Research Frontier: Artificial Intelligence

What's genuinely new and where the field is heading.

Active Frontiers

1. The Compressed Frontier Wave — Multimodal, Agentic, and Now Open-Weight (Apr 2026)

Status: Breakthrough | Key sources: Gemini 3, Muse Spark, DeepSeek V4, Latent reasoning, AI Landscape 2026 Key players: Google DeepMind (Gemini 3), Meta (Muse Spark), OpenAI (GPT-5.5), Anthropic (Claude Opus 4.7), DeepSeek (V4)

A coherent — and cost-disruptive — frontier-model release wave in April 2026, compressed into an ~8-day core window (Opus 4.7 Apr 16 → GPT-5.5 Apr 23 → DeepSeek V4 Apr 24):

• Google Gemini 3 (Apr 15): claimed multimodal leadership, improved agentic tool-use, native rich output (interactive visualizations).
• Meta Muse Spark (Apr 8): native multimodality (text/image/voice in a single transformer backbone), Contemplating mode for parallel sub-agent reasoning without sequential latency hit.
• Anthropic Claude Opus 4.7 (Apr 16): coding-precision + safety leadership ("Project Glasswing"); leads the hardest agentic-coding splits (SWE-bench Pro ~64%).
• OpenAI GPT-5.5 "Spud" (Apr 23): agentic versatility / knowledge work.
• DeepSeek V4 (Apr 24, open-weight MIT): 1.6T/49B-active MoE, 1M context, near-frontier capability (SWE-bench Verified ~80.6%, GPQA ~90–92%) at ~1/6th the cost of the closed peers.
• Architectural standard set: native multimodality is baseline; the differentiator shifted to cost, openness, and agentic robustness on the hardest tasks.
• Latent reasoning paper (Apr 23) challenges the CoT-as-reasoning frame: visible chain-of-thought may be post-hoc rationalization rather than the underlying computation.

Update (Jul 2026): the wave continued and its pricing logic inverted — see frontier §4. GPT-5.6 (Jul 9) split the flagship into three tiers; Kimi K3 (Jul 16) took the open-weight lead on size while raising price; Claude Sonnet 5 (Jun 30) shipped a 1M context with a tokenizer that bills ~30% more tokens for the same text.

What to watch: Independent benchmarks across the wave. Whether Kimi K3's open weights (due 2026-07-27) reproduce its claimed leaderboard positions. Whether open-weight competition remains cost-led (DeepSeek) or turns capability-led (Moonshot).

---

2. New Reasoning Architectures — Latent Loops + Memory Sparsity (2026)

Status: Active | Key sources: LoopFormer, Engram / Conditional Memory, Latent reasoning Key players: DeepSeek + PKU (Engram), ICLR-2026 authors (LoopFormer), broader looped/recursive-transformer community (SpiralFormer, SpiralThinker, MeSH)

Two architectural ideas are gaining traction as ways to add capability without simply scaling parameters:

• Latent / looped reasoning — LoopFormer (ICLR 2026) makes a looped Transformer's iteration count elastic and budget-conditioned via shortcut-consistency training, so the model spends more in-weights "thinking depth" when given more compute — deliberation that leaves no explicit token trace.
• Conditional memory as a sparsity axis — Engram (Cheng et al., Jan 2026) adds a constant-time static-lookup module orthogonal to MoE, with a U-shaped scaling law governing compute-vs-memory allocation. Gains skew toward reasoning (BBH +5.0, ARC-C +3.7) and long-context retrieval (NIAH 84.2→97.0).

Open problems:

• Do these axes hold at frontier (trillion-parameter) scale, and do they compose with hybrid attention?
• If deliberation moves into latent loops, what replaces CoT monitoring as a safety-auditing handle?
• Does conditional memory substitute for, or stack with, external agent-memory retrieval — and is the lookup table a knowledge-poisoning surface?

---

3. Inference Efficiency — Speculative Decoding × Quantization × KV-Cache

Status: Rapid progress | Key sources (all preprints): DSpark, Spec-AUF, Quantize the Target, Quantize the Drafter, OSAQ, VeriCache (full-text ingested); SpecMQuant / Speculative Decoding Meets Quantization, LaProx (abstract + reported-metrics, ingested 2026-07-24) Key concepts: Speculative Decoding, Low-Bit Quantization, KV-Cache Compression Key players: DeepSeek (DSpark — first-party production disclosure), LMCache / U. Chicago systems group (VeriCache), independent arXiv groups (Spec-AUF, OSAQ), Efficient Qwen Competition entrants

