PRISME -- literature verification log and novelty audit
UQO Working Paper 31 (Simon-Pierre Boucher). Audit date: 2026-10-29.
Companion files: papers.bib (77 entries), comparison_matrix.csv (67 rows).
Method. Every arXiv reference was resolved through the arXiv Atom API
(https://export.arxiv.org/api/query?id_list=...): title, full author list,
first-submission and latest-version dates, comment / journal_ref / doi
fields. Code and licences were read from the GitHub REST API
(api.github.com/repos/<owner>/<repo>, SPDX id) or from the LICENSE file.
Mechanism details for the closest neighbours were read from the arXiv HTML
full text (v-numbers given below). Nothing in this audit relies on memory
alone; where a venue is "common knowledge" but was not confirmed by a primary
source, it is flagged NOT verified here and in the note field of the
bib entry.
1. Per-reference verification log
1.1 The twelve requested core references
| # | Reference | What was fetched | Confirmed | Caveats |
|---|---|---|---|---|
| 1 | Universal Transformers, arXiv:1807.03819 | arXiv API; GitHub tensorflow/tensor2tensor |
Dehghani, Gouws, Vinyals, Uszkoreit, Kaiser; v1 2018-07-10, v3 2019-03-05; comment "Published at ICLR2019". Weight-shared recurrence over depth + ACT-style per-position halting. Code in tensor2tensor, Apache-2.0. | -- |
| 2 | PonderNet, arXiv:2107.05407 | arXiv API | Banino, Balaguer, Blundell; v1 2021-07-12, v2 2021-09-02; comment "8th ICML Workshop on Automated Machine Learning (2021)". Probabilistic halting with geometric prior, KL regulariser. | No official DeepMind code found. |
| 3 | Perceiver AR, arXiv:2202.07765 | arXiv API; GitHub google-research/perceiver-ar |
Hawthorne et al. (15 authors); v1 2022-02-15, v2 2022-06-14; comment "ICML 2022". Few latents (last positions) cross-attend causally to the long input; latents re-created per forward. Apache-2.0. | -- |
| 4 | Recurrent Independent Mechanisms, arXiv:1909.10893 | arXiv API; GitHub anirudh9119/RIMs |
Goyal, Lamb, Hoffmann, Sodhani, Levine, Bengio, Scholkopf; v1 2019-09-24, v6 2020-11-17. Top-k module activation per step; only active modules update. Repo exists, no licence file. | ICLR 2021 venue NOT verified from a primary source. |
| 5 | Mixture-of-Depths, arXiv:2404.02258 | arXiv API | Raposo, Ritter, Richards, Lillicrap, Humphreys, Santoro; v1 2024-04-02 (only version). Top-k token routing per layer with fixed capacity; unrouted tokens pass through the residual. | No official code found. |
| 6 | Huginn / Recurrent Depth, arXiv:2502.05171 | arXiv API; GitHub seal-rg/recurrent-pretraining |
Geiping et al. (9 authors); v1 2025-02-07, v2 2025-02-17; comment gives model tomg-group-umd/huginn-0125 and code repo. Apache-2.0. Zero-shot per-token exit by KL threshold. |
No venue listed on arXiv. |
| 7 | Mixture-of-Recursions, arXiv:2507.10524 | arXiv API; GitHub raymin0223/mixture_of_recursions |
Bae et al. (11 authors); v1 2025-07-14, v3 2025-10-25; comment gives code. Repo description says "(NeurIPS 2025)". Apache-2.0. | Venue taken from the official repo description, not from arXiv. |
| 8 | Ouro, arXiv:2510.25741 | arXiv API; HTML v5 (v6 HTML not rendered); project page ouro-llm.github.io; HF model cards via search | Zhu et al. (33 authors incl. Bengio, Eshraghian); v1 2025-10-29, v6 2026-10-11. Weight-tied stack, up to 4 loops; exit gate lambda_t = sigma(Linear(h_t)); Stage I entropy-regularised expected loss; Stage II (LM frozen): gate trained by BCE against labels w = sigma(k(max(0, L_{t-1}-L_t) - gamma)), k=50, gamma=0.005, i.e. labels derived from the model's own per-token loss improvement. Weights ByteDance/Ouro-1.4B, -2.6B (+Thinking) Apache-2.0 with research-only notice. | No official training code; community reimplementation rkstgr/LoopLM (no licence). No venue. |
