One-Line Summary

New research delivers built-in step scoring from reinforcement-learning post-training, near-10x decoding speedups, and evidence that combining many models has a measurable ceiling—plus a realistic multi-agent benchmark and a practical browser-automation repo.

Research Papers

Progress advantage: step-level scoring from RL post-training

This paper shows how to score each step an agent takes without training a separate reward model: reuse the probabilities from reinforcement-learning (RL) post-training to compute a “progress advantage.” In “Neglected Free Lunch from Post-training: Progress Advantage for LLM Agents,” the authors prove that the log-probability ratio between the RL-trained policy and its reference policy exactly recovers the optimal advantage function under a general stochastic Markov decision process (MDP). The result is an annotation-free, domain-agnostic signal that comes as a byproduct of standard post-training. ¹

Why this matters: building process reward models is hard for long-horizon, real-world agents where actions can be irreversible and environment feedback is noisy. The authors validate the progress advantage across three applications—test-time scaling, uncertainty quantification, and failure attribution—on five benchmarks and four model families. It consistently outperforms confidence-based baselines and, despite requiring no task-specific training, surpasses dedicated trained reward models. ¹

For teams running agents, this provides a drop-in step-level score you can compute from existing RL artifacts, which can help prioritize intervention, calibrate risk, and diagnose where runs go off the rails. The paper also offers practical guidance and analyses of when and how to adopt the signal in real agentic systems. ¹

A co-failure ceiling for combining language models

If you route, vote, or cascade across multiple large language models (LLMs), there is a hard ceiling on the gain: it cannot exceed one minus beta, where beta is the probability that all models are wrong on the same query. The usual metric—average pairwise error correlation—misses this tail risk. The authors give a Clopper–Pearson bound on beta, providing a finite-sample certificate of the largest possible gain before you even train a router. ²

Measured across 67 models from 21 providers, open-ended mathematics shows observed beta = 0.052 versus 0.023 under a 67-model Gaussian copula (about 2.5× underpricing of the all-wrong tail; 90% confidence interval (CI) 1.7 to 3.4; k = 17). On execution-graded code, beta = 0.079. Re-asking GPQA-Diamond questions in free-response (not multiple-choice) reopens the tail with beta = 0.127, and a five-judge panel reports kappa 0.73–0.92. At matched quality, heterogeneous low-correlation ensembles beat high-correlation Self-MoA (self–mixture-of-agents), but on checkable tasks in this pool, combining models rarely beats the best single model without a strong query-level routing signal. Gains come from models failing on different questions, not from simply adding more models. ²

JetSpec speeds up speculative decoding with parallel tree drafting

Speculative decoding (SD) speeds up generation by proposing multiple next tokens and verifying them, but it hits a scaling ceiling: bigger draft budgets only help if acceptance stays high and overhead stays low. JetSpec breaks this bottleneck by training a causal parallel draft head over fused hidden states from a frozen target model—delivering one-forward-pass efficiency while keeping branch-wise causal conditioning so trees remain consistent with the target model. ³

Across math, coding, and chat benchmarks on dense and Mixture of Experts (MoE) Qwen3 models, JetSpec converts larger draft budgets into longer accepted prefixes and higher end-to-end speedups. On Nvidia H100 graphics processing units (GPUs), it reports up to 9.64× speedup on MATH-500 and 4.58× on open-ended conversational workloads, with further latency gains shown via vLLM integration under realistic serving loads. ³

CoffeeBench evaluates long-horizon agents in a multi-agent economy

CoffeeBench is a benchmark that simulates a 90-day coffee supply chain with heterogeneous firms, forcing LLM agents to communicate, negotiate, transact, and manage cash, inventory, and pricing over time. The environment includes two farmers, two roasters, and two retailers; the evaluated model controls one roaster while the others use fixed reference agents, and the goal is to maximize cumulative net income. ⁴

Across recent open-weight and proprietary models, all outperform a passive baseline that takes no actions, and most achieve positive net income. Behavior analysis shows higher-performing models communicate more, while Claude Haiku 4.5 exhibits an “idle-drift” failure mode—repeatedly choosing inaction despite producing coherent plans. The authors release code and agent trajectories for follow-on research. ⁴

Open Source & Repos

Skyvern automates browser workflows with LLMs and computer vision

Skyvern is a GitHub project to automate browser-based workflows using large language models (LLMs) and computer vision. The repository landing page links to a website, docs, and community Discord, and positions the tool with the tagline “Automate Browser-based workflows using LLMs and Computer Vision.” ⁵

If your work involves repetitive web actions, Skyvern offers an AI-centered approach instead of traditional scripted bots. Teams exploring agentic automation in the browser can use it to prototype end-to-end flows with an LLM-driven controller. ⁵

Why It Matters

Today’s set of results points to a practical recipe for agentic systems: use post-training probabilities as a built-in step score (progress advantage) to decide when to push, pause, or ask for help, and deploy faster decoding paths like JetSpec to keep latency low without rewriting your base model. Together, these tackle two of the biggest production pain points—observability and speed. ¹

The co-failure ceiling reframes “more models” as a diversity problem, not a quantity race: unless models fail on different questions and you can route per query, ensembles will top out early. Meanwhile, CoffeeBench raises the bar on what we evaluate—long-horizon, multi-agent economics—so we can tell whether fixes that help in short tasks actually transfer to real operating conditions. ²