acceptodds
Under review as a conference paper at ICLR 2027

MOEMORY: LONG-TERM USER MEMORY AS A ROUTING POLICY ON FROZEN MIXTURE-OF-EXPERTS LANGUAGE MODELS

Abstract

Long-term language-model memory is commonly implemented by injecting retrieved tokens or updating parameters: injection supplies what to say. We study an orthogonal control mechanism: selection, which governs when to say it by storing user state as a bias over the routing logits of a frozen Mixture-of-Experts model. MoEMory estimates a compact per-user route state from history, then applies either an always-on static bias or a prompt-conditioned graph readout without replaying that history. The bias enters before native scoring and top-k, changing expert selection and weights while expert parameters remain frozen. We derive the geometry of this intervention—from native route containment and top-k stability to exact base recovery on the zero-bias branch—and measure its semantic behavior directly. In a cross-backbone evaluation, static routing raises domain disposition by +0.12 on GLM-4.5-Air, with +0.31 lift for matched memory– query domains and −0.07 for mismatches. Archived 30B comparisons show lower off-topic detector activation for routing at the tested strengths. A controlled coloading sweep separates static from prompt-conditioned behavior, and a timed 30B study measures compact persistent state without history-token replay. Together, these results establish expert routing as a compact control memory for model-realizable dispositions, complementary to content carriers.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM Memory, Moe Policy.

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_thVyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper treats long-term user memory as a routing policy on a frozen MoE: an additive bias on router logits, applied before top-k, either as a static vector estimated from native top-k selection frequencies or as a prompt-conditioned readout from a token–expert graph. The authors give results on route containment and top-k stability, and report lower off-topic intrusion than prompt, LoRA and memory-expert baselines on Qwen3-30B, disposition shifts on GLM-4.5-Air and Qwen3-Next-80B, and a serving cost no higher than the base model for the static variant.

Critical strengths
  1. Paired installation protocol. The paper installs a known memory and measures, against the no-memory model, both its intended effect and its appearance on unrelated prompts. Over-personalization has been studied before, but this pairing lets routing and content memory be compared on one criterion, benefit at low intrusion.
  2. Small training-free state. The static variant needs no training and a small state (24 KiB on 30B), and does not lengthen the context. If the benefit-with-low-intrusion claim holds, this is a practical advantage over prompt memory in context length and latency.
Critical weaknesses
  1. The evidence does not separate user preference from history topic. The static state contrasts expert frequencies on the user's history with those on a neutral corpus (Eq. 3), so what it stores is which experts the history's content activates. Each disposition effect is consistent with that reading: the selective-disposition lift (Sec. 5.2, Fig. 4) measures how often the model names a person from the memory's domain, the crossed design shows only that this depends on the query, and the style transfer on Qwen3-Next-80B (Table 2) is what a French-style history would induce. The paper itself calls the metric a topic-disposition lift, yet Sec. 6 describes the router as storing user-specific preferences, and no experiment installs histories with similar topics and opposite preferences, the one comparison that would tell the two apart. Either that comparison or a narrower description of the state, as a topic disposition, is needed to support this framing.
  2. The advantages over injection and over the static variant are not shown under matched conditions. Against injection: intrusion is compared on Qwen3-30B (Fig. 2) at each method's installed strength, which Sec. 5.1 notes does not match personalization benefit, while the positive effects in Table 2 come from GLM-4.5-Air and Qwen3-Next-80B, so benefit and intrusion are never reported for one backbone side by side. Because the detector is a lexicon and no uncertainty estimates are given, the gap of 0.007 against 0.449–0.624 is as consistent with weaker memory action as with better selectivity. Against the static variant: the co-loading curves (Fig. 5) compare one static vector with a graph that stores the history, and the graph's query dependence is only in direction, since any query sharing a literal token with it receives a bias of full strength β (Sec. 3.3, Eq. 4); the curves therefore do not show that query-dependent readout, rather than more stored state, is what retains the lift. Low intrusion at useful strength is the paper's central claim, so this is the main reason for my rating; a strength sweep on one backbone should be doable within the rebuttal period.
Questions for authors
  1. Can histories that share topic vocabulary but state opposite preferences be installed, and does the bias move behavior in the stated direction on unseen queries, compared with a shuffled-history bias of equal norm?
  2. On GLM-4.5-Air, how do target lift and off-topic intrusion compare when the strength of routing, prompt memory and a relevance-gated retrieval arm is swept, and which β underlies the reported results?
  3. Can the co-loading experiment include a baseline that stores one static vector per memory and selects it by query match, at the same storage as the graph?
  4. How many runs underlie Table 2 and Figs. 2, 4 and 5, what are their intervals, and do the intrusion scores in Fig. 2 hold under a non-lexical judge?
  5. Table 3 shows the static state faster than the no-memory model (first token 352.6 vs. 414.2 ms) although it adds a bias operation; what explains this, and was warm-up controlled across arms?
  6. Eq. 4 yields one vector over experts while the bias is layer-specific (Sec. 3.1). How is the readout assigned to each layer, and is the dynamic bias computed once from the prompt or updated during decoding?
2 comments

