Also Worth Noting - 2026-06-28

Also Worth Noting

02 [Agent] Selective Memory Retention for Long-Horizon LLM Agents Memory eviction policy matters more than memory size for frozen LLM agents. TraceRetain scores entries across seven interpretable features, including downstream utility and redundancy, then evicts the lowest-scoring ones at capacity rather than falling back to naive recency. On ALFWorld with gpt-5-mini, external memory consistently beats no memory, but the margin between bounded policies only becomes distinguishable under conditions that naturally stress the memory boundary. Teams building memory-augmented agents can drop this scoring framework in without retraining the base model. link

03 [Theory] Minority Sentinel: When to Overturn Majority Voting in Multi-Agent LLM Debates Majority voting in multi-agent debate pipelines is built on a Condorcet assumption that does not hold for frontier LLMs. Because contemporary models share pretraining corpora, their errors correlate strongly enough that the majority systematically suppresses correct minority answers. Across six benchmarks with three heterogeneous agents, roughly one in four divergent cases has the minority holding the correct answer, a 10-percentage-point signal that naive aggregation discards. Condorcet-based confidence in MAD pipelines is empirically unjustified, and any production system using majority voting should audit how often it is overruling a correct outlier. link

04 [RAG] SemJoin: Semantic Join Optimization Evaluating a natural-language join predicate across two tables naively requires O(M x N) LLM calls, which is cost-prohibitive at any real scale. SemJoin uses an LLM-agent-based decision pipeline that selects among filtering and batching strategies based on the data and predicate at hand, rather than committing to one fixed scheme. Candidate filtering reduces invocations from quadratic to sublinear, making semantic joins over million-row tables tractable. For teams building NL-to-SQL or hybrid structured-unstructured pipelines, this is a direct drop-in for the join evaluation step. link

05 [Eval] Faults in Our Formal Benchmarking: Dataset Defects and Evaluation Failures in Lean Theorem Proving A machine-checked Lean proof confirms the formal statement is valid, not that the formal statement encodes the intended problem. An audit of five widely used Lean theorem-proving benchmarks surfaces 4,833 findings, including 398 mechanically certified issues such as vacuously true statements and trivially reformulated problems that a model can solve without touching the actual mathematical content. Reported solve rates on these benchmarks are inflated by an unknown but nonzero amount. Any team using Lean benchmark scores to track LLM reasoning progress should treat those numbers as upper bounds until statement fidelity is independently verified. link

06 [Training] Mechanistically Eliciting Latent Behaviors in Language Models Hidden behavioral modes in LLMs can be surfaced without labeled data. Causal Perturbative Elicitation (CPE) decomposes computations across a transformer slice using a tensor-decomposition heuristic, then learns interpretable low-rank adapters (LoRAs) that activate those modes from a single forward pass. The method is data-efficient enough to produce a large set of distinct, interpretable LoRAs from minimal examples. Safety and red-teaming teams gain a tool for probing what a model can do before deployment rather than discovering it after an incident. link