Artificial Intelligence

The AI landscape in early 2026 is defined by a shift from building larger models to making AI systems smarter, more autonomous, and more interpretable. Three major consolidation papers published in Q1 2026 establish shared frameworks for understanding LLM agents: a three-layer agentic reasoning model (foundational → self-evolving → multi-agent), a three-paradigm tool-use framework (prompting → supervised → RL), and a 60-benchmark taxonomy for evaluation. Meanwhile, Google DeepMind's AlphaEvolve demonstrates that LLM-driven evolutionary search can beat 57-year-old algorithms, and Anthropic's Transformer Circuits Thread culminates in circuit tracing with attribution graphs that reveal end-to-end computational paths in production models.

April 2026 papers sharpen two critical questions: what actually triggers exploitation in deployed agents (goal reframing is the sole reliable trigger across 10,000 trials; 9 of 12 hypothesized attack dimensions fail), and where the real capability bottleneck lies in tool-using agents (navigation, not tool execution — best agents achieve only 37.2% on compositional multi-hop tasks despite near-perfect tool-use reliability). A third paper updates mechanistic interpretability with a comprehensive alignment integration survey. A fourth demonstrates that deployed personal agents with persistent state face architectural vulnerabilities (CIK taxonomy) that no model substitution can fix.

A parallel and architecturally incompatible story is playing out in world models. Two camps both claim the "world model" label: JEPA (Meta FAIR, LeCun) predicts abstract representations rather than pixels — V-JEPA 2 pre-trains on 1M+ hours of video and enables zero-shot Franka manipulation after <62h of robot data. Generative world models (DeepMind Genie 3, Sora-family, Wayve GAIA) predict pixels directly — Genie 3 is an 11B-parameter autoregressive transformer producing real-time 720p interactive worlds at 24fps with ~1 minute consistency. Two comprehensive surveys organize the schism rather than resolve it.

By mid-2026 the picture sharpens on three further fronts. The frontier compressed into a tight, cost-disrupted cluster: Claude Opus 4.7 (Apr 16), GPT-5.5 (Apr 23), and the open-weight DeepSeek V4 (Apr 24) shipped inside an ~8-day window, with V4 reaching near-frontier capability at roughly one-sixth the cost. New reasoning architectures propose adding capability without scaling parameters: looped/latent-reasoning models (LoopFormer) move budget-conditioned deliberation into weight-space with no token trace, and conditional memory (Engram) adds a static-lookup sparsity axis orthogonal to MoE. And a government safety body now publishes longitudinal capability trends: UK AISI's first Frontier AI Trends Report (Dec 2025) measures cyber-task time horizons doubling every ~4.7 months, self-replication capability rising from <5% to >60%, and universal jailbreaks present in every tested system.

July 2026 opens a third axis that this KB had not previously covered: the economics and engineering of inference itself — and it lands directly on the house thesis that the edge sits in the intelligence and harness layer rather than in serving or silicon. The first systematic study of LLM token pricing (3,237 models, 2020–2026) documents a ~600× price decline and decomposes it: ~103.7% attributable to software and architectural innovation, ~−0.9% to GPU hardware. The mechanisms behind that number are now visible directly — speculative decoding (DSpark: 60–85% faster per-user generation in a live production fleet, from a drafting architecture and a verification scheduler) and low-bit quantization (OSAQ: >40% lower 2-bit perplexity from a closed-form Hessian transform with zero inference overhead), which compound rather than substitute (6.978× when combined). All of it runs on unchanged weights and unchanged silicon.

The same month delivers the strongest counter-pressure on the thesis, and it deserves equal billing. Wang et al. audit automatic harness evolution and find it does not consistently beat simple test-time scaling under matched feedback and inference budget, and generalizes poorly to held-out tasks. That does not show harnesses don't matter — no arm tested harness-vs-no-harness, and test-time scaling is itself harness-layer work — but it does demolish the specific claim that scaffold advantage can be discovered automatically and transfers. Meanwhile Anthropic is shipping the harness as versioned, billable API surface (agent-memory header migration, per-session config override, Dreams) faster than the evidence supports. The harness layer is where cost and reliability are determined; the claim that it is a source of automatic, transferable capability gains just failed its first serious audit.

The 2026-07-23 compile then supplies the missing half of that story — the harness thesis measured, plus a new KV-cache lever and an evaluation correction. Three papers pit the harness layer against the model layer directly: Stop Comparing LLM Agents Without Disclosing the Harness finds harness-induced variance running 7.80× model-induced variance (18.48 vs 2.37 pp², 6/9 ranking reversals) in a controlled 3×3 SWE-bench factorial — the "edge is the harness, not the model" thesis measured in a controlled experiment, graded moderate (small n, self-measured, position paper). Harness-Bench (5,194 trajectories) and belief-divergence (harness alters an agent's beliefs even at matched terminal success) corroborate it on outcomes and internal state. So the thesis splits cleanly: "the harness beats the model" is now measured and standing; "the harness edge is auto-discoverable and transferable" is the part that failed. On the efficiency side, VeriCache opens the previously-empty KV-cache lane by making lossy cache compression lossless — using the compressed KV as a speculative drafter verified against the full KV (~4.3× throughput, output identical), the strong form of "levers compound" in which compression is speculation. But that pattern is now explicitly bounded: a flagged (not-yet-ingested) line, arXiv:2505.22179, reports speculative decoding and low-bit quantization can conflict — tree drafts on a 4-bit target can negate the quant gain unless made sequential — so "levers always compound" is recorded as too strong. Finally, How Inference Compute Shapes Frontier LLM Evaluation shows fixed single-budget scoring increasingly understates newer models, making inference budget a first-class evaluation axis and collapsing "how good is model X" and "how much does X cost" into one curve — the theoretical backbone of per-token → per-task repricing.