A cluster of 2026 papers treats speculative decoding, low-bit quantization, and now KV-cache compression as compounding levers, not separate tracks — and moves the problem from single-request latency to serving-fleet throughput economics:

• DSpark (DeepSeek, Jul 6) — semi-autoregressive drafting (a parallel backbone coupled to a lightweight sequential module, restoring intra-block dependency to fight acceptance decay) plus confidence-scheduled verification that tailors verification length per request from prefix-survival probability and engine throughput profile. 60–85% faster per-user generation at matched throughput vs. the MTP-1 production baseline, measured under live traffic in the DeepSeek-V4 serving system.
• VeriCache (May 17) — uses the compressed KV cache as a speculative drafter, verified against the full KV in parallel (drafting HBM-bound, verification interconnect-bound, so the paths overlap). Turns lossy compression lossless: up to ~4.3× throughput, output identical to full-KV (KL < 0.01 nats), ~25–40 accepted tokens/round vs 2–3 for small-model drafters; composes with EAGLE/MTP to 4.35×. Opens the KV-cache lane in this KB and is the strong form of "levers compound" — compression and speculation become one mechanism.
• Spec-AUF (Jul 2) — truncates the drafter's cross-entropy support at its first predicted failure, concentrating supervision on the prefix that actually survives verification. τ 2.40 → 2.61 (Qwen3-8B / DFlash, mean of six benchmarks, gain on every one); transfers to a second drafter family (Domino 2.56 → 2.68). No auxiliary objective, no verifier rollouts, no inference-pipeline change.
• Quantize the Target, Quantize the Drafter (Jul 5) — quantizes both target and drafter on a single A10G: 6.978× average speedup meeting required quality thresholds, 3rd overall in the Efficient Qwen Competition (ICML 2026 Workshop on AdaptFM). The clearest single datapoint that the two levers stack.
• OSAQ (May 6) — training-free, closed-form outlier suppression exploiting a stable low-rank Hessian null space; >40% lower perplexity than vanilla GPTQ at 2-bit when integrated with it, with zero inference overhead (transform folds into weights offline).
• LaProx (May 8) — broadens the lane from compression into eviction: reformulates KV-cache eviction as an output-aware, layer-wise matrix approximation with the first globally-comparable token importance score (model-wide, not per-head). Maintains performance at 5% KV-cache retention across 19 datasets (LongBench + NIAH), up to 2× accuracy-loss reduction vs prior SOTA under extreme compression. (abstract + reported-metrics grade)

Why this is the thesis frontier: every gain here comes from software on unchanged weights and unchanged silicon — a verification scheduler, a loss-support truncation, a linear-algebra transform, a cache-scheduling trick. It is the direct mechanism behind the cost decomposition in §4.

The compound-vs-conflict boundary (MEASURED + RESOLVED 2026-07-24): the compounding pattern is real but not universal — and the boundary is now measured, not flagged. arXiv:2505.22179 "Speculative Decoding Meets Quantization" (Zhang et al., May 2025, ingested 2026-07-24) measures that tree-draft speculative decoding (EAGLE-2) on a 4-bit-weight target conflicts and diminishes the quantization gain — verifying a tree draft costs significantly more than a single-token forward pass once the target is low-bit-weight — and resolves it: SpecMQuant's hierarchical small-model drafter converts tree drafts → sequence drafts, recovering 2.78× on 4-bit-weight Llama-3-70B (A100), +1.31× over EAGLE-2. The honest read, now measured: compounding holds for block/small-model drafters (Quantize-the-Target 6.978×, VeriCache's KV drafter), but drafter geometry interacts with target bit-width and at 4-bit a tree loses until made sequential. Weight-quantization only; 2-bit / activation-quant behaviour still open.

Open problems:

• Do these gains reach published API prices, or are they absorbed as provider margin?
• The compound/conflict boundary is now measured for 4-bit weights (tree drafts lose, sequence drafts recover — 2505.22179); does it hold at 2-bit and for activation/KV quantization, and is VeriCache's shared-weight KV drafter immune to the tree-vs-sequence tension?
• Does the Spec-AUF loss result hold at trillion-parameter targets (validated at 8B)?
• Is τ ≈ 2.4–2.7 near the block-drafter ceiling?
• Does OSAQ's "low-rank Hessian consistency" hold for MoE models at frontier scale, and is a perplexity-neutral transform also capability-neutral on long-tail behaviour?
• VeriCache keeps the full KV on CPU/storage alongside the compressed GPU copy — what is the context-length crossover where that storage overhead outweighs the throughput win, and does adaptive per-request draft length help heterogeneous compressors?

