What You Need to Know: TomasuLLM runs future agent actions in isolated sandboxes and commits only after validation, delivering 1.31x gains on SWE-bench Verified. Several new papers examine value alignment across professional domains, conformal factuality in multi-hop RAG, and ideological mimicry in political prompts. Builders should watch the agent runtime work and the alignment evaluation frameworks for immediate testing.

Top Story

TomasuLLM introduces a runtime that executes agent tool calls out of trajectory order while preserving task-execution correctness. It drafts future actions, runs them in copy-on-write sandboxes, traces dependencies, and commits results only after validation. The system improves benchmark means across three suites that include sub-second to minutes-long tool calls. On 100 SWE-bench Verified tasks it reaches 1.31x, on 28 Terminal-Bench 2.0 tasks it reaches 1.35x, and on 18 SWE-Marathon sessions it reaches 1.27x matched progress. Across 4,010 audited commit-validation records it records zero false accepts. The approach directly targets the idle time created by compilers, test suites, and repository commands. Source: arxiv.org

Model Updates

Large Language Models are Approximate Survival Estimators: arXiv NLP Frontier LLMs achieve cMAE within 10 percent of state-of-the-art random survival forests on several cancer types in the MSK-CHORD cohort and lower cMAE than RSF for prostate cancer. GPT-5.6-Sol is the example model cited. Feature ablations show LLMs prioritize clinical variables similarly to specialized survival models. LLMs still show inconsistent accuracy across cancer types and institutions and post lower c-index scores. Source: arxiv.org

The System Prompt Illusion: How Instruction Preambles Modify Computation in Language Models: arXiv NLP Across 17 instruction-tuned models from 1.5B to 72B parameters, persona and formatting instructions restructure intermediate representations while safety instructions produce changes statistically indistinguishable from a minimal baseline. Restrictive and explicitly permissive safety instructions engage near-identical pathways with mean CKA correlation 0.997. Safety penetration remains below 10 percent even at 70B-72B scale. Representational depth predicts behavioral effect size with Spearman rho of 0.761. Source: arxiv.org

Conformal Factuality Control for Multi-Hop Retrieval-Augmented Generation: arXiv NLP Split-conformal claim filtering applied to multi-hop RAG on HotpotQA, Natural Questions, and TriviaQA raises the fraction of fully supported retained claims to 95.80-97.20 percent at the 95 percent target. Only 4.41-31.09 percent of generated claims are retained and 9.70-51.40 percent of responses remain non-empty at that target. The evaluation covers Llama 3.1 8B and GPT-4o-mini. Source: arxiv.org

Framing the Narrative: Ideological Mimicry in Large Language Models: arXiv NLP Changing terminology alone reverses which side of an issue a model supports in 16.9 percent of matched comparisons across seven open-weight LLMs and ten contentious topics. Stated political ideology systematically shifts responses toward the user's position. The Poli-SHIFT dataset and evaluation framework support the measurements. Source: arxiv.org

ContextAdapt: Evaluating Contextual Adaptation and Value Alignment in LLMs: arXiv NLP Twelve LLMs achieve 95.6 percent mean appropriateness on honesty, autonomy, and confidentiality scenarios across medicine, law, finance, and national security. Correct domain-specific justification use varies from 25.6 percent to 76.9 percent across models. Varying stakes reveals severe but localised failures where models alter responses even though the underlying professional obligation remains unchanged. Source: arxiv.org

Agent & Tool Developments

EVOKE: Eliciting World Knowledge in Agents for Transferable Decision-Making: arXiv NLP EVOKE supplies goal diversity at fixed states during post-training to elicit pretrained world knowledge without additional world-model training. It evaluates across diverse tasks in three backbones and reports improved task performance, unseen environment generalization, and data efficiency. The method forces action preferences to change under alternative goals so that policies relying on contextual habits cannot order them correctly. Source: arxiv.org

Which Models Work Well Together? Measuring Heterogeneity for LLM Team Selection: arXiv NLP The heterogeneity-driven framework profiles individual capability plus decorrelation in error patterns and divergence in predictive behavior. It applies an efficient greedy search to select a small team from a candidate pool. Experiments across multiple benchmarks show consistent outperformance of quality-only baselines under controlled candidate pools and team sizes. Source: arxiv.org