| 9 | RecurTrace, arXiv:2609.03379 ("may not exist") | arXiv API; abs page; HTML v3; GitHub JensenYX/RecurTrace |
EXISTS. Wang, Feng, Shen, Xu, Wang, Wu; v1 2026-09-03, v3 2026-10-19. Loop Memory Attention: each looped layer attends over a window W (best W=3) of its own outputs from earlier loops at the same position; empty at loop 1 (returns zero). Halting head p_t = sigma(w^T z_t + b) on pooled state, sequence-level, BCE against oracle labels 1[min_{t'>t} L_{t'} < L_t - delta]; threshold tau chosen on held-out data. Baselines named: ACT, PonderNet, CALM, LoopUS-Conf, TaH-Mismatch. Code repo exists (Apache-2.0, created 2026-10-17). | No venue. delta value not read. |
| 10 | Elastic Latent Recurrence, arXiv:2610.12582 ("may not exist") | arXiv API; HTML v1 link scan; HF nvidia/Dory1.0-4B (HTTP 401) |
EXISTS. Ladhak, Acharya, Akle Serano, Sun, Ginsburg, Puvvada; v1 2026-10-08. Recipe: gated state update with learned step size, mixed loop counts per optimizer update, loop-index/budget conditioning (zero-initialised when added late), last-loop KV cache. Reports Koi-4B. | No code or checkpoint verified. Paper HTML links only to the nvidia/nemotron-post-training-v3 HF collection. Third-party blogs claim a public "Dory1.0-4B" release; the HF page returned 401 to the fetcher, so this is UNVERIFIED and must not be cited as the paper's release. |
| 11 | Titans, arXiv:2501.00663 | arXiv API; GitHub lucidrains/titans-pytorch |
Behrouz, Zhong, Mirrokni; v1 2024-12-31 (only version). Neural long-term memory updated at test time by surprise gradient; persistent memory tokens; MAC/MAG/MAL variants. | No official code; unofficial implementation is MIT. Cite year as 2024 (arXiv date) or 2025 consistently. |
| 12 | LayerSkip, arXiv:2404.16710 | arXiv API; GitHub facebookresearch/LayerSkip + LICENSE file |
Elhoushi et al. (13 authors); v1 2024-04-25, v4 2024-10-18; comment "ACL 2024", DOI 10.18653/v1/2024.acl-long.681. Code licence is CC BY-NC 4.0 (GitHub reports NOASSERTION; LICENSE file read). | -- |
1.2 Additional references (all verified via arXiv API unless stated)
Adaptive computation / looped / latent reasoning: Graves 2016 ACT
(1603.08983); Coconut (2412.06769, COLM 2025, code MIT); Giannou et al. 2023
(2301.13196; ICML 2023 venue NOT verified); Saunshi et al. 2025 (2502.17416,
ICLR 2025); Fan et al. (2409.15647, ICLR 2025); Relaxed Recursive Transformers
(2410.20672, ICLR 2025); Pause tokens (2310.02226, ICLR 2024); Latent tokens
(2505.12629); PonderLM (2505.20674, ICLR 2026, code Apache-2.0); Inner Thinking
Transformer (2502.13842); Encode-Think-Decode (2510.07358); Parallel Loop
Transformer (2510.24824); Think-at-Hard (2511.08577; ICML 2026 verified on
PMLR v306 pp. 32129-32154; title changed between v1 and the proceedings; code
thu-nics/TaH Apache-2.0); TaH2 (2609.35748; HTML v1 read: lookahead depth
supervision details below; abstract states code at thu-nics/TaH, no separate
TaH2 repo exists); LoopFormer (2602.11451, ICLR 2026); AdaPonderLM
(2603.01914); Latent lookahead (2603.20219); Frey et al. (2603.08391, ICLR 2026
workshop); LoopUS (2605.11011); Latent Recurrent Transformer (2605.26797);
LaTER (2605.07315, code Apache-2.0); Sharma & Vu (2606.24898);
Depth-Attention (2606.05014); DiscoLoop (2607.00341); Popescu et al.
(2607.20519); Logan (2607.14427, code MIT); Dau et al. (2609.19934); Huang et
al. (2609.21605); ReFlux (2609.32457; HTML v1 read; code gooogleshanghai/reflux
MIT); Olszowiec et al. (2610.18127).
