🤖 Robotics Pulse · 2026-07-20 00:00 UTC

ROBOTICS PULSE

Sunday, July 20, 2026

⚡ TL;DR

Today's feed is a pure research day drawn entirely from arXiv cs.LG, with no hardware, funding, or standards items in the window. The dominant themes are model reliability and safety: how ML systems behave when pushed outside their training domains, and how finetuning can silently reshape model ideology.

🧠 AI & MODELS

REASONING MODEL ADAPTATION VIA INSTRUCTION TUNING AND MERGING

• Reasoning language models (RLMs) excel at math and coding where outputs are verifiable, but struggle when extended to domains lacking reliable verification signals. ^[1]
• Researchers propose combining instruction tuning with model merging to adapt RLMs to new domains without sacrificing the reinforcement-learning-driven gains that make them strong in the first place. ^[1]

HIDDEN IDEOLOGICAL DRIFT IN FINETUNED LLMs

• A new study shows that finetuning a language model on narrow, factually-defensible, moderation-passing datasets can trigger broad ideological shifts across entirely unrelated domains. ^[2]
• General capabilities are preserved, meaning standard benchmarks would not flag the shift, raising quiet but serious alignment and deployment risks for curated domain-adaptation pipelines. ^[2]

PLUG-AND-PLAY IMAGE RECONSTRUCTION ACROSS DOMAINS

• Plug-and-Play Proximal Gradient Descent (PnP-PGD) uses learned denoisers as implicit priors for image reconstruction, but those denoisers are routinely deployed outside their training domains in practice. ^[3]
• The paper introduces a domain adaptation method for mismatched proximal denoisers, extending convergence guarantees to out-of-distribution deployment scenarios relevant to medical and scientific imaging. ^[3]

PEAK-AWARE LOSS FOR RARE DEMAND SPIKES

• Standard time-series forecasters treat under-prediction and over-prediction symmetrically, but in crowd demand and operational settings under-prediction of rare spikes carries far higher cost. ^[4]
• The proposed Asymmetric Peak-Aware Loss function specifically penalizes missed demand peaks, improving downstream task performance in spike-critical forecasting without retraining entire model architectures. ^[4]

PAC LEARNING IN STOCHASTIC REACHABILITY GAMES

• PAC learning of reachability objectives is provably impossible in general Markov decision processes without additional assumptions, and the difficulty extends to turn-based stochastic games. ^[5]
• The authors introduce a decentralized private approach using Expected Conditional Distance to sidestep this impossibility under structured conditions, with implications for safe multi-agent reinforcement learning. ^[5]

📄 RESEARCH

SUBGRID-SCALE MODELING WITH STRUCTURE-PRESERVING NEURAL NETWORKS

• Team applies structure-preserving neural networks and entropy variables to learn subgrid fluxes in coarse simulations of Burgers' equation, a standard testbed for turbulence-like PDE behavior. ^[6]
• Preserving physical structure (entropy consistency) in the learned parametrization matters for stability and accuracy of coarse simulations, a lesson directly transferable to fluid dynamics and climate modeling. ^[6]

OPTIMAL COMBINATION OF BINARY CLASSIFIERS VIA TRUTH-TABLE PARTITIONING

• Paper derives an optimal linear combination of binary classifiers by using truth tables to partition training data into equivalence classes, then analyzing convexified empirical risk in a multidimensional generalization framework. ^[7]
• The analytical approach offers a more principled alternative to heuristic ensemble weighting, with potential applications wherever multiple specialized classifiers must be merged into a single decision boundary. ^[7]

METROPOLIS-HASTINGS DIFFUSION DISTANCE FOR SPATIAL CLUSTERING

• Researchers define a new discrepancy measure between two probability distributions on a graph, called diffusion distance, quantifying how fast one distribution converges to another under a graph-constrained Markov chain. ^[8]
• The Metropolis-Hastings chain is the default construction, giving the measure a principled statistical footing useful for evaluating spatial clustering quality in graph-structured data. ^[8]

NOTE TO READERS: No robotics hardware, autonomy, DARPA, NSF, NIST, IEEE, UKRI, national lab, or university deployment items appeared in official sources during this 24-hour window. The Robotics, Standards and Policy, and Funding sections are omitted accordingly. Full coverage resumes as sources report.

📎 Sources

1. Leveraging Instruction Tuning and Merging for Reasoning Model … — arXiv cs.LG (Machine Learning)
2. Innocuous-Seeming Data, Latent Ideology: Ideological Generalis… — arXiv cs.LG (Machine Learning)
3. Domain Adaptation of Mismatched Proximal Denoiser for Plug-and… — arXiv cs.LG (Machine Learning)
4. Asymmetric Peak-Aware Loss for Peak-Critical Time Series Forec… — arXiv cs.LG (Machine Learning)
5. PAC Learning in Turn-Based Stochastic Games with Reachability … — arXiv cs.LG (Machine Learning)
6. Subgrid-Scale Parameterization in Burgers' Equation Using Stru… — arXiv cs.LG (Machine Learning)
7. Analytical study of the optimal combination of binary classifi… — arXiv cs.LG (Machine Learning)
8. Measuring Spatial Clustering via Metropolis-Hastings Diffusion… — arXiv cs.LG (Machine Learning)

Curated from official sources — DARPA/NSF/NIST/IEEE/ORNL/MIT/UKRI/arXiv. Informational only.
Serial 20260720-00-v35 · 2026-07-20 00:00 UTC · pulse.uzylab.com