---

4. Inference Economics — Tiered Price Decay and the Software Attribution

Status: Active, newly compiled | Key sources: Tiered Super-Moore's Law (paper), Anthropic Platform Release Notes (first-party), Kimi K3 (analysis), GPT-5.6 (news) Key concept: Inference & Token Economics Key players: OpenAI, Anthropic, Moonshot AI, DeepSeek

The first systematic empirical study of LLM token pricing (3,237 models via Epoch AI + 318 via OpenRouter + 62 milestones, 2020–2026) documents a ~600× price decline that is tiered, not uniform: economy-tier half-life 1.10 years, mid-tier 1.55 years (both faster than the 2-year Moore benchmark), while flagship/reasoning tiers resist — reasoning carrying a ~31.5× premium. A structural break in May 2024 (F = 5.74, p = 0.005) separates a technology-driven regime from a competition-driven one.

The number that matters: cost-decline attribution is ~103.7% software/architecture, ~−0.9% GPU hardware. Hardware improvement did essentially none of the work.

July 2026 then produced three divergent moves in one month, which any single "market price" figure would have hidden: OpenAI segmented (Sol/Terra/Luna at $5/$30, $2.50/$15, $1/$6), Moonshot raised price (K2.6 $0.95/$4 → K3 $3/$15), and Anthropic held quoted price roughly flat while shipping a tokenizer that bills ~30% more tokens for the same text.

Open problems:

• What does a tokenizer-normalized price series look like? Nobody has computed one, and without it cross-vendor "token efficiency" claims are not comparable.
• Does the competition-driven regime persist, or does K3's price increase mark capability-led repricing at the open-weight frontier?
• If reasoning effort becomes vendor-set (§5 / CoT page), does per-token pricing survive as a contract, or does the unit move to per-task? (Artificial Analysis already quotes Kimi K3 at ~$0.94 per task.)
• Does the ~31.5× reasoning premium hold per-lab and per-generation, or is it a cross-sectional artifact?

---

5. The Harness Layer — Measured to Beat the Model, Productizing Ahead of the Evidence

Status: Active, and now directly measured | Key sources: Stop Comparing LLM Agents Without Disclosing the Harness (paper), Harness-Bench (paper), Harness-Induced Belief Divergence (preprint), Rethinking the Evaluation of Harness Evolution (paper), Anthropic Platform Release Notes (first-party) Key concept: Agent Harness Evolution Key players: Zhang et al. / Yao et al. / Yi & Song (harness measurement), Wang et al. (harness-evolution audit), Anthropic (Managed Agents, Dreams), OpenAI (GPT-5.4 as tested backbone)

Two questions, answered in opposite directions — keep them separate. Does the harness layer beat the model? Measured, yes. Can that advantage be auto-discovered by search and does it transfer? Audited, no.

The harness beats the model — measured (2026-07-23 compile). Stop Comparing formalizes the Binding Constraint Thesis and measures harness-induced variance at 7.80× model-induced variance (18.48 vs 2.37 pp², 6/9 ranking reversals) in a controlled 3×3 SWE-bench Verified factorial (GPT-5.4 / Kimi K2.6 / GLM-5.1 × three human-configured harnesses); swapping the harness moved GLM-5.1 by 13.0 pp, swapping the model inside a fixed harness only 3.0–5.0 pp. It proposes an ETCSOVG Harness Card disclosure standard. Harness-Bench supplies the instrument — 106 tasks, 5,194 trajectories — and names execution-alignment decoupling (reasoning drifting from tool feedback / workspace state / contracts) as the dominant failure class, concluding performance must be attributed to model-harness configurations, not models. Belief-divergence shows the harness alters an agent's multi-step beliefs even at matched terminal success — outcome parity is not harness equivalence. This is the single most direct evidentiary support for the house "edge is the harness, not the model" thesis anywhere in the KB. Grade: moderate — controlled factorial, but small n (100 tasks, 2 runs), self-measured, position paper.

Automating the harness's design still fails its audit. Wang et al. audit automatic harness evolution and find two flaws that generalize beyond harness work: (1) a budget confound — harness evolution is itself an iterative feedback-driven search, so without matched feedback and inference budget "better harness" and "more search compute" are inseparable; (2) search/eval benchmark sharing — searching and reporting on the same benchmark inflates the number, so a held-out split is required. On Terminal-Bench 2.1 (GPT-5.4, Claude Opus 4.6) harness evolution does not consistently beat simple test-time scaling and generalizes poorly to held-out tasks. Scope discipline: no arm compares harness vs no-harness, human-designed harnesses aren't evaluated — it demotes automated scaffold search, not the harness layer. The two results are consistent: the harness is where outcomes are decided (Stop-Comparing), but its advantage is not cheaply auto-discoverable and transferable (Wang et al.).

