LLMs

This research introduces Fractional Decay KV-Cache (FD-KVC), a memory management algorithm designed for transformer-based dialog systems. Unlike uniform caching strategies, FD-KVC uses a dual-channel scoring system that combines cumulative attention importance with recency-weighted relevance. This approach allows the model to retain historically significant tokens while rapidly adapting to shifting conversation topics through temporal decay and reinforcement-inspired updates.

• Dual-channel scoring balances long-term importance with short-term recency.
• Adapts inference memory automatically as dialog topics evolve.
• Uses temporal decay and reinforcement updates for dynamic relevance.
• Outperforms coarse eviction heuristics in maintaining context fidelity.

Amazon Bedrock now hosts SpaceXAI Grok 4.6, a model optimized for coding, agentic workflows, and knowledge tasks. The service introduces cross-region inference to automatically distribute requests for higher throughput and reduced costs. Two distinct profiles are available: a US Geo variant ensuring data residency within the United States, and a Global profile for broader geographic distribution.

• Grok 4.6 is now available on Bedrock for coding and agentic tasks
• Cross-region inference auto-routes requests to lower latency and cost
• US Geo profile (us.xai.grok-4.6) enforces US-only data residency
• Global profile (global.xai.grok-4.6) serves requests from any region
• No manual capacity management required across multiple AWS Regions

This research isolates self-preference bias in LLM judges by evaluating narrative constraint selections rather than full text responses, thereby removing stylistic confounds. The study reveals that providing self- or other-labels induces bidirectional bias in ten different models during blind evaluations. These findings suggest that current evaluation metrics may conflate genuine preference with model-specific signatures hidden in the data.

• Labeling outputs as 'self' or 'other' introduces bidirectional bias in LLM judges.
• Evaluating constraints rather than full text isolates bias from stylistic features.
• Ten LLMs showed distinct bias patterns under blind evaluation conditions.
• Existing self-preference metrics may be confounded by recoverable model signatures.
• Evaluation reliability requires decoupling preference from stylistic fingerprints.

Researchers propose a sequence labeling approach for detecting token-level hallucinations in LLMs, moving beyond independent single-signal scoring. The method fuses text statistics, NLI entailment, and language model surprisal into a 33-dimensional feature stream processed by a BiGRU. This architecture achieves an AUC of 0.840 on RAGTruth, significantly outperforming logistic regression baselines by leveraging temporal context.

• Hallucinations are treated as temporally extended spans rather than isolated token errors.
• A 33-dim feature stream combines text stats, NLI, and surprisal without model internal access.
• BiGRU processing yields an 11-point AUC gain over independent logistic regression baselines.
• Temporal order is the primary driver of performance improvement in the fusion model.
• Method avoids reliance on model internals, enabling broader applicability across architectures.

A new position paper analyzes 500 Hugging Face model cards, arguing that existing transparency artifacts are insufficient for governing open-weight foundation models. The authors identify a safety gap in current regulatory approaches and propose a multi-layered framework. This approach combines model cards with acceptable use policies and licenses to better inform downstream developers.

• Existing model cards alone do not address distinct safety challenges of open-weight models.
• Effective governance requires integrating model cards, AUPs, and licenses.
• Current frameworks fail to provide adequate transparency for downstream users.
• Policy makers should consider multi-layered oversight for open-weight models.

Instruction-tuned models often bypass safety refusals for low-resource African languages like Yoruba and Hausa, despite enforcing them in English. This approach introduces Latent Space Refusal Anchoring, a training-free technique that extracts the refusal direction from English prompts and clamps it onto the residual stream during inference. The method successfully recovers safety mechanisms across Llama and Mistral architectures without requiring target-language retraining or labeled data.

• Training-free safety recovery avoids costly retraining for low-resource languages lacking labeled data.
• Refusal directions extracted from English are clamped to the residual stream to force compliance.
• Effective across diverse architectures including Llama-3-8B, Llama-3.1-70B, and Mistral-7B-Instruct.
• Addresses the specific failure mode where models comply with harmful requests in African languages.
• Mechanistic intervention targets the residual stream rather than modifying model weights.