Multi-step SWE
On the Multi-step SWE benchmark, the strongest of nine frontier models solves only 18.5% of 69 tasks.
Can Coding Agents Keep Building on Their Own Code Over Long Horizons?
AI Research Radar
Friday, November 6: 2 frontier papers out of 365 read.
Frontier
Papers rated as moving the field
How we chooseA Behavioural Framework for Predicting and Understanding Loss of Control in Frontier Artificial Intelligence Systems
A behavioural framework's loss-of-control score correlates with failure rates in contained long-horizon risk scenarios across 13 frontier models (Pearson r = 0.84).
This week in research
Oct 31 – Nov 6: 285 cleared, 18 frontier
- Security90◆ 3
- Agents49◆ 5
- Safety31◆ 5
- Robotics29◆ 3
- Society27
- Vision15◆ 1
- Inference14
- 4 moreTraining, Science, ML, Reasoning12
- Language9
- Security90◆ 3
- 8 moreSociety, Vision, Inference, Language, Training, Science, ML, Reasoning77◆ 2
- Agents49◆ 5
- Safety31◆ 5
- Robotics29◆ 3
Size: cleared papers◆ and tint: frontier papers9 papers not yet sorted into a field
Models under test
Papers this week that name the model
Labs publishing
Papers this week with an author at the lab
Also cleared
41 papers, by field
Security & privacy12
- Decision Models as Retrieval Gates Against Injection and PollutionGoogle
- Interactive Backdoor Attacks in LLM-Based Proactive Dialogue Systems
- SPILLWAY: Separating Target Admission from Sanitization to Defend Tool-Using Agents against Prompt Injection
- Breaking TraintoDefend via Adaptive Neuron Clustering in Cryptanalytic Model Extraction
- Dynamic Watermarking and Action-Conditioned Latent Dynamics for Secure Vision-Based Controllers
All 12 papers in Security & privacy
- JailBound: FOL-Guided Jailbreak Evaluation for Exploring LLM Behavioral Boundaries
- Embedding-Space Fingerprinting of LoRA Fine-Tunes: Recovering the Training-Data Domain from the Model Artifact Alone
- A Gingerbread House for LLM Agents: Source-Planned Honeypots That Lure Autonomous Attackers into an Isolated Shadow
- What the Middleman Delivers: Answer Behavior, Tool Permissions, and Billed Usage in LLM API Relays
- An Information Bottleneck for Post-hoc Memorization Mitigation in LLMs
- The Promise of Strategic Agentic Scambaiting
- Escalate or Close: How Fast-Tier Models Handle Held-Out Web Attack Families in Two-Tier Triage
Agents9
- The Memory Trap: Language Agents Overgeneralize from Benign InteractionsAnthropic
- A Multi-Modal Robotic Framework for Physical GUI Interaction Testing
- KG-AST for Cheaper CI/CD: Budgeted Test Selection and AI Bug Review
- The Last Mile to the Real World: Agent-Native Primitives as Reusable Building Blocks for Solving Real-World Tasks
- Signed but Not Safe: Requirement Integrity in Multi-Agent LLM Handoffs
All 9 papers in Agents
- DelegateEval: Evaluating Platforms for Personal External Service Agents
- Don't Prompt Your Agents. Equip Them: Skillware as a Modular Standard for Sovereign, Auditable, and Deterministic Capability Execution for Logical Systems
- Accepted but Wrong: Silent Failures of LLM Code Repairs after API Migrations
- Benchmarks as Software: Terminal-Bench and the Case for Continuous Benchmarks
AI & society5
- Reweighting is not Harmless: Understanding the Role of Dataset Characteristics
- Adversarial Multiplicity: Cherry-Picked Models and the Limits of Disclosure
- The Popular Get More Popular: Auditing Brand Bias in Generative Search
- Responsible GenAI Use During Active Armed Conflict: Evidence from Sudan's Emergency Response Rooms
- Who Can Evaluate AI Independently? A Framework for Analysing Evaluation Infrastructure