Q1. Do opposite preferences on the same topic move behavior in the stated direction?

Not consistently. We tested opposing preferences about the same topic against neutral-history and equal-norm wrong-user controls. Results at beta=3 are below.

Histories about statistics and computing state opposing preferences for length, formula-first explanations, bullets or answer-first order. Three variants per topic and attribute are tested on unseen query families. The 288 cases comprise 96 positive, 96 negative and 96 neutral histories, giving 24 positive/negative pairs per attribute with three seeds. Wrong-user histories are reassigned within each topic/attribute group at the same routing norm, beta=3.

AttributeCorrect-history pole differenceWrong-user pole difference
Length+0.010+0.024
Formulas before intuition0.000+0.069
Bullet formatting+0.069+0.028
Answer before reasoning−0.056+0.097

Correct histories have a larger effect on bullets but smaller effects on the other attributes. Every answer reaches the 192-token budget, so these scores describe generated prefixes. Appendix D.2 gives intervals and scoring rules.

The bullet differences are +0.069 [0.000, 0.139] for correct histories and +0.028 [−0.028, 0.097] for wrong-user assignments. This ordering alone does not establish a significant advantage. Answer-first moves in the opposite direction at beta=3. At beta=6, differences are −0.012 for length, 0.000 for formula order, +0.028 for bullets and +0.014 for answer order, all with intervals spanning zero. Base and neutral pole differences are zero because their inputs and states are shared.

The control permutes history-to-user assignments while preserving each history's words and order. We did not run word shuffling. Preference direction is not reliably recovered on these four GLM attributes, and the revision describes the observed behavior as history-conditioned dispositions.

The separate 80B tests find changes in language, register, formatting and reading level, with user-turn and assistant-turn histories compared directly. They use different history construction and scoring from this GLM control. Both experiments are reported. Our intended application is to retain a way of answering that the frozen model already knows.

Q2. What happens at matched benefit, with a gated retrieval arm?

We ran the GLM grid with base and relevance-gated retrieval. Routing produces much smaller target effects than prompting or active retrieval. No routing setting qualifies for a matched-positive-benefit comparison.

Eight held-out histories cover football, basketball, tennis and motorsport. Each setting has 144 target-domain and 384 unrelated cases, with separate validation histories and question families. We sweep four routing strengths, prompt record counts and retrieval thresholds. At beta=3, static gives +0.026 [−0.017, 0.083] and PPR +0.040 [−0.094, 0.218]. Three-record prompting gives +0.815 [0.681, 0.919], and active retrieval +0.797 [0.665, 0.910].

Retrieval inserts positively matched records at threshold zero and abstains at the higher tested thresholds, returning zero benefit and intrusion. PPR applies nonzero bias even on unrelated queries. Its low intrusion occurs with memory active. Section 5.1 reports both outcomes, with matching rules in Appendix D.1.