Frontier — What's Moving Now

• Inference efficiency is a software discipline, and now a quantified one — ~103.7% of the token-price decline attributed to software/architecture vs. ~−0.9% to hardware, with the mechanisms directly observable: verification scheduling, loss-support truncation, closed-form quantization. All on unchanged silicon.
• Speculative decoding × quantization compound — but not universally — 6.978× combined on a single A10G; VeriCache makes cache compression be speculation (~4.3×, composes with EAGLE to 4.35×). Bounded by a flagged conflict (arXiv:2505.22179): tree drafts on a 4-bit target can negate the quant gain unless made sequential. "Levers always compound" is now recorded as too strong.
• KV-cache lane opened — VeriCache turns lossy KV compression lossless by drafting from compressed KV and verifying against full KV in parallel: output identical (KL<0.01 nats), ~25–40 accepted tokens/round, up to ~4.3× throughput. First KV-cache source in the KB.
• The harness beats the model — now measured — Stop-Comparing puts harness-induced variance at 7.80× model-induced variance (controlled 3×3 SWE-bench factorial, 6/9 ranking reversals); Harness-Bench (5,194 trajectories) and belief-divergence corroborate. The direct evidentiary support for "edge is the harness." Separately, harness evolution (automating the design) still fails its matched-budget test — the two are consistent.
• Inference budget is an evaluation axis — fixed single-budget scoring understates newer models (12 models × 7 benchmarks); capability should be reported as a function of inference-time compute. Collapses "how good" and "how much" into one curve — backbone of per-task pricing. Pairs with the harness-evolution matched-budget floor.
• Token pricing is tiered, not monolithic — economy half-life 1.10yr, mid 1.55yr, flagship resists; ~31.5× reasoning premium; a May-2024 structural break from technology-driven to competition-driven pricing. July 2026 showed three labs moving in three different directions in one month.
• Open-weight strategy bifurcated — DeepSeek competes on cost (~1/6th), Moonshot's Kimi K3 competes on capability while raising price ($0.95/$4 → $3/$15). "Open weights" no longer predicts "cheap."
• Tokenizers are an unpoliced pricing lever — Anthropic's ~+30% tokenizer expansion is an effective price rise invisible on a price sheet, and it silently invalidates cross-vendor "token efficiency" claims.
• The buyer is losing control of deliberation spend — Anthropic removed manual thinking budgets (400 on attempt); Moonshot shipped "max"-only reasoning. This compounds the latent-reasoning threat to CoT auditing with an economic one.
• World models split into JEPA vs. generative camps — V-JEPA 2 and Genie 3 occupy incompatible architectural bets under the same label. The schism is the frontier.
• Agent safety goes empirical — Goal reframing confirmed as sole exploitation trigger; CIK taxonomy quantifies persistent-state vulnerabilities (64–74% ASR from single-dimension poisoning).
• Navigation is the real agent capability gap — Tool execution is near-perfect; compositional multi-hop navigation drives 27–52% of agent failures.
• Mechanistic interpretability breakthrough — Circuit tracing with attribution graphs reveals computational paths in production models. Named a 2026 breakthrough technology.

Concept Map

Concepts

Entities

Timeline

See timeline.md for chronological developments (1969 through July 2026).

Research Frontier

See frontier.md for active research directions, breakthroughs, and knowledge gaps.

Sources

Note: the Sources table above lists key/representative sources; the authoritative full registry (55 entries) is ../raw/_sources.json.

Evidence-handling note for this KB: sources 38–41 were full-text-ingested via arXiv HTML5 on 2026-07-23 (local raw/*.md files now exist; the earlier abstract-only compile was directionally confirmed, no claim reversed). Sources 42–46 (the 2026-07-23 discovery pass) were read at abstract + partial full-text depth — Stop-Comparing and VeriCache at full HTML5, the rest abstract-grounded (belief-divergence magnitudes, Harness-Bench per-backend grid, and the inference-compute per-benchmark curves are owed). Sources 33–37 are marked summary-derived in the registry except 35 (Anthropic release notes), a complete first-party fetch. Flagged-but-unread: arXiv:2505.22179 (spec-decoding × quantization conflict) is cited on the Speculative Decoding / Low-Bit Quantization / KV-Cache pages as a tension, not yet ingested — evidence pending.