AI News Brief Roundup: 2026-10-28 Morning

AI News Brief: Morning roundup, 2026-10-28

Summary

This morning's feed was dominated by research on how to make LLM agents better, and on how to measure them honestly. Several papers tackle agent self-improvement, memory and evolution, while others show that headline scores can hide real failures, from radiotherapy planning decisions to vending-machine economics and chain-of-thought faithfulness. Retrieval, reasoning-training and evaluation methods also got fresh attention. On the industry side, Amazon's guidance on AI tooling, Anthropic's reported change to its coding interviews, and Legora's use of OpenAI models showed how fast AI is reshaping work. For readers, the common message is to judge systems by their decisions and real-world outcomes, not just benchmark numbers.

1. Researchers used AlphaCNOT to find a cheaper way to swap distant qubits under limited connectivity, needing at most 6d-4 CNOT gates versus 6d-3 for SWAP routing. They turned it into a provable synthesis method, showing AI can reveal reusable compilation principles.
2. OpenGray is a benchmark for radiotherapy planning agents with known correct decisions. It found good aggregate scores hid failures: agents rarely escalated impossible dose goals, and 24% of feasible plans missed a hard goal, suggesting agents need an explicit stop-and-review action.
3. Guided-RLVR adds a lightweight controller that injects steering tokens during sampling, countering the collapse of RLVR rollouts onto narrow reasoning paths. It outperforms existing RLVR methods on math and code benchmarks while producing more diverse reasoning.
4. This paper shows that evaluating chain-of-thought faithfulness only on correctly answered cases biases results. Across four deployments the selection gap was positive (+7.4 points in the main run), so it recommends evaluating all traces where possible.
5. SpectRAG builds a weighted graph of document chunks and uses spectral propagation to pick context within a token budget. It improves evidence recovery for multi-hop and cross-document questions, though with trade-offs in early precision and compute cost.
6. Natural Hard-Negative Mining turns original positive passages into hard negatives under altered instructions, creating a roughly 7K-instance training set. Models from 568M to 4B parameters trained on it match or beat much larger retrievers at following instructions.
7. Dwell reframes a vision-language model as its own latent world model, letting agents learn from imagined outcomes. It performs best among trainable methods across five visual environments while cutting imagination compute, since predicting action-useful futures suffices.
8. A philosophical review questions claims that LLMs have "functional" pleasure, pain or emotions. It argues key aspects of valence remain untested and current experiments fail to separate valence from preferences, value judgments and role-play.
9. An Amazon engineering leadership quote says that if AI tooling doesn't help an engineer's work, documenting why is a legitimate finding and shows good judgement. It signals a more nuanced stance on AI adoption.
10. ProactMem organizes agent history into recent events, trajectories and verified behavioral regularities, so agents can anticipate needs. On the new ProactiveStream benchmark it reaches 52.6% F1, a 21.4% relative gain over the best baseline.
11. Legora, a legal-tech company, uses OpenAI models, ChatGPT and Codex to speed up legal workflows and contract analysis. OpenAI presents it as an example of an agentic operating system freeing lawyers for higher-value work.
12. AutoVend tests LLM agents running 20 vending machines for 180 simulated days. Four of five agents fell short of a simple rule-based baseline, as inventory and execution costs eroded gains, exposing weaknesses in long-horizon economic decision-making.
13. GUI-JEPA picks mobile GUI actions in latent space without decoding tokens. It runs 2.8-4.0x faster than a fine-tuned Qwen2.5-VL-7B agent while keeping 86-92% of its accuracy with only 6% of the parameters.
14. Noise Steering uses reinforcement learning to control wind turbine yaw and power so noise limits near protected areas are met. It loses only about 2% power on average versus 17% for the industry-standard noise-reduced operation.
15. Evolving Agent Teams uses a supervisor that restructures coding agents when they hit local optima. It produced a CUDA kernel 1.2x faster than FlashInfer's production decode kernel and showed gains on AMD, NVIDIA, Apple silicon and robotics tasks.
16. This paper proposes a policy-iteration algorithm for zero-sum partially observable games that updates policies using a single supporting hyperplane rather than a full value function. It often improves runtime-exploitability trade-offs against PBVI, HSVI and CFR+ at longer horizons.
17. Reporting from Cisco Partner Summit in Vancouver says the event shifted focus from selling Cisco technology to how partners make money in the AI era, as transactions stop being the main event. Cisco arrived with strong momentum.
18. RubricAtlas builds type-level evaluation rubrics from 3,079 Chinese financial reports across 99 research types, to grade deep research agents. It shows single reports cover only about half of a type's questions, and the judge pipelines agree well on rankings (tau 0.87).
19. PIP-ToM improves theory-of-mind reasoning by admitting only evidence accessible to the relevant agent, avoiding perspective leakage. With GPT-4o it reaches roughly 92-95% accuracy on MuMA-ToM, MMToM-QA and Hi-ToM.
20. This framework represents mathematical arguments as source-grounded dependency graphs with measures like depth and bulk, using LLMs to propose links for human review. The authors offer an evaluation design rather than empirical evidence of benefits.
21. Extending the GEPA optimizer to propose and evaluate batches of candidates per step cut wall-clock time by half or more. Held-out scores also rose, for example from 49.0% to 60.0% on HoVer.
22. RuHarness co-evolves evaluation rubrics and agent harnesses, continually reassessing behavioral weaknesses from execution trajectories. It beats harness-evolution baselines, especially on unseen tasks, because supervision keeps pace with the agent's changing flaws.
23. A Latent Space newsletter item reports that Anthropic is ending coding interviews, with the tagline "hunger beats tenure". The framing is that if coding is solved, testing it in hiring no longer makes sense.
24. A lightweight knowledge-graph "learner digital twin" supports early-warning alerts in education. It keeps predictive performance of raw-table features while linking each alert to attributions, evidence items and raw platform records for traceability.
25. Population-Centered Evolution improves self-evolving agents by drawing on successes and failures across the whole population, not just a chosen parent. It reaches 72.1% on SWE-bench Verified and 96.7% on Polyglot, beating a group-evolution baseline.

