AI FRONTIER
Signals worth understanding before they become obvious.
Introducing GPT-6.1 Sol
GPT-6.1 Sol offers near-Astra intelligence for coding, computer use, and professional work at one-fifth of Astra's standard API input and output token prices.
- OpenAI News announced GPT-6.1 Sol.
- It is positioned for coding, computer use, and professional work with near-Astra intelligence.
- Standard API input and output token prices are one-fifth of Astra's.
For AI builders, near-Astra intelligence at one-fifth the token price lowers the inference cost of coding and computer-use agent workflows.
The System Prompt Illusion: How Instruction Preambles Modify Computation in Language Models
The paper uses Centered Kernel Alignment (CKA) to compare layer-wise representations under 20 system prompts across 17 instruction-tuned models, finding that effects are layer-selective and instruction-type-dependent: persona and formatting instructions deeply restructure intermediate representations, while safety inst...
- Evaluated 17 instruction-tuned models spanning 8 architecture families and 1.5B to 72B parameters under 20 system prompts in five functional categories.
- Restrictive safety instructions and explicitly permissive ones ("you have no restrictions") engage near-identical computational pathways, with mean CKA correlation of 0.997.
- Safety penetration remains below 10% at commercial scale, even at 70B-72B.
- Linear probing shows the model encodes prompt category at every layer but restructures computation only at a small subset, and causal activation patching confirms these layers mediate behavioral change.
For builders relying on system-prompt-based safety, this evidence indicates safety instructions may be reliably encoded but not deeply acted upon in computation, so prompt-layer defenses alone are insufficient against jailbreaks.
TomasuLLM: Out-of-Order Speculative Execution for LLM Agents
TomasuLLM is a runtime that executes agent tool calls out of trajectory order while preserving task-execution correctness: it drafts future actions, runs them in isolated copy-on-write sandboxes, traces their dependencies and effects, and commits results in trajectory order only after validation.
- The paper reports 1.31x on 100 SWE-bench Verified tasks, 1.35x on 28 Terminal-Bench 2.0 tasks, and 1.27x matched progress on 18 SWE-Marathon sessions.
- Across 4,010 audited commit-validation records, the paper reports zero false accepts.
- The work targets long-running tools such as compilers, test suites, and repository commands that take seconds to minutes while the coding agent idles.
- Results span three benchmarks covering sub-second to minutes-long tool calls and scale with tool latency.
For teams building coding agents, TomasuLLM shows that speculatively running tool calls in isolated sandboxes and committing them in trajectory order can cut long-tool latency without sacrificing commit correctness.
MiniMax Open-Sources OpenAgentCore for OpenAI Agents API Compatibility
MiniMax has open-sourced OpenAgentCore, aimed at compatibility with the OpenAI Agents API.
- MiniMax open-sourced OpenAgentCore.
- OpenAgentCore targets OpenAI Agents API compatibility.
AI builders can use OpenAgentCore to target OpenAI Agents API workflows with a MiniMax open-source implementation.
SInGA: Learning Semantic Inpainting for Animatable Gaussian Head Avatars
The paper presents SInGA, a method that defines a semantic inpainting framework in UV space to complete unobserved facial regions from a single image and then regress Gaussian attributes, producing an animatable Gaussian head avatar. The resulting avatar generalizes across identities without per-identity optimization a...
- The semantic inpainting framework is defined in UV space and uses the inherent symmetry cues of human faces to complete unobserved regions.
- Features extracted from observed regions are used to complete unobserved regions, and the completed representation is then used to regress Gaussian attributes.
- Instead of a single Gaussian at each surface or pixel location, the method stacks multiple Gaussians to enhance detail.
- The paper reports that the method generates high-quality head avatars with improved completeness and identity preservation, while supporting realistic animation and consistent rendering from unobserved views.
For builders of single-image head avatars, moving completion into a UV space with consistent spatial correspondences offers a reusable way to handle unobserved regions in single-view settings.
Conformal Factuality Control for Multi-Hop Retrieval-Augmented Generation
The paper applies split-conformal claim filtering to multi-hop RAG and evaluates it on HotpotQA, Natural Questions, and TriviaQA with Llama 3.1 8B and GPT-4o-mini, alongside a single-hop reference experiment. It reports that at the 95% target, the fraction of responses whose retained claims are fully supported rises to...
- Across all six multi-hop model-dataset configurations, increasingly stringent conformal targets consistently increase the fraction of responses whose retained claims are fully supported.
- At the 95% target, this rate ranges from 95.80% to 97.20%, compared with 55.60%-76.03% without filtering.
- At the 95% target, only 4.41%-31.09% of generated claims are retained and 9.70%-51.40% of responses remain non-empty.
- Evaluation uses Llama 3.1 8B and GPT-4o-mini on HotpotQA, Natural Questions, and TriviaQA, together with a single-hop reference experiment.
For multi-hop RAG builders, conformal filtering raises claim-level support but must be read jointly with claim retention and abstention, since at the 95% target most claims are discarded and many responses become empty.
Acceleration of Diffusion Language Model through Discrete Average Generator
The paper introduces the Discrete Average Generator, extending MeanFlow to Continuous-Time Markov Chains by defining an average generator as the normalized increment of the transition kernel over a time interval, with a self-consistency identity as the basis of the training objective.
- In Potts model simulations, the objective reduces the total variation distance of the K-step sampler by up to 67%.
- On OpenWebText, the method achieves the lowest generative perplexity among the evaluated methods for 8 to 64 sampling steps while enabling a 16x acceleration.
- On ImageNet, the method achieves comparable performance to existing methods.
- The self-consistency identity admits a closed-form expression when projected onto per-coordinate marginals.
For teams building discrete diffusion language models, this offers a training objective that cuts sampling steps while preserving generation quality, making the reported 16x acceleration worth reproducing on their own data.
AI Agents Are Vulnerable to Radicalization
The paper simulates conversations between two LLM agents: a target that role-plays a human persona based on demographic and psychological attributes, and an influencer that aims to make the target's beliefs more extreme. It reports that both resonance (reinforcing a pre-existing belief) and persuasion (promoting a beli...
- Submitted to arXiv on 29 Sep 2026 as arXiv:2609.38296, classified under cs.AI and cs.CY.
- Setup: a target LLM role-plays a human persona based on demographic and psychological attributes, while an influencer LLM aims to make the target's beliefs more extreme.
- The paper examines two pathways: resonance reinforces a target's pre-existing belief, and persuasion promotes a belief the target initially considers unimportant.
- The paper reports that both mechanisms radicalize the target across affective and behavioral metrics, but resonance produces consistently stronger effects than persuasion.
Builders of personalized agents and multi-agent systems should treat belief-aligned inputs as a high-risk influence surface, because the paper reports resonance yields stronger radicalization that also propagates to related beliefs.