OpenAI
2602
80
TRIAGE: Reliability- and Load-Conditioned Escalation for Two-Tier LLM Serving with a Semantic Evidence Judge
Two-tier LLM serving escalates selected requests from a small model to a large one, but prompt-based routers cannot see whether retrieved evidence supports an answer, and they treat the GPU cost of an escalation as constant. We study TRIAGE, which conditions escalation on a post-generation reliability state and on a load-dependent GPU price, and we report both what held and what did not on one A40 serving Qwen2.5-7B and 32B (AWQ) under VLLM. Escalation cost rose 21.9x from 1 to 10 requests/s but showed no monotone dependence on KV-cache headroom, and a pre-registered claim that the reliability state detects corrupted evidence (AUROC>=0.70) was falsified at 0.571, although any prompt-only router scores exactly 0.500 by construction. Post hoc, a semantic evidence judge raised detection to 0.846 (0.926 on never-served items) and predicted wrong small-model answers at AUROC 0.90-0.96 on RGB and three RouterBench model pairs; the resulting cascade stayed within about 3 pp of GPT-4 at 79.5% lower API cost for GPT-3.5->GPT-4, yet fell below random routing for open-model pairs with 2-3x price ratios. End to end, the 7B->32B pair showed only a +2.88 pp held-out gap, so the allocator rightly declined to escalate and the headline is a null that a small pre-launch gate failed to expose.
Discussing politics with a chatbot: How political beliefs shape prompt formulation for generative AI
By generating comprehensive and highly personalized answers to user prompts, AI chatbots such as ChatGPT and Gemini can serve as alternatives to search engines and substantially transform online information seeking. However, little research has examined how users interact with these tools to find information, particularly about politics. In this experiment, we examine how Swiss citizens (N = 563) formulate prompts to search for migration-related issues using AI chatbots and whether this process is affected by selective exposure. We find evidence of attitude-consistent prompting, whereby participants who were more critical of migration were more likely to formulate prompts with an anti-migration stance. This effect was stronger for a broad topic (migration in general) than for a specific one (anti-migration popular initiative put to a vote in Switzerland). However, exposure to an attitude-consistent AI response did not further amplify selectivity in prompting. We also did not find evidence that either AI attitudes or thinking styles moderated selectivity in prompting. These findings contribute to the understanding of selective exposure via AI chatbots and the use of generative AI for political information-seeking.
Compile as Hypothesis: A Headroom Typology for Prompt Optimization on Real Reasoning Models
Automated prompt compilers search for or rewrite prompts under the implicit assumption that optimization consistently improves downstream task performance. We evaluate this assumption on frontier reasoning models through rigorous paired, same-day measurements. Across four optimizer families (genome search, profile-guided mutation, GEPA, and ESPO) and eight tasks, structured search yields no statistically significant accuracy gains on strong reasoners. However, a discrete structural genome provides essential architectural stability, preventing catastrophic bimodal seed-collapse (where unconstrained search collapses toward the judge floor in 3/5 runs). Through 190-problem evaluations on knowledge-intensive benchmarks, we establish a three-dimensional headroom typology: prompt optimization cannot compensate for intrinsic knowledge deficits or physical token truncation, but successfully resolves protocol and formatting errors. On gpt-6, while free-text reflective search (GEPA) yields a +.14 gain on a task where structured search remains flat, qualitative ablation reveals that ~.035 stems from judge-rubric reconstruction while ~.105 reflects genuine instruction refinement. Our pre-registered AutoAPC-Select deployment gate matches the oracle policy in 5/7 groups and rescues seed-collapse arms with zero regret. Finally, we expose silent backend-mixing in commercial APIs during outage windows and provide automated per-call audit verification tools to ensure empirical evaluation fidelity.