DEdit: Iterative Draft Editing for Speculative Decoding: arXiv NLP DEdit adds iterative token-to-token editing to a diffusion-based drafter so later predictions repair earlier errors using bidirectional context. ProposalMix mixes draft predictions with ground-truth tokens based on first-pass confidence during training. On Qwen3-4B and Qwen3-8B it reaches macro-average speedups of 5.72x and 5.97x over autoregressive generation under greedy decoding. Source: arxiv.org

Recovering Off-Policy Supervision for Speculative Decoding: arXiv NLP Anchor-Label Relabelling replaces corpus labels with distributions from greedy target rollouts. In-Rollout Anchors places draft blocks inside target rollouts to expose the drafter to target-generated context at no extra target cost. The framework increases greedy accepted length by up to 36.5 percent over DFlash on fixed vision-language and text corpora. Source: arxiv.org

Practical & Community

NinaXander: Feasibility and Limits of Composing Frozen Language Models Across Architecture Families via a Shared Latent Space: arXiv NLP NinaXander connects layers of frozen models from different families with a single trained shared-latent adapter. The configuration combining the first five layers of Pythia with the remaining 27 layers of RWKV reduces Transformer KV cache by 84.4 percent with accuracy not significantly different from RWKV alone. No composed model matched parent model Pythia on multiple-choice accuracy. Source: arxiv.org

Doc2LoRA Provides Decodable Representations of Scientific Ideas: arXiv NLP Doc2LoRA generates a LoRA adapter for each paper so every point in the embedding space corresponds to an LLM that can be queried in natural language. On American Physical Society papers the average-of-subfield instruction produces labels closer to official names than five baselines. A small invertible transform makes the embeddings competitive for search while remaining mappable back to an LLM. Source: arxiv.org

Evaluating Whether LLMs Can Reliably Connect the DOTs?: arXiv NLP A multi-domain benchmark of approximately 9.2K instances tests narrative infilling across encyclopedic text, commonsense stories, news articles, and visual narratives. Gemma-2-2B achieves the highest qualitative score of 4.02 out of 5, outperforming models over ten times larger. Explicit chain-of-thought reasoning yields only a 0.6 percent marginal improvement. Source: arxiv.org

KlinikeBench: Evaluating Language Models Beyond Diagnostic Accuracy: arXiv NLP KlinikeBench contains 333 clinician-authored tasks each with a virtual patient, clinical tools, and task-specific success criteria. More than 35 clinicians contributed to case authoring and evaluation. Best-performing models succeed on less than 30 percent of tasks even though diagnosis accuracy reaches 90.7 percent. Source: arxiv.org

Under the Hood: Speculative Decoding Tradeoffs

Speculative decoding lets a small drafter propose tokens that a larger target verifies in parallel, yet the quality gain disappears above roughly seventy billion parameters because the drafter's error rate begins to offset the parallel verification benefit. The core mechanism works by accepting a prefix only when every drafted token matches the target's output distribution, so a single early mismatch forces the entire block to be discarded. Adding iterative editing passes, as in DEdit, lets later predictions supply bidirectional context that repairs earlier errors without restarting the draft. This raises accepted length but adds a second forward pass per editing round, trading roughly two hundred milliseconds of latency for the observed five point seven times speedup on Qwen3-4B. The practical decision rule is to use plain speculative decoding when the drafter matches the target family and context length stays under eight thousand tokens; switch to editing or anchor-relabelling once the drafter is drawn from a different family or the corpus contains many long tool calls.

Things to Try This Week

• Try TomasuLLM on SWE-bench Verified tasks to measure the 1.31x latency reduction on long-running tool calls.
• Run ContextAdapt scenarios on any of the twelve evaluated models to check domain-specific justification rates.
• Test EVOKE post-training on a held-out environment to observe gains in unseen generalization without extra world-model training.
• Apply Doc2LoRA embeddings to a set of APS papers and query the resulting LoRA adapters for subfield descriptions.

On the Horizon

• Additional results from the HalluScoring 2026 shared task on Arabic hallucination detection are expected in the coming weeks.
• Further scaling experiments on NAMOH sparse attention are anticipated from the same research group.
• Expanded KlinikeBench tasks covering additional clinical specialties are under development by the authoring clinicians.

💬 Reply to this email — Patrick reads every one.

Share: X · LinkedIn · WhatsApp

Forwarded this email? Subscribe here — it's free.