The gap that defines the frontier: while the research lands, productization accelerates. In three weeks Anthropic shipped a breaking agent-memory-2026-07-22 header migration, per-session harness override (model/prompt/tools/MCP/skills swappable without touching the base agent), and Dreams expansion to Fable 5 / Sonnet 5. The harness is sold as versioned infrastructure faster than harness disclosure (the ETCSOVG Harness Card) has been adopted by any leaderboard.

Open problems:

• Does the 7.80× ratio hold outside coding and above n=100? It is a within-study SWE-bench measurement, not a constant.
• Will any leaderboard (SWE-bench, Terminal-Bench, HAL) adopt harness disclosure or variance decomposition — or does non-disclosure persist because it flatters model marketing?
• Is there a direct head-to-head of an automated-search harness vs a well-designed human harness under matched budget? (Stop-Comparing measures human harnesses; Wang et al. measure automated search; nobody pits them against each other.)
• Does the harness-evolution null result replicate on SWE-bench, GAIA, OSWorld, and other backbones?
• If outcome-matched harnesses still diverge in belief (arrival + growth), which downstream decisions actually flip, and how large is the divergence on tasks longer than the coding stress-tests?
• Does per-session harness override turn production traffic into an implicit harness search — reintroducing the overfitting risk with real users as the feedback signal?

---

6. Agentic Reasoning & Autonomous Agents

Status: Rapid progress Key papers: Agentic Reasoning for LLMs, From LLM Reasoning to Autonomous Agents, Agentic Tool Use in LLMs Key players: Google DeepMind, OpenAI, Anthropic

The field is consolidating around a unified understanding of how LLMs function as agents. Three major survey/review papers published in Q1 2026 indicate the research community is moving past fragmented explorations toward standardized frameworks. The three-layer model (foundational → self-evolving → multi-agent) and three-paradigm tool-use framework (prompting → SFT → RL) are becoming canonical.

Open problems:

• Long-horizon interaction (multi-step plans spanning hours/days)
• Multi-agent governance and safety
• Benchmark-to-deployment gap

---

7. World Models — JEPA vs. Generative Schism

Status: Rapid progress, open architectural debate Key papers: V-JEPA 2, Genie 3, LeWorldModel, Tsinghua Survey, Embodied AI Survey Key players: Meta FAIR, Google DeepMind, Yann LeCun

The "world model" label now covers two architecturally incompatible approaches both claiming to be the path beyond LLMs. JEPA (Meta FAIR, LeCun) predicts abstract representations, not pixels — V-JEPA 2's 1M-hour video pre-training 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's 11B-param autoregressive transformer produces real-time 720p interactive worlds at 24fps with ~1 minute consistency. The next 12-24 months will pressure-test which approach scales.

Open problems:

• Can generative world models scale from minute-long to hour-long consistency?
• Does LeWM's simplification (stable end-to-end pixel JEPA with only two losses) hold at V-JEPA 2 scale?
• Which approach produces better robotic control on long-horizon tasks?
• How should hierarchical (symbolic + visual) world models be trained end-to-end?
• What's the right evaluation metric — pixel fidelity misleads; physical consistency is not yet standardized

---

8. Mechanistic Interpretability

Status: Rapid progress Key papers: Mechanistic Interpretability, Anthropic Circuit Tracing, Mech Interp for LLM Alignment Key players: Anthropic, OpenAI, Google DeepMind

Named as a 2026 breakthrough technology. The progression from individual feature identification (2024) to complete reasoning path tracing (2025-2026) represents a qualitative leap. Anthropic's Transformer Circuits Thread (2021-2026) has produced the field's deepest results: a mathematical framework for circuits, sparse autoencoders extracting millions of monosemantic features from Claude 3 Sonnet, and circuit tracing with attribution graphs that reveal end-to-end computational paths. Circuit tracing tools are now open-sourced.

A 2026 survey (Naseem, Macquarie) provides the clearest synthesis yet of how interpretability techniques translate to alignment applications: four technique categories (observational analysis, feature discovery, circuit discovery, causal intervention) map to distinct alignment objectives. The roadmap emphasizes automated interpretability, cross-model generalization, and interpretability-first architectures.

Open problems:

• Scaling circuit tracing to trillion-parameter models
• Making tools accessible beyond specialist researchers
• Detecting alignment failures proactively, not reactively
• Cross-model generalization of interpretability findings
• Dual-use risk: interpretability tools enabling improved deception
• New (Jul 2026): if deliberation budgets become vendor-set and non-configurable (§4 / §5), can an external auditor obtain a trace long enough to audit — or does CoT monitoring become a capability only the lab retains?