CompSkill: Compiling Execution Experience for Skill Retrieval
LLM agents increasingly rely on skill documents, yet most retrievers rank them by description alone and ignore what earlier executions revealed, even though a relevant skill can hurt execution. We present CompSkill, which indexes skills with learned 512-d keys and compiles matched same-task comparisons into skill offsets and capped key residuals (KeyOpt), so experience changes which document the executor receives, never its text. Under a logistic outcome model with unknown task difficulty, matched comparisons cancel that difficulty and only discordant ones carry information; on support executions, predictions improve with each discordant task. An experience hypergraph keeps each comparison with its executor and document versions, recalls missed skills and grounds a cited review. Compact keys raise SkillRet NDCG@10 from 64.0 to 70.9 at one eighth of the dimension. With one document, CompSkill completes 256 and 260 of 332 ToolQA tasks with GPT-5.6-luna and DeepSeek-V4.1-Flash, respectively, more than every baseline, and stays on par with the strongest baselines on AppWorld and ALFWorld. On 332 newly generated ToolQA tasks, it completes 256 with GPT-5.6-luna, significantly more than BM25, SkillRet-8B, Tool2Vec and RankGPT; over all 844 ToolQA tasks, it completes 24 more than the strongest baseline. Selection matters beyond relevance: pairing the right guide with a same-database sibling instead of another database's guide costs 8 of 96 tasks. Code, models and support data: https://github.com/satoshinji2992/CompSkill
Auriga Technical Report: A Task-Agnostic Agent Harness for Zero-Shot Embodied Manipulation
Vision-language models can interpret instructions and plan manipulation tasks, yet struggle when acting directly as robot policies. On RoboDojo, a large planning-execution gap separates high-level task decisions from physical execution: removing arm execution raises GPT-6 Astra's success rate from 22.5% under the reference control interface to 76.2%. We present Auriga, a task-agnostic agent harness that closes 68.3% of this gap in zero-shot manipulation without target-task trials or resets. In Auriga, the model makes all task decisions over an explicit task belief and acts through either direct tool calls or synthesized multi-step Python programs, while the harness supplies calibrated 3D measurements, finger-referenced dual-arm control, and a single admission gateway to the robot. Before motion, the model inspects a visual trajectory preview while simulation is paused; during execution, the harness checks declared stage tolerances and grasp bounds, stopping the robot on divergence and returning the measured discrepancy for in-place recovery. Developed across 13 cross-task capability iterations and frozen for evaluation, Auriga ranks first on the RoboDojo leaderboard with a 57.7% success rate (64.0 Score), more than doubling the reference harness on the same model, and remains first under scene randomization while learned policies drop by up to 33.9%. On RoboCasa365, the same frozen harness reaches a 73.2% success rate, outperforming VLA-assisted agents by up to 23.1% on composite tasks. Project page: https://auriga.stelledge.com/ Code: https://github.com/lose4578/Auriga
OpenAI reportedly lowers funding target for new round
OpenAI’s latest funding round doesn't seem to be living up to expectations.
Multi-Estimate Elicitation Does Not Uniformly Debias Language Model Judgments
Does eliciting five weighted estimates within a single response change how a language model responds to information? Across anchoring, hindsight, base-rate, conjunction, and status-quo/default tasks, we compare direct numerical answers with a multi-estimate response format, contrasting source framings, explicitly random cues, and probability-consistency constraints. We evaluate three completed configurations--gpt-6.1-sol, claude-sonnet-5.5, and Qwen3-8B (reasoning disabled, analyzed descriptively). In the first two configurations, direct elicitation yielded near-zero irrelevant-anchor indices and status-quo effects, while remaining responsive to informative sources and informed advice. Of 23 preregistered confirmatory tests, six rejected after Holm adjustment; however, all six absolute point estimates fell below their prespecified smallest effect sizes of interest without establishing equivalence. Moreover, while separate versus joint elicitation shifted conjunction judgments, interpretation is constrained by between-call noise. Ultimately, multi-estimate elicitation failed to produce uniform debiasing or response attenuation across tasks and models, though post hoc correlations between candidate spread and anchoring error motivate independent predictive validation. Materials: https://github.com/asj252/llm-judgment-studies
Conic Certification of Transformer Attention
Attention is the computational core of large language models, yet the verifiers that bound how far an attention head's output can move under input perturbations give no guarantee on how loose their bounds are. A weak certificate therefore cannot be told apart from a fragile model. We develop a conic programming theory of attention certification that supplies this guarantee. We prove that the difficulty lies not in the softmax but in the bilinear interaction of a varying query with the keys. This interaction makes certification NP-hard over boxes and inapproximable once the attention logits can move by a constant amount. A fixed query or listed token substitutions, by contrast, make the exact worst case efficiently computable. An exponential cone moment relaxation then certifies a head with an error bounded quadratically in the perturbation radius by computable quantities, an order that cannot be improved. We also show that relaxations that treat a network one operation at a time, linear bound propagation among them, pay on second-order terms a sum of per-operation maxima instead of one coupled maximum. Computational results on GPT-2 indicate that the theory carries over to a trained model. Against exact worst cases with the query fixed, the certificate attains the predicted second-order accuracy. It is tighter than linear bound propagation on every head, and the looseness of bound propagation tracks the predicted per-operation excess.
