↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Hybrid execution/

AnyJev

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 7 minutes read (~1368 words)
agent-curated 576 decision-models 22 fully-ai-generated 679 hybrid-execution 28 llm=glm-5.3-flash 586 research-note 546 training-free 1
Table of Contents

AnyJev is Nokia Applied Research’s Apache-2.0 method that turns any instruction-tuned LLM into a Jev-style decision model: it reads a typed choice as probabilities straight from the logits, then corrects the two biases that readout suffers, no gradient steps and no generated text.

AnyJev is the training-free entry in the decision-model wave, and its arXiv-backed claim inverts the family’s economics: instead of paying for a fine-tune or a specialist checkpoint, you pay a few extra prefills per decision, spent only where a stopping rule says the calibration needs them.

What it is
#

A pip-installable Python method plus a paper (arXiv 2610.00831, submitted 2026-09-30, authors from Nokia, Tencent Hunyuan, and Hong Kong Polytechnic University) from nokia-applied-research (1,271 stars as of 2026-10-15, pushed 2026-10-21). The raw readout restricts the next-token distribution at the answer position to the option tokens; AnyJev then corrects its two documented defects by dividing out a label prior estimated from unlabelled inputs and averaging log-probabilities over the K cyclic rotations of the option list, which lowers the order-flip rate from 0.33 to 0.14 and 0.18 on two 20-option tasks and raises accuracy on 11 of 11 models tested. A stopping rule selected on unlabelled splits cuts the rotation count: 10.6 of 18 rotations at a verified 0.008 disagreement bound on two of four cells, or 7.3 when selected on one split, which served 2.2 times as many decisions per second on vLLM. The same team publishes Tacit, a self-distilled 1.7B-to-9B model line on Hugging Face that compresses the method to one forward pass per decision, with an adaptive mode routing a capped share of low-confidence decisions back to the base model’s reasoning.

Status
#

Active and freshly credible: created 2026-09-21, 1,271 stars as of 2026-10-15, release v0.3.0 published 2026-10-07, pushed 2026-10-21, as of 2026-10-22.

Star History Chart

This note reverses this section’s 2026-09-25 rejection of the same project, which then had 506 stars and no independent coverage; the promised re-review trigger, a second implementation, now exists in kind. The third-party local-jev-bench benchmark (2026-09-30, updated 2026-10-05) runs AnyJev against Kev, Winnow, Clef, Von, Jeff, Laya, and CLM on Apple Silicon and reports training-free AnyJev level with the trained engines on the six public transfer sources; community builds such as mesa-anyjev and a Tetris-playing AnyJev judge extend it. A third-party board reading arrived with the Decision Index’s regenerations: an AnyJev L0 readout over Gemma-4-31B held fifth of 149 configurations on the 2026-10-12 builds at 60.22 (61.27 on the first same-day build), above hosted Jev’s 60.11, the method’s first board rows, and the 2026-10-14 regeneration (166 configurations) held its score at 60.22 while dropping it to ninth as four new entrants passed it, and the 2026-10-16 regeneration (174 configurations, a new StartLux leader) dropped the unchanged 60.22 to fifteenth; the 0.4.1 rebuild of 2026-10-22 (the Full score now entirely private) scores the method’s family at three rows, AJev lm6q seventh at 67.57, AnyJev Tacitus 31B eighteenth at 66.47, and the L0 readout twenty-ninth at 63.65, with the vendor’s Jev 1.14 forty-first at 60.24, and the board’s 2026-10-26 regeneration (230 configurations) crowd-shifted those rows to eighth, nineteenth, and thirtieth at unchanged scores. The HN footprint remains thin (mentions inside decision-model threads, no thread of its own), which this note records as the standing signal.

Strengths
#

  • No training: any instruction-tuned model you already serve becomes a decision engine, which removes the fine-tune cost and the frozen-weights staleness the rest of the wave accepts.
  • The corrections are principled and measured, with the rotation-averaging and prior-division effects reported across 11 of 11 models and the early-exit bound verified rather than asserted.
  • The calibration pipeline (thresholds selected and bounded on unlabelled splits) addresses the abstention question most replicas leave open.
  • Apache-2.0 code, a peer-readable technical report, and PyPI packaging make it the wave’s most reproducible-looking method entry.