---

9. Agent Safety & Security — Now Empirical (+ a Government Baseline)

Status: Rapid progress (newly empirical) Key papers: Agentic AI Security & Red-Teaming, AI Safety, Alignment, and Interpretability in 2026, Exploitation Surface Taxonomy, OpenClaw Analysis, AISI Frontier AI Trends Report Key players: OPIT/Cohorte AI (Mouzouni), UC Santa Cruz / NUS / Tencent / ByteDance (Wang et al.), UK AI Security Institute

Agent safety moved from theoretical taxonomy to large-scale empirical measurement in April 2026. Three strands:

Exploitation surface is narrow. 10,000 trials across seven models show 9 of 12 hypothesized attack dimensions fail to trigger exploitation. Goal reframing (puzzle/CTF framing) is the only confirmed cross-model trigger — 38–40% exploitation on Claude Sonnet 4 while GPT-4.1 achieves complete immunity (0/1,850 trials).

Deployed agents face architectural vulnerabilities. Poisoning any single CIK dimension (Capability, Identity, Knowledge) raises attack success from 24.6% to 64–74%, regardless of backbone. The evolution-safety tradeoff is empirically documented: protections blocking 97% of malicious injections also prevent 93% of legitimate updates.

A government safety body now publishes longitudinal capability trends. UK AISI (Dec 2025), 30+ models 2022–2025: cyber-task time horizons doubling every ~8 months (accelerating to ~4.7); RepliBench self-replication <5%→>60%; frontier models surpassing the biology-PhD baseline; 40× rise in jailbreak effort — yet universal jailbreaks present in every tested system.

Open problems:

• Mechanistic basis for GPT-4.1's exploitation immunity
• Detecting goal reframing at inference time via CoT monitoring
• Architectural designs separating learning-update from injection pathways
• Cross-platform generalization of CIK vulnerability findings
• Cross-dimension attack chaining (lower bound on actual risk)
• Closing the safeguard gap that leaves universal jailbreaks in every deployed system
• Whether capability-doubling rates (cyber ~4.7mo) outpace safeguard hardening
• New (Jul 2026): do platform-imposed memory constraints (shallow depth, server ordering, segment-exact prefixes) reduce or merely relocate the Knowledge-dimension poisoning surface?

---

10. Tool-Chain Navigation — The Real Capability Gap

Status: Early stage, high impact Key papers: Amazing Agent Race Key players: University of Minnesota, Yonsei University, Google DeepMind

The field has been measuring the wrong thing. Six existing benchmarks average 55–100% linearity — invisible to the compositional failures that dominate real multi-step tasks. The Amazing Agent Race introduces DAG-structured Wikipedia navigation tasks where agents must branch independently and aggregate results. Best agents achieve 37.2% accuracy; navigation errors (27–52% of failures) dwarf tool-use errors (<17%).

Architectural efficiency matters: Claude Code matches Codex CLI at 37% using 6× fewer tokens, decoupling token budget from task performance. Read alongside §5, this is the strongest evidence in the KB that harness differences are real — and that they show up as cost at equal capability rather than as capability.

Open problems:

• Training approaches targeting fork-merge compositional reasoning
• Generalization from Wikipedia to broader multi-hop domains
• Detecting and filtering shortcut solutions (88% bypass rate at extreme difficulty)
• Planning architectures (tree search, DAG decomposition) for navigation improvement

---

11. Evolutionary Code Generation

Status: Early stage, high impact Key papers: AlphaEvolve Key players: Google DeepMind, AlphaEvolve, Gemini

AlphaEvolve demonstrates that LLM-driven evolutionary search can discover algorithms that beat human-designed ones (Strassen, 57 years). The move to semantic evolution (Gemini 2.5 Pro rewriting logic, not just parameters) is a qualitative shift. Production deployment recovering 0.7% of Google's global compute validates commercial viability.

Note (Jul 2026): AlphaEvolve and harness evolution (§5) are both automated search over artifacts, but they are not the same bet — AlphaEvolve searches against automated evaluators with ground truth (does the algorithm compute the right answer, faster?), where the harness-evolution audit's overfitting critique bites much less hard. The distinction is worth preserving: automated search works where the objective is verifiable.

Open problems:

• Extending beyond problems with automated evaluators
• Discovering fundamentally new paradigms vs. optimizing within known frameworks
• Interpretability of discovered algorithms
• Scientific hypothesis generation (beyond math/code)

---

12. Agent Evaluation Standardization

Status: Steady progress, newly disciplined Key papers: From LLM Reasoning to Autonomous Agents, Agentic Tool Use in LLMs, Amazing Agent Race, Rethinking Harness Evolution, Stop Comparing, Harness-Bench, How Inference Compute Shapes Evaluation

The ~60 benchmark taxonomy and three-tier evaluation framework represent meaningful consolidation. The Amazing Agent Race revealed a structural gap: existing benchmarks are 55–100% linear, blind to compositional navigation failures.

