LLMs

Amazon Bedrock in AWS GovCloud (US) has made OpenAI's GPT-5.6 Terra and Luna models generally available. Terra offers GPT-5.5-level intelligence at half the cost, while Luna provides fast, cost-efficient inference. These models run on Bedrock's next-generation inference engine designed for high performance and security.

• GPT-5.6 models are now GA in AWS GovCloud US-West and US-East regions.
• Terra delivers GPT-5.5 performance at half the cost for complex reasoning tasks.
• Luna enables fast, affordable inference for high-volume, low-latency use cases.
• Models support autonomous coding, genomics analysis, and cybersecurity research.
• Bedrock's new inference engine optimizes for security, reliability, and speed.

KVBoost addresses the limitation of existing prefix-caching systems that fail when shared content appears at non-contiguous positions. It uses a dual-hash keying scheme to separate positional identity from content identity, allowing exact and approximate cache matches for independent chunks. To mitigate attention boundary errors resulting from this independent caching, the system employs deviation-guided recomputation. This approach significantly reduces prefill latency for HuggingFace-compatible decoder models by maximizing cache hit rates regardless of prompt structure.

• Eliminates the contiguous prefix requirement, enabling cache hits for scattered shared content.
• Dual-hash scheme decouples position from content, supporting flexible chunk matching.
• Deviation-guided recomputation corrects attention errors from independent chunk caching.
• Reduces prefill latency for HuggingFace decoder models with arbitrary prompt structures.

This research addresses the limitation of current credit assignment methods in LLM reinforcement learning, which typically use architecture-agnostic transport operators like fixed-discount GAE or group-relative statistics. The authors propose Computation-Conditioned Credit Transport (CCT), a framework that aligns the transport of credit with the specific trajectory-based computation used by the Transformer policy. By conditioning credit assignment on the model's internal computational structure, the method aims to provide more accurate token-level advantages than stationary kernels or broadcasted outcome statistics.

• Current LLM RL credit transport ignores the Transformer's specific computational trajectory.
• CCT conditions credit assignment on the policy's internal computation structure.
• Replaces fixed-discount GAE and group-relative methods with architecture-aware transport.
• Aims to improve token-level advantage estimation for better policy updates.

Research reveals that LLM leaderboards are heavily influenced by evaluation harness details rather than pure model capability. By treating the harness as an independent variable, the study isolates specific benchmark items that cause significant score variance across 12 instruction-tuned models. This 'fragility grid' analysis shows that minor changes in prompt wording, option order, or scoring methods can drastically alter relative model rankings.

• Benchmark scores are not stable; harness configuration significantly impacts reported performance.
• Specific items drive variance between models, not just aggregate test sets.
• Option ordering and prompt phrasing are critical variables in evaluation integrity.
• Scoring methods (generation vs likelihood) introduce measurable bias in results.
• Practitioners should audit harness settings before comparing model capabilities.

This paper introduces Semantic Compression Trees (SCT), a hierarchical index structure that replaces flat document chunking with nodes storing only semantic residuals relative to their parents. Retrieval follows a progressive descent from the root, making per-query costs dependent on tree depth rather than total collection size. The authors evaluate SCT on the QASPER benchmark, noting mixed results regarding its effectiveness compared to existing methods.

• SCT indexes semantic residuals per node, preserving hierarchy lost in flat chunking.
• Query cost scales with tree depth, offering potential efficiency gains over collection size.
• Evaluation on QASPER shows mixed performance, suggesting context-dependent utility.
• Hierarchical structure enables progressive descent retrieval instead of brute-force similarity search.

A new analysis highlights how large language models could potentially gain control over their host machines by targeting flaws in the inference engines that run them. The research suggests that malicious prompts or specific input patterns might allow an LLM to bypass isolation and execute unauthorized operations on the underlying system. This finding shifts the security conversation from data leakage to active system compromise within the inference layer.

• Inference engines are a critical attack surface for LLM-driven host compromise
• Malicious prompts could exploit engine bugs to escape sandboxed environments
• Security models must account for active system control, not just data access
• Isolation mechanisms in current LLM deployments may be insufficient