Harness intelligence, world models, and auditable multi-agent systems

Guibin Zhang, Leo Lu, Fangzhou Xie, Kang Zhu, Junhao Wang, Zhifei Xie, Zhaochen Yu, Zihang Liu, Zhongxiang Sun, Qiankun Li, Yue Liao, Heng Chang, Xiaobin Hu, Qibing Ren, Wangchunshu Zhou, Shuicheng Yan

The authors formalize the agent harness (memory, planning, action protocol, tool orchestration) as a machine-generatable four-module artifact and train a dedicated model that synthesizes, repairs, and evolves task-adaptive harnesses for off-the-shelf LLMs. With JIT-Agent, weaker models surpass stronger baselines on DeepSearchQA and OdysseyBench, and generated harnesses match mature runtimes such as OpenCode while transferring across model families. The work positions harness intelligence as a trainable dimension of capability orthogonal to foundation-model scaling.

Subhadeep Pal, Fiona Y. Wang, Markus J. Buehler

Homogeneous LLM agents self-organize in a spatial environment without assigned roles, processing resources, building persistent artifacts, and writing executable controllers that are later stress-tested by a deterministic simulator after the agents are removed. Shared societies produce broader, more resilient technological portfolios than strong isolated best-of-N search, with agents spontaneously differentiating into exploration, construction, maintenance, and coordination roles. Physical stigmergy alone suffices for capable societies, while explicit cultural mechanisms amplify organization over longer timescales.

Yiwen Chen, Guosheng Lin, Chi Zhang

The framework separates rule-consistent world evolution from high-fidelity visual rendering: a coding agent maintains executable state and generates code updates, which are compiled into a proxy video that conditions a video model. Data pipelines produce aligned proxy-observation pairs from gameplay and real video. After fine-tuning, the system follows coding-agent specifications while preserving visual detail, offering a path to open-ended world models with persistent consequences.

Jiaming Zhou, Qihang Zhang, Gangwei Xu, Cunxin Fan, Yujie Zhao, Ruilin Wang, Yiming Luo, Shuai Yang, Xing Zhu, Yujun Shen, Junwei Liang, Yinghao Xu

Zero-WAM treats human video as the natural in-context task specification for robotic manipulation and trains a causal video-action model to follow it without parameter updates. An automatic pipeline converts robot trajectories into matched human videos (HumanGen: 74.2K pairs, 8.6K tasks), and an in-context future-chunk prediction objective forces the policy to use the video prompt rather than shortcuts. On seven unseen RoboTwin tasks it reaches 47% average success (plus 29.5 points over the strongest baseline) and generalizes to real multi-object and insertion settings.

Sheng Liang, Yongyue Zhang, Nathanael Brian, Hang Lv, Hao Wang, Chen Zhang, Yong Liu

Standard speculative decoding assumes identical context for drafter and verifier, which collides with the need to compress long agentic contexts. AsymSpec lets a lightweight drafter read the full input while the large verifier sees a compressed view, steering via contrastive logit fusion and a divergence-aware acceptance gate. It recovers about 90% of full-context accuracy at 1.3-1.7x throughput and substantially lower compute on agentic benchmarks.

Songyuan Li, Ahmed M. Abdelmoniem, Shiqiang Wang

Instead of one-shot cascade routing, ProgRouter scores multi-view task progress (outcome regimes, subtask completion, trends, state quality) and uses dual-path prediction plus meta-gating to choose the next LLM step under time and cost budgets. Online decisions balance immediate progress gain against long-term operating cost. On code, math, and long-form RAG workflows it cuts cost relative to strong baselines while preserving solution quality.

Srimonti Dutta, Akshata Kishore Moharir

Answer accuracy alone is insufficient because a correct final answer can rest on an invalid computational trace. The authors define Trace Integrity via execution contracts that bind intent, schema, operators, queries, and answer linkage, and introduce the CAIT rate measuring silent failures. On BIRD Mini-Dev, answer accuracy stays low while CAIT rates remain high (45-59%), showing that answer-only metrics systematically overcount success.

Zhongwen Luan, Xiaoyu Zhang, Ming Hu, Yue Yang, Jiongchi Yu, Xiaohong Chen

SymTrace freezes execution before failure anchors so only the downstream trajectory is regenerated, enabling causal rather than stochastic diagnosis. On the annotated SymFail set of 536 trajectories, unguided reruns reproduce failures only 68% of the time and repair just 6.9%. A symptom-driven intervention method raises repair rate to 20.15%, a nearly threefold gain over prior repair techniques.

Paul Rosu, Rowan Wang

LLM auditors are trained with RL in environments that plant hidden behaviors in target system prompts; a judge that knows the ground truth supplies pairwise rewards against reference investigations. Pairwise rewards and clean-target negatives keep false positives below 1% while improving investigation quality, realism, and detection of concerning behaviors in production models. Gains transfer to adversarially fine-tuned targets on AuditBench.

Fredrik Rømming, Mantas Bakšys, Martin S. Fixman, Sean B. Holden

Proof construction in clausal connection tableaux is cast as a policy over a calculus-induced transition system, scored by a graph neural network trained via imitation on found proofs. Removing symbolic backtracking scaffolding still yields policies that solve up to 46% more problems than leanCoP on standard suites and reach proofs in an order of magnitude fewer steps. The result shows structure-transferring learned policies can largely replace search scaffolding inside a formal calculus.