Slots / workspaces / memory: Slot Attention (2006.15055, NeurIPS 2020, code Apache-2.0); Shared Global Workspace (2103.01197, ICLR 2022); Perceiver (2103.03206, ICML 2021); NTM (1410.5401); Memory Networks (1410.3916); Sparse Access Memory (1610.09027, NIPS 2016); Memformer (2010.06891); RMT (2207.06881, NeurIPS 2022; repo has no licence file); Token Turing Machines (2211.09119, CVPR 2023, scenic Apache-2.0); Memory Layers at Scale (2412.09764); Registers (2309.16588; ICLR 2024 NOT verified).
Long context / recurrence: Transformer-XL (1901.02860, ACL 2019, Apache-2.0); Compressive Transformer (1911.05507; ICLR 2020 NOT verified); Mamba (2312.00752, Apache-2.0).
Sparse routing: Shazeer et al. 2017 (1701.06538; ICLR 2017 NOT verified); Switch Transformers (2101.03961, JMLR).
Value / off-policy: Ernst, Geurts & Wehenkel 2005 (verified on JMLR, vol. 6, pp. 503-556); DQN workshop paper (1312.5602); Dudik, Langford & Li 2011 (1103.4601, ICML 2011); Bottou et al. (1209.2355; JMLR 2013 NOT verified); Swaminathan & Joachims CRM (1502.02362; ICML 2015 NOT verified).
Early exit / test-time compute: CALM (2207.07061, NeurIPS 2022 oral); speculative decoding (2211.17192, ICML 2023 oral); Snell et al. (2408.03314); Physics of LMs 2.1 (2407.20311).
Zero-init identity insertion: ReZero (2003.04887); ControlNet (2302.05543; ICCV 2023 NOT verified); Flamingo (2204.14198, NeurIPS 2022).
Building blocks: SwiGLU (2002.05202); RoFormer (2104.09864); GQA (2305.13245, EMNLP 2023); RMSNorm (1910.07467, NeurIPS 2019); Kaplan (2001.08361); Chinchilla (2203.15556; NeurIPS 2022 NOT verified).
1.3 Things that could not be verified or were dropped
nvidia/Dory1.0-4Bas the public release of ELR / Koi-4B: HF page not readable (401); only third-party blogs. Not in the bib.- "Solve the Loop: Attractor Models for Language and Reasoning" (arXiv 2605.12466) surfaced in a search result list only; abstract not fetched, so it is omitted.
- Official code for PonderNet, Mixture-of-Depths, Titans, Memformer, Switch Transformers, ACT, pause tokens, Shazeer 2017: none found; the bib says so.
- Venues flagged "NOT verified" above are plausible but were not confirmed by a primary source during this audit; either confirm them before submission or cite the arXiv version.
- No MCP OpenAlex access (server failed to connect); DOIs are therefore given only where the arXiv record carries one (LayerSkip) or the JMLR page was read.
2. Claim-by-claim novelty table
Legend for status: novel = no predecessor found after targeted search; partially novel = each ingredient exists, the specific combination or formulation does not; not novel = a predecessor implements the same mechanism; unknown = search inconclusive.