The 2026-07-23 compile hardens the methodological floor into a three-part disclosure standard. An agentic result is only interpretable if it discloses (1) the harness, (2) the search/inference budget, and (3) held-out generalization — and reports capability across a budget range, not at one point. Three converging findings:

• The harness is an undisclosed variable that can dominate the model — Stop-Comparing's 7.80× harness:model variance ratio and Harness-Bench's 5,194-trajectory attribution analysis make "which model scored X" unanswerable without stating the harness; the proposed ETCSOVG Harness Card is the disclosure mechanism, adopted by no leaderboard yet.
• Inference budget is an evaluation axis, not a hidden constant — How-Inference-Compute shows fixed single-budget scores increasingly understate newer models (12 models × 7 benchmarks); capability should be reported as a function of inference-time compute. This is the mirror of the matched-budget correction: unequal budgets inflate a method's apparent gain (Wang et al.), while a fixed low budget deflates a model's apparent ceiling (McFadyen et al.).
• Matched feedback/inference budget + held-out splits remain the portable floor for any feedback-driven agentic result — applying to research and vendor claims alike, and vendor claims essentially never disclose budget.

Open problems:

• Benchmarks that expose compositional reasoning (not just linear tool execution)
• Predictive validity: which benchmark performance predicts deployment robustness?
• Safety-oriented metrics beyond capability measurement
• Evaluating multi-agent collaboration scenarios
• Re-running the existing agentic literature under matched-budget, harness-disclosed, budget-swept controls
• Whether any leaderboard adopts the harness-disclosure + budget-sweep standard, or whether non-disclosure persists

---

13. Agent Memory Architectures

Status: Early stage in research, shipping in production Key papers: Memory for Autonomous LLM Agents, A-MEM: Agentic Memory, Anthropic Platform Release Notes

Agent memory is formalized as a write-manage-read loop with five mechanism families: context-resident compression, retrieval-augmented stores, reflective self-improvement, hierarchical virtual context, and policy-learned management. A-MEM (NeurIPS 2025) demonstrates Zettelkasten-inspired agentic memory outperforming fixed-structure baselines across six foundation models.

New (Jul 2026): memory crossed into versioned, billable platform surface. Anthropic's agent-memory-2026-07-22 beta header replaced managed-agents-2026-04-01, fixing server-defined ordering, restricting depth to 0/1, and requiring segment-exact path_prefix. The constraint shape reveals the design bet: hierarchical, path-addressed, shallow reads optimized for bounded retrieval cost rather than arbitrary client query.

Note: Memory is also an attack surface. The OpenClaw paper shows Knowledge-dimension poisoning achieves the highest attack success rate (74.4%), and defenses that protect memory also block legitimate learning.

Open problems:

• Continual consolidation without catastrophic forgetting
• Causally grounded retrieval (beyond similarity-based)
• Separating learning-update from external-injection pathways
• Multimodal embodied memory (visual, spatial, proprioceptive)
• Whether platform retrieval constraints rule out whole mechanism families — notably policy-learned management, which needs control over retrieval order

---

Recent Breakthroughs

Predictions & Trends

• Consolidation year: Q1 2026 saw three major survey papers — the field is converging on shared frameworks and vocabulary
• Safety-capability tension: Interpretability research is accelerating because capabilities are outpacing understanding
• Evolutionary AI as infrastructure: AlphaEvolve's production deployment suggests evolutionary code generation becomes a standard optimization tool — where the objective is verifiable
• Protocol standardization: ACP, MCP, A2A protocols indicate multi-agent systems moving from research demos to interoperable infrastructure
• Agent safety going empirical: The shift from theoretical threat models to measured exploitation rates marks methodological maturation
• Navigation training as next frontier: The compositionality gap will drive targeted training for DAG-structured reasoning
• Interpretability-first architectures: Retrofitting transparency is harder than designing for it
• Deployed agent security as product requirement: CIK findings suggest real deployments need security audits, not just capability evaluations
• NEW — Efficiency is a software discipline, and the evidence is now quantitative: ~103.7% of the token-price decline attributed to software/architecture vs. ~−0.9% to hardware, with the mechanisms (verification scheduling, loss truncation, closed-form quantization) directly observable in this compile's preprints. Expect the serving stack, not the accelerator, to remain the locus of cost improvement.
• NEW — Matched-budget becomes the credibility test: Following the harness-evolution audit, expect "did they control for inference budget and hold out tasks?" to become the standard first question of any agentic claim — and expect most vendor claims to fail it, since budgets are never disclosed.
• NEW — The buyer is losing control of deliberation spend: Anthropic removed manual thinking budgets; Moonshot shipped "max"-only reasoning. If this continues, per-token contracts weaken and per-task pricing (already quoted by Artificial Analysis for K3) becomes the natural unit.
• NEW — Open-weight strategy is bifurcating: DeepSeek competes on cost, Moonshot on capability-at-price. "Open weights" has stopped predicting "cheap."
• NEW — Tokenizers are an unpoliced pricing lever: a ~30% tokenizer expansion is an effective price rise invisible on a price sheet, and it silently invalidates cross-vendor token-efficiency comparisons.

