervision. A lightweight gating module predicts step-wise fusion weights, linking trajectory-level correctness to generation without token-level labels or expert retraining. Across diverse reasoning and code benchmarks, DLLG consistently outperforms strong routing, heuristic ensembling, and parameter-merging baselines across model scales, highlighting learned logit-level fusion as a robust and scalable paradigm for integrating specialized experts.

Consolidating Rewarded Perturbations for LLM Post-Training

Post-training of language models is commonly framed as a sample-score-update loop implemented by gradient descent. A recent line of work, exemplified by RandOpt, relocates this loop to weight space, sampling Gaussian perturbations around a pretrained model and ensembling the top-K rewarded specialists at inference. While competitive with PPO and GRPO under matched training compute, this prediction-level ensemble incurs K forward passes per test example and does not extend cleanly to free-form generation. We ask whether the rewarded population can instead be folded into a single deployable model, replacing the inference-time ensemble with one consolidated update. A split-half analysis over 25 model-task pairs reveals reproducible low-rank structure in every case. We turn this geometry into CoRP (Consolidating Rewarded Perturbations), a gradient-free operator that combines reward-weighted aggregation, compatibility-aware reweighting, and a held-out validation gate, with no gradient flowing through the language model. Across five language models from 0.5B to 8B and five tasks covering math, code, and creative writing, CoRP improves the base model by 8.1 points on average. Using one tenth of RandOpt's perturbation budget, CoRP exceeds single-inference RandOpt by 6.5 points and recovers more than half of the gain of the 50-pass majority-vote ensemble, at one forward pass per test example.

LATTE: Forecasting Peer Anchored Preference Trajectories for Personalized LLM Generation

Personalized generation with frozen large language models requires a conditioning signal that is both compact and current. Existing personalization methods typically retrieve or summarize user histories in text, or compress them into static latent profiles and soft prompts. These approaches are efficient, but they treat a user's past behavior as an aggregate profile and therefore mix stable identity, recent drift, and item content in the same representation. We propose LAtent Trajectory Tracking and Extrapolation (LATTE), a framework that represents personalization as forecasting a peer anchored relative preference state. For each historical session, LATTE subtracts a time masked baseline formed from comparable users who responded to the same item, producing a state that measures how the target user differs from peers under a shared item context. A lightweight sequence predictor then forecasts the next state in this trajectory, and a State to Token Bridge injects the forecast into a frozen instruction tuned LLM through a single anchored soft token. We provide a latent factor analysis showing when peer anchoring cancels shared item variation and why temporal forecasting trades off stale averages against noisy recent states. Experiments on Amazon Reviews 2023 and MemoryCD show that LATTE consistently outperforms retrieval, summary memory, static latent profiles, difference aware latent profiles, and soft prompt compression baselines. On Amazon Reviews 2023, LATTE improves average ROUGE-L from 0.219 for a static latent profile and 0.245 for the strongest added latent compression baseline to 0.259. Additional pairwise comparisons and diagnostic analyses suggest that the improvement is mainly due to forecasting user-specific trajectory information, rather than merely adding a soft prompt interface.

TapSampling: Inference-Time Sampling with a Task-Progress-Understanding Verifier for Robotic Manipulation

Existing embodied control research demonstrates remarkable performance improvements by scaling training data and model size. We instead explore inference-time strategy as an alternative axis. Non-deterministic generative models, such as diffusion and autoregressive models, have been widely adopted in the field of embodied control. However, the single-shot inference paradigm limits their performance. In this paper, we propose \textbf{TapSampling}, a plug-and-play framework for inference-time sampling. First, we introduce an Action-VAE that represents actions in a low-dimensional latent space by mapping policy-generated initial actions into a compressed posterior distribution, from which any number of latent samples can be drawn and decoded into candidate actions that approximate the true action distribution. Second, we formulate action verification as task-progress outcome prediction, using the intrinsic sequential structure of robotic datasets to train a semantically grounded verifier for interpretable action selection. Furthermore, TapSampling is a policy-agnostic framework. Extensive experiments in both simulated and real-world environments demonstrate that our method substantially improves multiple generalist policies without further policy finetuning. Code and models are available at the project page.

Beyond the Target: From Imitation to Collaboration in Speculative Decoding

Speculative decoding (SPD) accelerates large language model (LLM) inference by letting a smaller draft model propose multiple future tokens that are verified in parallel by a larger target model. The dominant SPD paradigm treats the target model as the sole reliable teacher, accepting a draft token only when it exactly matches the target prediction. This design implicitly assumes that the target is always the better choice at every position. In practice, this assumption does not hold. Although the draft is the weaker model overall, it is not uniformly inferior at the token level. In a meaningful fraction of cases where draft and target disagree, the draft's choice is the one that leads to the correct final answer. Inspired by this, we introduce \textbf{Collaborative Speculative Decoding (CoSpec)}, a generalization of SPD that no longer treats the target model as the sole token-level authority. CoSpec trains an arbitration policy via reinforcement learning to decide whether to accept tokens from the draft or target model, selectively accepting draft tokens at mismatches when doing so is likely to yield a correct final answer. Experimental results show that CoSpec maintains substantial speedups while surpassing target-only performance. By shifting the emphasis from imitation to collaboration, CoSpec suggests a new perspective on speculative decoding.

One-Way Policy Optimization for Self-Evolving LLMs

Reinforcement Learning with Verifiable Rewards (RLVR) has become a promising paradigm for scaling reasoning capabilities of Large Language Models (LLMs). However, the sparsity of binary verifier rewards often leads to low efficiency and optimization instability. To stabilize training, existing methods typically impose token-level constraints relative to a reference policy. We identify that such constraints penalize deviations indiscriminately; this can flip verifier-determined direction when the policy attempts to outperform the reference, thereby suppressing gains. To resolve this, we propose One-Way Policy Optimization (OWPO), a method based on the principle of decoupling optimization direction from update magnitude. In OWPO, the verifier dictates the update direction, while the reference policy serves only to adjust the magnitude. Specifically, OWPO applies asymmetric reweighting: it performs Accelerated Alignment for inferior deviations (where the policy lags behind the reference) and Gain Locking for superior deviations (where the policy surpasses the reference). Furthermore, by incorporating iterative reference updates, OWPO creates a ``Ratchet Effect'' that continuously consolidates gains. Experimental results demonstrate that OWPO outperforms strong baselines, including DAPO, OPD, and MOPD, breaking the bottleneck of fixed priors to enable continuous self-evolution without reliance on external reference models.