| Proposed claim | Closest source(s) | Common mechanism | Actual distinction | Evidence required | Status |
|---|---|---|---|---|---|
| C1. Selected local writes into a temporary slot workspace (16 slots, reset at every token, initialised S0 = e + P h_t, 1-2 slots rewritten per action) | Shared Global Workspace (Goyal et al. 2022); RIMs (2019); Perceiver / Perceiver AR (2021/2022); Slot Attention (2020); ReFlux (2026); SAM (2016); pause / latent / register tokens | Small bank of latent vectors read by cross-attention from a larger source (Perceiver, Slot Attention); write access restricted to a few selected slots or modules (Global Workspace top-k writers, RIMs top-k mechanisms, SAM sparse writes, ReFlux sparse routes); extra ephemeral computation per prediction whose outputs are discarded (pause tokens, latent tokens, registers) | Global Workspace / RIMs / SAM / TTM keep the workspace across time steps; PRISME's bank is non-persistent (reset per token) and persistence is replaced by a read-only source bank of past backbone states (like Transformer-XL's cache, not a writable memory). Perceiver / Slot Attention update all latents each iteration; PRISME updates 1-2. ReFlux selects routes per block, shared by all tokens, and writes into the residual stream, not into a per-token slot set. Pause/latent tokens are dense and unselected. No paper found that combines an ephemeral per-token slot bank with editor actions that rewrite only selected slots. | Ablations: (i) 16-slot sparse edits vs. dense update of all slots at equal FLOPs; (ii) reset-per-token vs. carrying slots across tokens; (iii) slot bank vs. plain looping of h_t (Huginn/MoR-style). Must show the sparse-write variant is on a better Pareto frontier, not merely different. | partially novel |
| C2. Action-conditional counterfactual value supervision: labels g(s,a) = loss(no edit) - loss(after edit a) collected by executing paired edit / no-edit branches, with recorded propensities; heads for cost, uncertainty, harm; greedy Q = g - lambda c - kappa sigma, Q(STOP)=0 | TaH2 (You et al., 2026-09): "lookahead depth supervision", online per-token labels = loss gain L_m - L_{m+1} obtained from a no-gradient lookahead branch, cost-sensitive BCE; Ouro Stage II (2025-10): gate fit to sigma(k(max(0, L_{t-1}-L_t) - gamma)); RecurTrace (2026-09): oracle label 1[min future loss < L_t - delta]; TaH (2025-11): token-mismatch oracle; ACT/PonderNet (compute penalty); DQN / fitted Q (greedy argmax over predicted action values); Dudik 2011, Bottou 2013, Swaminathan & Joachims 2015 (logged propensities) | Supervising a stop/continue decision with the realised change in next-token loss from actually running the extra computation is now established practice (Ouro, TaH2, RecurTrace, all 2025-26). Greedy selection by predicted value with STOP at value 0 is standard Q-learning. | All three predecessors supervise a binary continue/stop decision (one action). PRISME's label is conditional on which of 25 actions is taken (which slot or pair to edit), i.e. a counterfactual value per action, and the paper records sampling propensities of the logged actions, which none of the three do (TaH2 explicitly has no importance weighting; Ouro/RecurTrace use deterministic oracles). The cost, uncertainty and harm heads are additions. Whether propensities are actually used (IPS / DR correction) must be stated, otherwise logging them is not a contribution. | Show that action-conditional labels beat (a) a binary continue/stop decider in TaH2 style applied to the same slot editor, and (b) learning the controller end-to-end from the LM loss (MoD / ReFlux style). Report off-policy evaluation of the learned controller with the logged propensities (DR estimator) to justify the logging. | partially novel (binary version: not novel; action-conditional + propensity-logged: no predecessor found) |
| C3. Differential readout U(read(S_final) - read(S0)) added to h_t with U = 0 at init, giving an exact backbone output at zero edits | ReZero (2020), ControlNet zero-conv (2023), Flamingo zero-init tanh gates (2022); ReFlux (no active edge => frozen backbone unchanged; additive Delta payload); Mixture-of-Depths (unrouted tokens bypass exactly); TaH / TaH2 / ETD (skipped tokens keep first-pass output); RecurTrace (LMA returns zero at loop 1); ELR (gated update with learned step size, zero-init conditioning, "one-loop tax") | Zero-initialised additive branch so that the untouched model is reproduced exactly at initialisation; exact skip path when the adaptive component is not invoked; subtracting a baseline read so that "no edit" contributes exactly zero | The differential form (read after minus read before) guarantees the identity for all training time, not only at initialisation, because the two reads cancel exactly when no slot was edited; ReZero/ControlNet/Flamingo lose exactness once the gate is trained, and ELR obtains one-loop parity only empirically. ReFlux has the same all-time property (zero active edges => unchanged) and so does MoD's residual skip, so the property itself is not new; the differential-read implementation is a modest variant. | State precisely what is exact: output equality with the backbone co-trained with the editor, not with an independently trained backbone. Show that the fast path really costs zero extra FLOPs (controller cost included) and that the Pareto curve passes through the backbone point. | not novel as a mechanism (minor engineering variant); do not list as a standalone contribution |
