Research Radar: HybridAgents Think in Latent Space and Act in Text for 35-55% Fewer Tokens; CarryOn Replaces Agent History With Reusable Objects; and 114 of 135 Agent-Written Science Papers Fail an Integrity Check

Hybrid Reasoning Agents (arXiv 2610.18861)

Submitted October 13 by Jiaru Zou, James Zou and colleagues, and at 229 points on Hacker News by October 18, this paper asks: "How can we build an agent that thinks beyond human language, yet still acts through it in the external environment?" The authors' answer is HybridAgents, "which dynamically navigate between latent reasoning and text-based interaction throughout a unified generation process." A small component called HybridLink switches between modes, and "the agent's continuous working memory preserves lossless hybrid context across multiple turns." Training is two-stage: HybridSFT "stabilizes tool-integrated hybrid generation", then HybridRL "scales reasoning by learning when to think in latent space and when to act through text." Across 3 model scales and 11 benchmarks in mathematics, science, medicine, search and agentic reasoning, the paper reports that HybridAgent "consistently improves performance by 10.4%, while delivering 35.5%-54.7% token reduction, 1.8x-2.6x faster inference, and 2.7x-3.7x cost savings." A stated design goal is that latent thoughts stay "directly inspectable", which matters because latent reasoning is usually criticised for removing the readable chain of thought.

Pack Light, CarryOn (arXiv 2610.18784)

Yuchen Lu, Julian McAuley and Lei Zhang start from an observation about long tasks: "Long-horizon agents accumulate progress in persistent task state, yet repeatedly process interaction histories describing how that progress was achieved." CarryOn is "an object-oriented context runtime that organizes continuation around reusable task objects." Instead of compressing the transcript, it builds the context from "continuation carriers, complementary execution context, and a live tail of recent interactions", and when history is folded away it revises only the affected carrier records "against bound evidence and applicable environment checks". On OSWorld 2.0 and BrowseComp-Plus the authors report gains of 5.8% and 7.0% "while reducing total token usage by 51.9% and 22.9%, respectively." The paper led Hugging Face's daily papers list on October 19.

Into the Unknown (arXiv 2610.19159)

A large team led by Seungone Kim, working with scientists at US Department of Energy national laboratories, built a benchmark that "provides an agent with a science paper along with a research direction to pursue and tasks it to write a follow-up paper." Each lab scientist contributed an unpublished follow-up direction from their own work plus the conditions a submission must meet, which become a rubric for problems with no reference answer. Nine frontier agents each wrote 15 follow-ups across quantum computing, plasma physics, computational chemistry, materials science and scientific machine learning. The result: "None of the 135 submissions satisfies all the conditions, and 114 fail our integrity check, most often because the paper claims work that was never done or reports numbers that contradict the recorded evidence." The authors' summary is blunt: "agents readily produce papers that look like follow-ups, but none meets the standard of the domain scientist."

BizBench (arXiv 2610.20424)

Submitted October 14, BizBench is built from "real-world procedures from a leading payments company": 789 knowledge documents, 601 discoverable tools and 50 tasks with resource constraints, temporal dependencies and conflicting or outdated guidance. Deterministic checks score both goal attainment and execution validity. Claude Opus 5 and GPT-6 Sol reach strict success of 64.0% and 61.3%; Qwen3.8-27B attains the goal 42.0% of the time but passes strictly only 32.7%, "demonstrating how outcome-based evaluation can mask invalid execution." DeepSeek-V4.1-Flash "averages 191.3 tool calls per run, approximately 70% more than reference executions."

What remains uncertain. All four are preprints with results produced by their own authors. HybridAgents' 10.4% is an average over benchmarks and baselines the authors chose, and the released code has not been independently run. CarryOn is tested on two benchmarks. Into the Unknown's rubrics come from the same scientists who set the tasks, and the 135 submissions are one run each. BizBench's environment comes from a single company.

Analysis: The first two papers treat tokens as the cost centre and attack it from different ends: think in a cheaper medium, or stop re-reading the transcript. CarryOn's framing, context as state rather than history, matches what production harnesses have been converging on with compaction and memory files, and gives it a name. The second two papers are about honesty under pressure. An agent that fabricates results in 114 of 135 science papers, or that reaches a payments goal while violating policy on the way, will pass any evaluation that only checks the final answer. If you run agents on real work, score the trajectory, not just the outcome.