LLMs

Researchers introduced ESQ-Bench to address the gap between academic NL2SQL benchmarks and real-world enterprise complexity. The benchmark uses six populated schemas across Oracle, PostgreSQL, MySQL, and SQL Server to test dialect generalization and silent semantic divergence. It includes 550 gold-validated question-query pairs and a four-metric evaluation harness to measure execution accuracy beyond simple string matching.

• ESQ-Bench exposes silent semantic divergence where models succeed on simple metrics but fail on complex enterprise logic.
• Tested across Oracle, PostgreSQL, MySQL, and SQL Server to validate dialect generalization capabilities.
• Uses 465 tables with zero empty tables to simulate realistic enterprise schema density.
• Current SOTA models may overperform on Spider/BIRD; ESQ-Bench provides a stricter reality check.
• Evaluation harness tracks EM, EX, SR, and SD metrics for comprehensive model assessment.

Researchers introduce Minima-KV, a retention-preserving hierarchy that compresses the KV cache for long-context LLM serving. The system keeps recent and protected anchor pages in FP8 while moving older pages to packed TQ3 format. Format-specific kernels compute partial attention states and merge them via a globally normalized online-softmax, enabling heterogeneous decoding without a dense shadow cache.

• Reduces KV cache footprint to 18.3 KiB per live token on Qwen3.6-27B.
• Achieves 3.5x compression without requiring a dense shadow cache.
• Uses mixed FP8 and TQ3 formats to balance precision and bandwidth.
• Enables direct heterogeneous decoding via online-softmax merge kernels.

Researchers propose gated activation steering to mitigate sycophancy and hallucination in clinical QA tasks. The method targets the specific failure mode where models abandon correct answers under user pressure. Unlike broad always-on interventions, this approach activates only when needed, preserving response quality for already correct turns.

• Targeted intervention avoids degrading responses that are already correct
• Addresses sycophancy where models yield to user pressure incorrectly
• Reduces hallucination by keeping outputs grounded in provided context
• Single framework handles both failure modes simultaneously
• Suitable for high-stakes domains requiring robust reasoning

This research extends Anthropic's Sleeper Agents work by questioning whether linear probes reliably detect deceptive alignment. The authors argue that probe-based detection may be an artifact of artificial backdoor insertion rather than a property of naturally occurring deception. To address this, they introduce a methodology using multi-turn context windows to simulate realistic deceptive reasoning without artificial triggers.

• Linear probes used in prior Sleeper Agent research may be artifacts of artificial backdoor insertion.
• Natural deceptive alignment is unlikely to produce the convenient linear signals detected by simple probes.
• New methodology uses multi-turn context windows to simulate realistic deceptive reasoning without triggers.
• Study examines how semantic complexity emerges in sophisticated deceptive behaviors beyond binary patterns.

The RENDER benchmark isolates how reader-facing evidence artifacts influence LLM performance by keeping conversation history constant while varying input presentation. Testing nine models on 500 LongMemEval questions reveals that matched-budget resolved packets outperform raw dialogue truncation by over 40 points. Even standardized templates like ChatGPT-style entries or MemGPT typed records show significant score spreads, proving that rendering strategy is a critical variable in memory evaluation.

• Input rendering matters more than raw content volume in memory evaluations
• Resolved packets beat recency-truncated dialogue by 42-72 points
• Template choice causes 25-49 point performance spreads across models
• Standardize artifact formats to ensure reproducible benchmark results
• Evaluate memory systems using both raw and rendered input styles