Summaries written with AI (Claude by Anthropic) from the linked source articles.

Sources in this roundup

Most-mentioned keywords

Sources

1. AI-Assisted Discovery of CNOT Decompositions for Quantum Permutations on Constrained Topologies
2. OpenGray: a benchmark for evaluating the decisions of radiotherapy planning agents
3. Guided-RLVR: Guided Exploration in Reinforcement Learning with Verifiable Rewards
4. Correctness Conditioning as Selection in Chain-of-Thought Faithfulness Evaluation
5. SpectRAG: Weighted Chunk Graphs for Token-Budgeted Context Selection in Multi-Hop Retrieval-Augmented Generation
6. Natural Hard-Negative Mining from Positives for Instruction-Following Retrieval
7. Vision-Language Agents Can Learn in Their Own Latent Worlds
8. Do bots feel bad? Assessing empirical evidence for functional valence in LLMs
9. Quoting Terk Terkowitz
10. Proactive Memory for Event-Driven LLM Agents
11. Legora is building the agentic operating system for legal work
12. AutoVend: Benchmarking LLM Agents on Long-Horizon Autonomous Vending Operations
13. GUI-JEPA: Candidate-Conditioned JEPA Latents as a Fast Action Policy for Mobile GUI Agents
14. Noise Steering: Exploiting Wake Steering and Acoustic Directivity for Noise-Constrained Wind Farm Control via Reinforcement Learning
15. Evolving Agent Teams
16. Conservative Policy Iteration For Zero-Sum POSGs
17. Cisco Partner Summit: Five insights that will define the channel in the AI era
18. RubricAtlas: Mining Type-Level Rubrics from Expert Reports to Evaluate Deep Research Agents
19. Who Knows What – and What Does It Support? Perspective-Bounded Evidence Calibration for Multimodal Theory of Mind
20. Dependency Graphs of Mathematical Arguments: Representation, Structural Measurement, and Proof Review
21. Batching the Reflective Optimization Loop: Parallel Proposals Make GEPA Faster and Better
22. When the Critic Falls Behind: Co-Evolving Rubrics and Harnesses for Continual Agent Self-Improvement
23. [AINews] Anthropic ends Coding Interviews: "Hunger beats tenure"
24. An Incremental Knowledge-Graph Learner Digital Twin for Evidence-Grounded Early Warning
25. The View from Nowhere: Population-Centered Agent Evolution