Knowledge Gaps

Areas where the KB needs more sources:

Compiled 2026-07-23 (5 sources) — residual gaps these left open:

• Speculative decoding × quantization conflict (bounding the "compound" framing) — [CLOSED 2026-07-24 — ingested + resolved] arXiv:2505.22179 "Speculative Decoding Meets Quantization" (Zhang et al., May 2025) is now ingested (abstract + reported-metrics grade) and compiled into the Speculative Decoding / Low-Bit Quantization pages. Tree-draft EAGLE-2 on a 4-bit-weight target diminishes the quant gain (tree-draft verification costs significantly more than a single-token forward pass); SpecMQuant's hierarchical tree→sequence drafter recovers it — 2.78× on 4-bit-weight Llama-3-70B (A100), +1.31× over EAGLE-2. Was the KB's highest-priority flagged ingest. Residual (still open): 2-bit / activation-quant / non-A100 behaviour, and whether VeriCache's shared-weight KV drafter is immune.
• Belief-divergence quantitative results — belief-divergence paper compiled at abstract-grounding only; arrival/growth magnitudes and per-task numbers owed. Suggested search: "harness belief divergence BIWM arrival growth results".
• How-Inference-Compute per-benchmark grid — inference-compute-eval paper compiled at abstract level; the per-model capability-vs-budget curves are the load-bearing evidence (they turn "buy an outcome at budget B" from thesis into price) and were not captured. Suggested search: "inference compute frontier evaluation token budget curves SWE math medicine".
• VeriCache adaptive drafting + multi-compressor deployments — compiled into KV-Cache Compression, but current scope assumes static draft length, one compressor, uniform bit-precision; the adaptive/heterogeneous case is open. Suggested search: "adaptive KV cache draft length speculative verification 2026".
• Harness Card / disclosure adoption — the ETCSOVG disclosure standard is a proposal; does any leaderboard (SWE-bench, Terminal-Bench, HAL) adopt harness disclosure or variance decomposition? Suggested search: "SWE-bench harness disclosure card variance decomposition adoption 2026".
• KV-cache lane broadening — eviction addressed 2026-07-24: LaProx (arXiv:2605.07234) compiled into KV-Cache Compression as the first eviction source (5% KV retention across 19 datasets, output-aware global token scoring) — a strong alternative to the earlier-named EVICPRESS (2512.14946). Still open: rematerialization (XQuant) and long-range KV-drift (2604.26412). Suggested search: "KV cache rematerialization XQuant drift 2026 arxiv".
• Harness-Bench per-backend grid — Harness-Bench compiled at abstract depth; the enumerated model backends and per-configuration completion/quality/efficiency numbers are owed. Suggested search: "Harness-Bench model backends completion quality trajectory results".

Opened by this compile (2026-07-22) — still high priority:

• Tokenizer-normalized price series — no source computes price per character or per word; without it, every cross-vendor efficiency claim in this KB is unfalsifiable. Suggested search: "tokenizer normalization LLM pricing comparison 2026", "characters per token frontier models benchmark"
• Harness-evolution replication — the null result is one benchmark, two backbones. Suggested search: "harness evolution agent scaffold SWE-bench matched budget 2026", "agent scaffold search generalization held-out 2026 arxiv"
• Human-designed vs. automated harness comparison — substantially addressed 2026-07-23: Stop Comparing (human-configured H₁/H₂/H₃, harness variance 7.80× model variance) and Harness-Bench (5,194 trajectories across model-harness pairings) both measure human-configured harness effects. Still open: a direct head-to-head of an automated-search harness vs a well-designed human harness under matched budget. Suggested search: "automated harness search vs human scaffold matched budget 2026"
• GPT-5.6 primary system card — context window and full benchmark suite (SWE-bench, GAIA) still missing; openai.com returned 403. Suggested search: "GPT-5.6 system card model card Sol Terra Luna", "GPT-5.6 SWE-bench context window"
• Kimi K3 independent reproduction — open weights due 2026-07-27; every benchmark claim currently unverified, and the vendor-eval vs. Arena.ai conflict is unresolved. Suggested search: "Kimi K3 open weights independent benchmark reproduction", "Moonshot LatentMoE Kimi Delta Attention technical report"
• Moonshot AI primary source — moonshot.ai/blog was an nginx placeholder at ingest; no first-party technical documentation held. Suggested search: "Moonshot AI Kimi K3 technical report arxiv"
• OSAQ full results grid — only the 2-bit headline figure is held; per-model perplexity, corpora, and other bit-widths unknown. Suggested search: "OSAQ outlier self-absorption quantization results 3-bit 4-bit"
• Do serving-layer efficiency gains reach published prices? — the central open question linking §3 and §4. Suggested search: "inference cost pass-through API pricing margin 2026", "Epoch AI token price index 2026 update"
• Terminal-Bench 2.1 primary documentation — used as the key testbed but not itself an ingested source. Suggested search: "Terminal-Bench 2.1 benchmark paper agentic terminal tasks"
• Speculative decoding at frontier scale — Spec-AUF validated at 8B; no source covers block-drafter behaviour at trillion-parameter targets. Suggested search: "speculative decoding trillion parameter MoE drafter 2026 arxiv"

Carried forward (still open):

• Framework comparison benchmarks — suggested search: "LangChain CrewAI AutoGen comparison benchmark 2026"
• Embodied AI benchmarks — suggested search: "embodied AI benchmark VLA evaluation 2026"
• Multi-agent safety and governance — suggested search: "multi-agent AI governance accountability framework 2026"
• NVIDIA Cosmos world model deep dive — suggested search: "NVIDIA Cosmos world foundation model 2026 arxiv"
• V-JEPA 2 long-horizon manipulation follow-ups — suggested search: "V-JEPA 2 AC multi-step manipulation 2026 arxiv"
• GAIA-3 technical details — Wayve press release is public; arXiv paper pending
• GPT-4.1 immunity mechanism — suggested search: "GPT-4.1 safety training scope constraint agent exploitation 2026"
• Navigation-targeted training — suggested search: "compositional multi-hop agent planning training DAG 2026"
• Claude Opus 4.7 / GPT-5.5 primary technical reports — secondary coverage only; suggested search: "Claude Opus 4.7 system card 2026", "GPT-5.5 model card 2026"
• Looped/recursive transformer family — SpiralFormer, SpiralThinker, MeSH pending — suggested search: "looped transformer latent reasoning 2026 arxiv"

Closed:

• Multimodal agents · Agent safety and alignment · Anthropic interpretability research · Agent memory architectures · Agent safety empirical exploitation · Tool-chain navigation gap · Mech interp × alignment integration · World models for robotics · LeCun 2022 position paper · Wayve GAIA / commercial AV deployment · Sora / Veo physics-consistency critique · Frontier capability-trend baseline (government) · Apr-2026 frontier wave incl. open-weight · New reasoning architectures — latent loops + memory sparsity
• Inference efficiency — speculative decoding / quantization — addressed 2026-07-22, full-text ingested + extended 2026-07-23, boundary resolved 2026-07-24 (DSpark, Spec-AUF, Quantize-the-Target, OSAQ compiled into Speculative Decoding and Low-Bit Quantization; local raw files exist; the compound-vs-conflict boundary is now measured + resolved — 2505.22179 / SpecMQuant ingested, tree→sequence recovers the gain at 2.78×; residual 2-bit / activation-quant behaviour noted above)
• KV-cache compression lane (was named-but-empty) — opened 2026-07-23, broadened into eviction 2026-07-24 (VeriCache compiled into the new KV-Cache Compression concept; eviction added via LaProx (5% retention, output-aware global scoring); rematerialization / drift remain, see residual gaps)
• Inference economics / token pricing — addressed 2026-07-22, extended 2026-07-23 (Tiered Super-Moore + GPT-5.6 + Kimi K3 + Anthropic notes + How-Inference-Compute compiled into Inference & Token Economics; per-benchmark capability-vs-budget grid owed)
• Agent harness / scaffold layer — "harness > model" measured — addressed 2026-07-22, measured 2026-07-23 (Wang et al. audit + Anthropic notes, then Stop-Comparing 7.80×, Harness-Bench, and belief-divergence compiled into Agent Harness Evolution; harness-disclosure adoption + belief-divergence magnitudes owed)