| C4. The combination C1 + C2 + C3 | ReFlux (sparse selected additive writes with exact bypass, but per-block routing and LM-loss training); TaH2 (counterfactual loss labels + exact first-pass bypass, but whole-state dense iteration, binary action); RecurTrace (oracle stop + attention over own trajectory); MoR (token-level routed recurrence as the Pareto baseline) | Each neighbour shares two of the three ingredients. | No work found that has all three: per-token ephemeral slots with 1-2-slot edits, an action-conditional counterfactual value controller, and an exact zero-budget identity. The architecture-level claim therefore survives as a combination, but only if the empirical Pareto improvement over MoR / Huginn / a TaH2-style binary decider on the same backbone is demonstrated at matched FLOPs including controller and editor cost. | Head-to-head at matched training tokens and inference FLOPs against: dense-slot variant, binary-decider variant, Huginn-style global recurrence, MoR. Report quality-compute curves, not single points. | partially novel (novel combination; contribution is empirical) |
| C5. Fixed action catalog: 16 single-slot edits + 8 fixed-pair edits + STOP (25 actions, max 8 per token) | Switch / Shazeer top-k expert routing; RIMs top-k mechanisms; MoD top-k tokens; DQN discrete action sets | Discrete finite action set with greedy selection and a capacity cap | A particular design choice (which pairs, why 8, why max 8 actions). Nothing in the literature forbids or anticipates this exact catalog, but a fixed catalog is a hyper-parameter, not a mechanism. | Sensitivity analysis over catalog size / pair set / max actions; justify pairs (e.g. learned vs. fixed pairs). | not novel (do not claim as a contribution; present as design) |
3. How much novelty survives, and what the contribution statement should say
Frank assessment. The three mechanisms the paper proposes to claim are each preceded by well-verified work, and the field moved fast between 2025-10 and 2026-10. Learned stopping supervised by realised loss reduction from an executed extra branch is now standard: Ouro's Stage II (2025-10) fits its exit gate to per-token loss improvements, TaH2 (2026-09) calls exactly this "lookahead depth supervision" with a no-gradient counterfactual branch for already-stopped tokens, and RecurTrace (2026-09) trains a halting head on an oracle "does more depth still reduce loss" label. Exact zero-budget bypass is structurally present in Mixture-of-Depths, in TaH/TaH2 (skipped tokens keep the first pass) and in ReFlux (no active route leaves the frozen backbone unchanged), and zero-init identity insertion is a decade-old trick (ReZero, Flamingo, ControlNet). Sparse competitive writes into a small slot workspace go back to RIMs and the Shared Global Workspace, and ReFlux (2026-09) already writes a sparse, router-selected difference signal additively with exact bypass.
What survives: (i) the workspace being ephemeral per token while the only persistent object is a read-only projected source bank -- no neighbour has this split; (ii) the value controller being action-conditional over which slot(s) to edit, with per-action counterfactual labels and recorded propensities, instead of a binary continue/stop -- no neighbour does this; (iii) the combination. C3 should be demoted to an implementation property ("by construction the zero-edit path returns the backbone output exactly") rather than a contribution, and C5 to a design choice with a sensitivity study.
Suggested contribution statement (three items, no more):
- Architecture. A per-token, non-persistent 16-slot workspace edited by a shared sparse editor that rewrites one or two selected slots per action, reading a causal, read-only bank of projected backbone states; the readout is differential so the zero-edit path is exactly the backbone.
- Supervision. An action-conditional value controller trained from counterfactual edit / no-edit branches executed offline, with logged propensities, predicting per-action loss gain, cost, uncertainty and harm; actions chosen greedily with STOP at zero. Position this explicitly as the multi-action generalisation of the binary loss-gain supervision of Ouro (Stage II), TaH2 and RecurTrace, and cite them as such.
- Evidence. Quality-compute Pareto curves at matched training tokens and inference FLOPs against dense-slot, binary-decider, global-recurrence (Huginn-style) and Mixture-of-Recursions baselines on the same backbone, with ablations isolating sparse selection, action-conditional labels and the exact bypass.
If the empirical gains over a TaH2-style binary decider on the same slot editor are small, the honest framing is "a unifying formulation" rather than "a new mechanism".