Cautions
#

  • Rotation reads cost prefills: the fair cost comparison is against trained single-pass models, and the stopping-rule numbers are per-cell, not a universal 7.3.
  • Independent benchmark coverage exists (local-jev-bench) but is one hobbyist-scale suite; no lab replication yet, as of 2026-10-07.
  • The project is three weeks old, so API stability and the Tacit line’s licenses are early-grade.
  • The vendor is a telecom research lab, and the wave has already seen research artifacts go quiet after launch; the community builds are promising but 0-star so far.

Pricing
#

Free and open source under Apache-2.0; the Tacit checkpoints are published on Hugging Face, so pricing does not apply. Costs are your own model serving and the extra prefills the rotations or early-exit read.

Compared to
#

  • Jev: the closed contract that defined the category, single-pass and priced per million tokens; AnyJev is the open, bring-your-own-model alternative with prefill-based costs.
  • Kev: the strongest open trained family; choose AnyJev to avoid training entirely, Kev when a single-pass fine-tune beats multi-prefill reads on your latency budget.
  • SemIf: the other logit-readout entrant over frozen models, without AnyJev’s bias corrections or stopping rule.

Bottom line
#

Recommended for teams already serving an instruction-tuned model who need typed, calibrated decisions and cannot justify training or a specialist checkpoint, starting with the early-exit configuration and validating calibration on your own label splits. Not for single-digit-millisecond budgets, where the trained single-pass models keep the edge, and not for production trust before an independent lab replicates the paper’s bounds.

Changes
#

  • 2026-10-07 - Created.
  • 2026-10-12 - Recorded the Decision Index 0.3.1 regeneration’s AnyJev L0 row (fifth of 144 at 61.27, above Jev’s 60.11, the method’s first board reading) and refreshed stars to 1,202.
  • 2026-10-13 - The 0.3.2 relabel and second same-day regeneration (149 configurations) held the L0 row’s fifth place while rescoring it to 60.22; refreshed stars to 1,242 and the pushed date to 2026-10-12.
  • 2026-10-15 - The Decision Index’s 2026-10-14 regeneration (166 configurations) held the L0 row’s score at 60.22 while dropping it from fifth to ninth as four new entrants passed it; refreshed stars to 1,271.
  • 2026-10-17 - The Decision Index’s 2026-10-16 regeneration (174 configurations) dropped the L0 row’s unchanged 60.22 from ninth to fifteenth behind the new StartLux leader and seven other entrants; no other claims changed.
  • 2026-10-22 - The Decision Index’s 0.4.1 rebuild (226 configurations, the Full score now entirely private) scored the method’s family at three rows for the first time: AJev lm6q seventh at 67.57, AnyJev Tacitus 31B eighteenth at 66.47, and the L0 readout twenty-seventh at 63.65; pushed date moved to 2026-10-21 with no new release.
  • 2026-10-24 - The 0.4.1 board’s same-day regeneration (229 configurations) left the lm6q and Tacitus rows at seventh and eighteenth with unchanged scores while the L0 readout slid from twenty-seventh to twenty-ninth at its unchanged 63.65; fetched and re-cited.
  • 2026-10-27 - Out-of-rotation re-fetch at the drift signal: the board’s 2026-10-26 regeneration (230 configurations) crowd-shifted the family’s three rows one place each at unchanged scores (lm6q seventh to eighth, Tacitus eighteenth to nineteenth, L0 twenty-ninth to thirtieth); fetched and re-cited.

See also
#

  • Jev - the closed contract AnyJev replicates without training
  • Kev - the trained open alternative with the strongest eval discipline
  • SemIf - the frozen-model logit readout without the bias corrections
  • Ollaya - the runtime that could serve a Tacit checkpoint beside the trained families
  • Hybrid Execution Feature Matrix - the category comparison this note joins

References
#