Agentic Perception, RL Optimization, and Bayesian Inference Advances

Today’s selection of 8 noteworthy AI/ML papers from arXiv, covering semifactual credit assignment for LLM reasoning, disagreement-guided multi-agent evaluation, minimal-harness ML engineering agents, agentic visual perception, LLM-driven LiDAR attack planning, RL-guided sparse solver optimization, exact incremental BPE, and amortized Bayesian inference for multilevel models.

1. Semifactual Credit-Augmented Policy Optimization

Authors: Junshu Pan, Zhizhang Fu, Shulin Huang, Yiran Ding, Zifan Cheng et al. | Categories: cs.LG, cs.AI, cs.CL Link: arxiv.org/abs/2609.40360

The authors show that LLM reasoning is sensitive to task-irrelevant prompt features, and that suppressing high-drift token candidates during decoding improves accuracy without weight updates — exposing a limitation of GRPO’s uniform token-level advantage assignment. They introduce SCAPO, a GRPO variant that measures token probability drift under semifactual interventions and down-weights advantages for unstable tokens during early training. On Qwen3-4B-Base and Qwen3-1.7B-Base, SCAPO beats GRPO by 5.63 and 4.17 points on AIME 2024-2026, with the best results on most math and all out-of-distribution benchmarks tested.

Takeaway: A clean causal story turned into a practical fix: reward tokens that stay stable when irrelevant prompt details change, and reasoning generalizes better. The gains at both model scales make this a drop-in idea worth testing.

2. JuryFlow: Disagreement-Guided Human-in-the-Loop Multi-Agent Evaluation

Authors: Mufeng Yang, Junwei Yu, Yepeng Ding | Categories: cs.CL, cs.HC Link: arxiv.org/abs/2609.40103

JuryFlow treats inter-judge disagreement as a precise, claim-level signal rather than noise, decomposing responses into atomic claims and building a disagreement graph scored by verdict entropy. A human acts as a structural guide — resolving a single focal disagreement — after which the correction propagates along graph edges and crystallizes into reusable rubric entries all judges inherit. On MT-Bench and LLMBar, JuryFlow improves agreement with gold labels over single-judge and majority-vote baselines, with ablations isolating the contributions of targeted re-evaluation, propagation, and rubric induction.

Takeaway: Majority voting throws away exactly the information you most need. Reframing the human as a structural guide rather than a labeler is a smart, scalable answer to the LLM-as-judge reliability problem.

3. How Much of a Harness Does a Strong Agent Need for Autonomous ML Engineering?

Authors: Kirill Brilliantov, Alejandro Hernández-Cano, Emmanuel Abbé | Categories: cs.AI Link: arxiv.org/abs/2609.40303

Under equal time budgets and the same frontier LLM backbone, the authors find that state-of-the-art open-source MLE agent harnesses provide no advantage over a single minimal-harness coding agent with direct read, write, and bash access. Large-scale systematic ablations suggest multi-agent orchestrators and retrieval subagents become redundant in the coding-agent setting. They conclude that hand-crafted harness machinery yields poor returns on current MLE benchmarks, with the backbone as the primary performance driver.

Takeaway: A useful corrective to harness complexity creep. If your elaborate orchestration layer isn’t beating a single bash-enabled agent, the effort is better spent elsewhere — though the conclusion is scoped to today’s MLE benchmarks.

4. EviRover: Reinforcing Agentic Perception Beyond a Glance

Authors: Kaixuan Fan, Kaituo Feng, Tianshuo Peng, Yilei Jiang, Manyuan Zhang et al. | Categories: cs.CV, cs.AI Link: arxiv.org/abs/2609.40230

The authors formalize “perception under insufficient evidence” — queries that one-shot visual prediction cannot resolve — as an agentic process that gathers information beyond a single glance. They build dedicated data pipelines (EviRover-SFT-5K, EviRover-RL-12K) and EviLens, a 688-instance human-verified benchmark, then train EviRover with SFT followed by agentic RL. The 4B model outperforms its backbone by 30 points on average on EviLens, rivaling proprietary models, with gains transferring to WebEyes, conventional perception benchmarks, and a 15-point improvement on BrowseComp-VL.

