AI FRONTIER
Signals worth understanding before they become obvious.
LRMoE: Demand-Side Load Re-Calibration for Efficient MoE Training on Memory-Constrained Nodes
LRMoE is a memory-aware MoE training system that jointly recalibrates backward demand and balances the retained workload, using capacity-aware routing and physical compaction to reduce backward traffic, computation, and execution-buffer demand. The paper reports that on Qwen3-30B-A3B and GLM4.5-106B-A12B it reduces cum...
- The paper reports that on Qwen3-30B-A3B and GLM4.5-106B-A12B, LRMoE reduces cumulative MoE-layer training time to fixed-loss targets by 21.2-44.8% relative to SYMI.
- During forward execution, it plans retention quotas and expert placement under resource constraints while preserving all routed computations.
- During backpropagation, a load recalibration metric estimates each token-expert pair's contribution to the expert update and selects pairs within these quotas.
- A capacity-aware routing algorithm distributes retained pairs among eligible instances of their original experts, and physical compaction forms smaller communication payloads and variable-size grouped GEMMs.
For teams training sparse MoE on memory-constrained GPU nodes, LRMoE shows that selecting token-expert pairs by contribution during backpropagation can cut backward traffic and buffer demand without changing forward routed computations.
Detecting Covert Collusion between AI Agents through Partner-Outsider Positive Listening
The paper proposes a behavioural test, Partner-Outsider Positive Listening (PoPLi): a partner using a secret code reacts to the exact wording of a message, whereas an uninvolved agent placed in its seat does not. PoPLi is the difference between the two agents in how much their action changes when the message is reworde...
- Message-only monitors did not separate the conditions: one reading the speaker's messages answered "no" in all 50 markets, and one reading every seller's messages flagged 9 of 10 markets with no collusion.
- PoPLi detects the use of a private channel, not collusion as such, and it cannot detect collusion through incentives alone or codes carried by meaning.
- The method is computed after the fact by replaying logged prompts, and the paper is marked as work in progress.
For builders of multi-agent systems, PoPLi offers a log-replay behavioural audit that can surface a private channel behind innocent-looking messages, while leaving incentive-only and meaning-carried collusion out of scope.
WebAnything: From the Wild Web to Scalable Agent Environments
WebAnything is a fully automated framework that reconstructs source-grounded, reusable Web environments from real URLs without per-site manual implementation or repair after initial setup and authorization. The paper reports constructing 244 website artifacts across 14 application domains and 53 categories, supporting ...
- The paper reports constructing 244 website artifacts across 14 application domains and 53 categories, supporting 12,984 executed and accepted tasks.
- Compared with human-assisted WebHarbor, WebAnything achieves higher aggregate scores on three fidelity metrics, with 12 wins, two ties, and no losses in a blind audit.
- An Explorer-Builder collects browser-observable structures, resources, and interaction outcomes, while an independent verifier checks frozen candidates and guides local repair or targeted source recapture.
For agent builders, this shows that resettable Web training and evaluation environments can be generated automatically from real URLs without per-site engineering, with reported fine-tuning gains quantified on specific benchmarks.
ThinkJev: One Thought, Every Budget for Calibrated Typed Decisions
ThinkJev builds a fast and a slow system from one open hybrid-thinking model: System 1 reads the option-letter probabilities when the model answers directly, and System 2 reads the same letters after the model has thought on the same cached prompt, so escalating repeats no work and one sampled thought yields the answer...
- At a 5% target error rate, Qwen3.5-27B answers 24.5% of 8359 test decisions with System 1 alone at a realized error of 3.9% (Qwen3.5-4B and 9B: 6.5% and 0.0%).
- For the remaining decisions, stopping each thought as soon as its answer is confident reaches 81.6% accuracy versus 82.7% for always thinking, using 72% of the thought tokens.
- The paper states no training is required: a statistical test on labeled dev data keeps accepted System 1 answers under a target error rate, and code and data generators are available.
For builders of typed-decision services, this suggests reusing one cached prompt and its letter probabilities on a single open hybrid-thinking model, then deciding per decision whether to think under a statistical guarantee, to cut thought-...
Measuring Where Intent Gets Lost in Agentic Smart Contract Synthesis
The paper designs an execution-based evaluation that replays the same transactions on an AI-written contract and the original it should reproduce to measure how often they agree, and proposes an elicitation-first architecture that asks the client a short checklist before writing code.
- The paper reports that checking only the functions every system includes hid about half the difference between pipelines.
- The paper reports that asking first mattered far more than reviewing or patching afterwards across three AI models, while automated security fixes often break requested behavior.
- The paper reports that common methods such as AI judges and standard analysis tools miss obvious mistakes, and scoring choices matter too.
- The authors release their tools, 600 tasks and all results; the paper was accepted at ACM TheWebConf2027.
For agentic code generation, elicit missing requirements before generation and validate against an independently written executable answer key rather than relying on post-hoc review or automated security patching.
OpenFrontiers: A Collection of Open Research Problems and a Benchmark for Scientific Rediscovery
OpenFrontiers combines a structured corpus of 12,151 open research problems with a retrospective benchmark of 200 mathematics problems, 110 resolved in 2026 and 90 still open, each carrying a Lean 4 statement checked by automated probes and two independent reviews.
- The corpus records problem statements, context, progress, and other information for 12,151 problems.
- The benchmark contains 200 problems, 110 resolved in 2026 and 90 still open, each with a Lean 4 statement checked by automated probes and two independent reviews.
- A subset HQ-50 further filters based on quality score, etc.
- The evaluation protocol grades formal proofs with two independent checkers in an isolated sandbox under closed-book, frozen-retrieval, and live-web conditions.
Because the benchmark separates formal grading across closed-book, frozen-retrieval, and live-web conditions, builders can distinguish genuine rediscovery from retrieval of published solutions.