4. Near-neighbours published after 2025-06 (chronological)
| arXiv | Date (v1) | Work | Why it is a neighbour |
|---|---|---|---|
| 2507.10524 | 2025-07-14 | Mixture-of-Recursions (Bae et al.; NeurIPS 2025) | token-level routed recursion; Pareto-frontier baseline; code Apache-2.0 |
| 2510.07358 | 2025-10-08 | Encode, Think, Decode (Koishekenov et al.) | iterate selected layers; per-token adaptive depth |
| 2510.24824 | 2025-10-28 | Parallel Loop Transformer (Wu et al.) | latency-neutral looping; KV sharing |
| 2510.25741 | 2025-10-29 (v6 2026-10-11) | Ouro (Zhu et al.) | loss-improvement-supervised exit gate (Stage II); weights Apache-2.0 |
| 2511.08577 | 2025-11-11 | Think-at-Hard (Fu et al.; ICML 2026) | selective extra latent iteration at few tokens; oracle-distilled decider; code Apache-2.0 |
| 2602.11451 | 2026-02-11 | LoopFormer (Jeddi et al.; ICLR 2026) | budget-conditioned elastic depth |
| 2603.01914 | 2026-03-02 | AdaPonderLM (Song et al.) | token-wise learned halting in pondering LMs |
| 2603.20219 | 2026-03-03 | Latent lookahead (Noci et al.) | extra latent compute at selected positions |
| 2603.08391 | 2026-03-09 | Adaptive loops + memory banks (Frey et al.; ICLR 2026 workshop) | learned halting + gated memory |
| 2605.07315 | 2026-05-08 | LaTER (Li et al.) | supervised halting of a latent phase; code Apache-2.0 |
| 2605.11011 | 2026-05-10 | LoopUS (Park et al.) | confidence head for early exit in recast looped LLMs |
| 2605.26797 | 2026-05-26 | Latent Recurrent Transformer (Huang et al.) | cross-token latent memory from a source layer, one forward per token |
| 2606.05014 | 2026-06-03 | Depth-Attention (Zeng et al.) | attention over earlier-layer keys at the same position |
| 2606.24898 | 2026-06-12 | Readout blind spot (Sharma & Vu) | readout design in looped models |
| 2607.00341 | 2026-07-01 | DiscoLoop (Fu et al.) | looping discrete + continuous states |
| 2607.07128 | 2026-07-08 | Distributed Sparse Interventions (Ernst et al.) | sparse latent interventions (static) |
| 2607.20519 | 2026-07-08 | Diagnosing halting gates (Popescu et al.) | evidence that jointly trained gates can underperform post-hoc readouts |
| 2607.14427 | 2026-07-15 | Per-token fixed-point convergence (Logan) | training-free exit beats trained router; code MIT |
| 2609.03379 | 2026-09-03 | RecurTrace (Wang et al.) | oracle loss-reduction halting; loop-time memory attention; code Apache-2.0 |
| 2609.19934 | 2026-09-17 | Beyond depth truncation (Dau et al.) | evaluation of depth utilisation |
| 2609.21605 | 2026-09-18 | Trading depth for time (Huang et al.) | recurrent transformers |
| 2609.32457 | 2026-09-26 | ReFlux / Write Back the Delta (Ye et al.) | sparse selected additive writes of a difference signal, exact bypass; code MIT |
| 2609.35748 | 2026-09-28 | TaH2 (You et al.) | lookahead (counterfactual) loss-gain supervision of a per-token decider |
| 2610.12582 | 2026-10-08 | Elastic Latent Recurrence (Ladhak et al.) | one-loop-tax removal; gated zero-init update; no code verified |
| 2610.18127 | 2026-10-13 | Monitoring latent reasoning (Olszowiec et al.) | analysis of looped LLM halting |
Searches run (WebSearch, 2026-10-29): "latent reasoning looped language model learned halting oracle supervision"; "Think-at-Hard selective latent iterations"; "slot workspace sparse writes per-token latent edit"; "counterfactual / intervention supervision value controller latent edits next-token loss"; "test-time latent computation adaptive recurrence per token learned stopping"; "ephemeral / scratch latent tokens per prediction"; "value head / Q-value selects which hidden state to edit STOP action counterfactual loss reduction propensity". The last two returned no paper combining per-action counterfactual labels with latent-state editing, which is the basis for the "no predecessor found" entries in Section 2. Absence of evidence after seven targeted searches is not proof of absence; a final check on arXiv listings for cs.LG/cs.CL in the week before submission is recommended.