Takeaway: Moving perception from a single forward pass to an interactive evidence-gathering loop is a compelling reframe, and the transfer results suggest the skill generalizes rather than overfitting the new benchmark.

5. TACTIC: Temporal and Context-Aware LLM Tactical Planning for Roadside LiDAR Attacks

Authors: Yiming Gao, Shaocheng Luo | Categories: cs.RO, cs.AI, eess.SY Link: arxiv.org/abs/2609.39969

TACTIC uses a multimodal LLM to coordinate state-adaptive roadside LiDAR attacks under a gray-box threat model, relying only on an attacker-operated perception stack and roadside imagery to build a semantic scene graph. The system selects and configures two primitives — push-away range shifting and phantom-obstacle braking — grounded in physically feasible operating regions, with asynchronous reasoning/execution overlap to handle MLLM latency. Across 280 randomized CARLA trials, the full policy achieves a 100% collision rate versus 35% for a fixed rule and 75% for a restricted LLM, with joint physical-and-image input reaching 100% success versus 65% and 75% for single modalities.

Takeaway: A sobering security result: scene-aware tactical planning exposes context-sensitive LiDAR failure modes that fixed attack policies simply miss. The modality ablation is the key evidence that traffic context, not just measurements, drives success.

6. Reinforcement Learning-Guided Graph Transformations for SpTRSV Optimization

Authors: Buse Yılmaz | Categories: cs.DC, cs.LG Link: arxiv.org/abs/2609.40159

Sparse triangular solve suffers from data dependencies that limit parallelism, and existing graph transformation heuristics are manually designed and hard to adapt. This work formulates graph transformation as a sequential decision-making problem where an RL agent learns matrix-dependent policies. On real-world sparse matrices, the approach achieves average reductions of 23% in level count and 29% in the coefficient of variation of level costs while rewriting only 0.82% of rows, with policies transferable to unseen matrices via curriculum learning and fine-tuning.

Takeaway: The RL approach trades slightly less aggressive level reduction for the best balance of competing objectives — and does so while touching under 1% of rows. The zero-shot generalization analysis is refreshingly candid about where transfer breaks down.

7. Component-Weighted Centroid Search for Exact Incremental BPE

Authors: Harshit Verma, Rex Ying | Categories: cs.DS, cs.LG Link: arxiv.org/abs/2609.40016

Building on Jiang and Gong’s (2026) exact incremental BPE algorithm, the authors modify only the centroid search, weighting each interval by the size of the recursive component it selects so moves telescope to O(log t) time per appended byte and O(n log t) over an n-byte stream, with identical BPE semantics and asymptotic space. They construct a normalized proper BPE family where count-balanced search requires Θ(log² t) probes while weighted search needs Θ(log t), and a Rust implementation matches the predicted probe counts.

Takeaway: A precise, well-scoped theory improvement — and notably honest that on ordinary vocabularies the queried degrees are small, making this a worst-case guarantee rather than an average-speed win.

8. Amortized Bayesian Inference on Multilevel Models of Arbitrary Structure

Authors: Daniel Habermann, Andreas Bulling, Stefan T. Radev, Paul-Christian Bürkner | Categories: stat.ML, cs.LG, stat.CO Link: arxiv.org/abs/2609.40024

The authors present a general method for amortized Bayesian inference on multilevel models of arbitrary structure, automatically deriving valid posterior factorizations and matching neural network architectures from a DAG-specified generative model. Graph expansion and inversion produce an inverse graph determining how inference networks are stacked and conditioned, amortizing over group counts and observations per group while preserving all conditional independence and exchangeability assumptions. Across three case studies, the method closely matches gold-standard samplers on models with more than 6,500 parameters while reducing inference to a near-instant forward pass after training.

Takeaway: Refusing to simplify the dependency structure is what sets this apart from faster-but-lossy alternatives, and matching gold-standard samplers at this scale makes it a genuinely practical tool for hierarchical modeling.

This content was generated with AI assistance. Paper information sourced from arXiv.