Distinguishing among reasoning levels of GPT 5.5 Pro via a grammar lens benchmark
Does the level of test-time processing effort affect a language model's computational capability? A formal grammar-lens benchmark method, derived from the theory of computation, generates novel data at specified levels, each successively more computationally powerful than the levels below. Using this data, we tested GPT-5.5, and GPT-5.5-Pro-Medium, -High, and -XHigh, showing incrementally increased acceptance of test data: 16, 40, 43, and 52\%, respectively, illustrating the benchmark's ability to discern fine-grained differences among models' added test-time reasoning levels. Additionally, as test-time compute effort increases, model responses diverge from human responses to the same data. This increasing disagreement may turn out to be broadly characteristic of more capable language models, introducing the potential of a general capability-versus-alignment conflict.
Agentic Context Management for Multi-Turn Human-AI Conversations
LLMs degrade in multi-turn conversation partly because of context pollution: the model's own earlier outputs stay in context and anchor later turns, so an early misreading of the user's intent can persist after the user corrects it. Assistant omission (AO), dropping every assistant message, removes this pollution but fails when interaction is referential: later turns refer back to drafts, decisions, and tool-call results that exist only in assistant turns. We propose Agentic Conversation Context Curation (AC3), a training-free method in which a separate analyzer (1) consolidates the task specification from the user's messages alone, (2) audits the assistant's prior outputs against it, and (3) edits the context, either appending the analysis or discarding the assistant side and rebuilding the context from the specification. A second finding: pollution is contagious. An analyzer that sees the assistant's reasoning while extracting the specification can do worse than no intervention, and instructing it to ignore them did not close the gap, so we exclude assistant turns structurally. On Lost in Conversation (replaying difficulty-selected instances), where AO is an oracle by construction, AC3 closes 55-84% of the gap between full context and AO on three of four tasks and exceeds AO on the fourth, and an LLM that faithfully summarizes the history does not reproduce the gain even with more compute. On real WildChat conversations, a GPT-5-mini judge prefers AC3's next turn over both full context and AO in every configuration of our main table, though this preference depends on the judge; on CollabLLM results are mixed. In stateful tool use (tau2-bench), AO collapses to 0% on every model; AC3 does not collapse, though it falls slightly behind full context. Detector studies show that Reset removes injected pollution but is not span-selective: the gain comes from rebuilding the specification.
To Jev or Not to Jev: When Can System-1 Decision Models Replace LLM Judges?
Decision models such as TypeSafe's Jev answer typed questions about a text (yes/no, multiple choice, or an ordinal score) with calibrated probabilities in a single forward pass and without generating text, which makes them an inexpensive candidate for LLM-as-a-judge workloads. We evaluate 14 System-1 decision models, selected with a pre-stated protocol, against four generative judges (Claude Haiku 5.5, GPT-5.6 with and without reasoning, and Claude Opus 5.5 with thinking) on eleven established judge benchmarks. Five of the 14 are statistically indistinguishable from GPT-5.6 without reasoning on all five pairwise and best-of-4 benchmarks, and the two strongest, DReX-Max and d1-ultra, are significantly more accurate than Haiku on JudgeBench, LLMBar and RewardBench 2. Conditioned on a reference answer generated once per prompt, eleven models reach 90.3-93.4% on JudgeBench, statistically indistinguishable from Opus 5.5 and GPT-5.6 with reasoning, and on PPE best-of-16 every reference-guided System-1 judge matches GPT-5.6 with reasoning at roughly one-sixteenth of its cost per prompt. Several System-1 judges are better calibrated than every generative judge and correlate more strongly with human helpfulness ratings. Robustness to instructions injected into the judged response is a property of the individual model rather than of the System-1 design: an appended note to the evaluator reverses 0-5% of the correct verdicts of five models but 17-49% for the other nine. Two general System-1 leaderboards that contain no judging tasks predict judging accuracy (Pearson r = 0.95-0.96), a well-calibrated System-1 first stage matches Opus on LLMBar while escalating 17% of items, and same-size controls show that the correctness deficit of a 4B decision model is a reasoning gap rather than a size gap.
Managing Amazon Bedrock costs end to end: A playbook for FinOps and platform teams
This post presents a five-stage operating model for managing Amazon Bedrock cost: attribute every dollar to an owner, make every request observable, enforce budgets, optimize platform defaults, and operate a review cadence. Written for FinOps and platform engineering teams.