Sunday edition with a ~3-day window (arXiv papers submitted Nov 11–12). Highlights: models that learn when to reason (HDM cuts ~80% of tokens), Singularity-Bench for agents that build AI systems, two routes to self-improving agents (MAPE, ACE-RL), plus findings on synthetic data, streaming ASR and TTS voices for agents. No notable new model release today that was not already covered.
Large Language Models
1
The Value of Thinking: Learning When to Reason in Hybrid Decision Models
Proposes zero-or-T thinking: one model either answers from implicit reasoning in a single forward pass or runs full autoregressive reasoning costing T tokens. A Hybrid Decision Model uses a frozen backbone for both modes and a light self-routing block that estimates the value of thinking, invoking the slow mode only when that value exceeds a per-request price. CDPO trains the router from episode-level returns for agentic tasks. HDM-4B slightly beats the baseline on JevBench with much lower latency, and HDM-27B cuts token cost by about 80% on MiniWoB.
Why it mattersAdaptive reasoning usually just shortens chain-of-thought; this makes not generating at all a priced, first-class option. It links the emerging System One decision-model line to reasoning models, and CDPO extends the idea to agents without per-step labels.
Human or Synthetic? What a Fixed Budget Buys for Post-Training
Compares human and synthetic preference data for post-training (SFT and DPO) under dollar-matched budgets on helpfulness and summarization. Synthetic data is far more cost-efficient at low-to-moderate budgets, reaching high win rates with as little as $200–400. On helpfulness, human data closes the gap only at roughly 100x larger budgets; on summarization it never does within the tested range. Human sources also vary: one widely used source clearly underperforms another even after domain restriction and deduplication.
Why it mattersGives a cost-denominated answer to a question most teams settle by intuition. Small teams can start synthetic and buy human data only where it demonstrably pays, and should not treat human data as a uniform gold standard.
From Large Language Models to Efficient System One Models via Adaptive Computation
Studies how to turn LLMs into fast System One decision models, which swap the LM head for a decision head but still run every layer. Decision information concentrates in middle layers, and the last six layers can be removed without losing accuracy. Independently trained smaller models mostly imitate the full model, suggesting efficiency comes from sharing computation across depths. The method jointly trains the full model and prefix sub-models via self-distillation with a light gate at each exit, matching or beating per-depth models on seven decision benchmarks.
Why it mattersEarly exit on decision heads is a cheap path to latency savings without training a separate small model per budget. The layer analysis is useful for anyone distilling LLMs into classifiers, judges or routers.
An open-ended benchmark for autonomous AI development with six end-to-end tasks: pre-training, mid-training, post-training, inference, the agent harness and cluster orchestration. Agents build core systems from a minimal start and are compared with human-engineered baselines on identical hardware. Red-teaming across 9 of 10 OWASP Top 10 for Agentic Applications categories exposed reward-hacking holes in the Harbor framework, leading to the Breakwater extension that blocks all tested attacks. Across 11 agents (7 frontier models, 2 harnesses), Opus 5.5 with Claude Code matches or beats the human baseline on all six tasks (Singularity-Index 1.25); the others average 0.58.
Why it mattersIt measures the capability most relevant to recursive self-improvement end to end rather than through toy ML tasks. The reward-hacking findings matter as much as the leaderboard, since evaluating autonomous AI R&D needs tamper-resistant harnesses.
Discovering Effective Agent Memory Designs through Program Evolution
MAPE uses program evolution to design agent memory systems, starting from minimal seed programs with a standardized lifecycle interface and an iterative diagnose-and-propose loop. On LoCoMo and LongMemEval it gains up to 12 points over hand-designed baselines at about 35% of the serving cost, and transfers across backbones and benchmarks. Two recurring patterns emerge: keep raw conversations as retrievable evidence, and compose retrieved evidence by query intent; adding the composition module to five existing memory systems improved all of them. Optimizing cost shifts compute from LLM-based memory construction to selective query-time processing.
Why it mattersAgent memory has been hand-engineered with many competing designs; automated search beats them and yields transferable principles. The keep-raw-evidence, compose-at-query-time lesson is directly actionable.
Experience-Handoff Reinforcement Learning for Recursive Self-Improvement
Scalar-reward RL discards rich environment feedback such as error logs, while agent harnesses use it through playbooks that are never learned. The authors fix the Agentic Context Engineering (ACE) harness and train one model in three roles (action generator, insight reflector, playbook curator), scoring each playbook edit by its effect on the next attempt's reward (ACE-RL). The authors report the highest held-out reward versus PPO, GRPO, ERL and end-to-end PPO through the harness, staying ahead of GRPO up to 8 attempts despite training with 3. The abstract gives no specific figures.
Why it mattersIt bridges weight updates and in-context experience, two self-improvement routes usually treated separately. If it holds up, models should be trained to write better notes for their future selves, not just to act better.
Beyond Word Error Rate: A Failure Taxonomy of Streaming Dictation ASR
Argues that WER on recorded files misses what matters in live dictation. The authors streamed a test set to 11 ASR systems (106 English dictations in quiet, fan noise and whisper conditions, 11 other language sets and 400 code-switched recordings) and sorted errors into four levels. Zoom Scribe and GPT-Live-Transcribe tie on quiet English (7.4% vs 7.3% WER) but split on whispered speech (34.7% vs 8.8%). With only English selected, Gemini, Deepgram and self-hosted Whisper lose the non-English part in 86–99% of mixed recordings, and putting Whisper behind a Silero VAD raises whisper WER from 10.1% to 51.4%. Code, recordings and transcripts are released.
Why it mattersHeadline WER can hide whole failure classes, often in the pipeline (VAD, endpointing, language settings) rather than the model. A practical checklist for anyone deploying streaming ASR in voice agents.
Human-Likeness and Conversational Preference Diverge for Synthesized Agent Speech in Customer-Service Calls
Compares two listener judgments of synthesized agent speech: how human it sounds and whether listeners would want to talk to it. Paired-choice tests on 120 actor-improvised customer-service turns compared six commercial TTS systems (eight voices each) and pitted systems against the actors' recordings. Rankings diverged between the two judgments; listeners chose a system over an actor in 19% of human-likeness trials, 36% of preference trials and 49% for the most preferred system. Variation among a system's voices was about three-quarters of the variation between systems, and none of three automatic naturalness scores predicted preference.
Why it mattersVoice-agent teams often pick TTS by naturalness scores, which this shows can mislead about what callers prefer. Voice choice within a vendor matters nearly as much as vendor choice, and faster speech with fewer pauses is favored.