Q3. Can the co-loading test separate query dependence from stored state?

We added a query-selected static vector bank and a fixed readout of the same PPR graph. The fixed readout changes only the seed. On generic queries its lift is 0.000 throughout, while PPR ranges from −0.056 to 0.000. At eleven distractors on country-constrained queries, PPR gives +0.037 versus −0.037 for fixed readout. PPR's advantage depends on the query type.

The bank stores one static state per history and selects by IDF-cosine query matching. It returns zero bias without overlap and breaks ties by history hash. Fixed-graph readout retains the PPR transition matrix and seeds all retained history tokens uniformly, holding graph storage fixed.

Capped sum adds independently normalized signatures with a per-layer cap of 2 beta, without query selection. Its lift is +0.333 at four and seven distractors and +0.222 at eleven on generic queries. Concatenated-history static stays near zero. Appendix D.3 reports all five loads, six generic and eighteen country-constrained cases per load, three seeds and history/query-resampling intervals.

At eleven distractors, the stored numeric arrays are:

StateBytes
Single static state23,040
Query-selected vector bank276,480
PPR / fixed graph1,465,472

These sizes exclude the bank's text retrieval index. Appendix D.3 gives serialized sizes. Only the fixed-graph/PPR pair holds complete graph storage exactly equal.

The revised cost description includes graph storage and query-time propagation. Static state has fixed size once installed.

Q4. How many runs support the added tables, and do the intrusion results hold with a non-lexical judge?

Appendix D.1–D.4 gives strengths, decoding settings, intervention layers and interval definitions. Each added point is linked to its generated answers. Sampling is summarized below.

Added experimentHistories/questions and repetitionsInterval unit
GLM joint benefit/intrusionEight test histories, 144 target + 384 unrelated cases per setting, five seedsDomains, histories and shared query families after seed averaging
Same-topic preferences24 paired cases per attribute, three seeds, positive/negative and neutral conditionsSix synthetic topic/history variants after seed averaging
GLM co-loadingFive loads; six generic + eighteen constrained cases per load, three seedsHistories and queries after seed averaging
80B response attributes12 held-out history pairs per attribute, three generations per query, seed 0History pairs

The added GLM records cover 214 main jobs and 15 capability jobs, totaling 61,248 response rows. Steering supplies 12 separate runs. Original figures retain their own protocols and seed counts. Appendix D.3 also reports the three existing 80B capped-sum seed curves individually.

The semantic evaluator gives zero mean intrusion for static/PPR beta=3, compared with 0.888 for three-record prompting and 0.493 for active retrieval. Its ordering agrees with the lexical detector on the new GLM test.

Prompting has per-response intrusion 0.242 and at-least-one-in-five intrusion 0.646. Active retrieval gives 0.169 and 0.404. The latter counts cases with any detection across five answers. The evaluator sees no method label or installed preference pole. Unique question/answer/rubric triples are judged once and scores mapped to originating responses. All frontier scores parse. The 25 invalid unique preference judgments are excluded with valid counts reported. The evaluator has no human calibration.

Q5. Why was static faster than base, and was warm-up controlled?

We repeated the 30B timing test with six discarded warm-ups per arm and three independent process starts. Each process uses four prompts and five repeats, randomizing arm order within blocks and synchronizing CUDA.

ProcessStatic minus base: first token (ms)Static minus base: turn (ms), 95% CI
First−7.07+24.32 [20.82, 27.85]
Second−6.95+27.23 [24.28, 30.08]
Third−7.30+24.88 [20.47, 29.11]

Static retains a smaller first-token advantage, whose cause remains unresolved. Full-turn time increases by about 1.4%. Table 5 uses these repeated measurements, with intervals in Appendix D.5.

Table 5 reports medians of three process-level medians. No-memory first-token/turn times are 109.3/1799.1 ms, versus 101.8/1825.1 ms for static at 32 tokens. The paired estimates above match prompts and repetitions across arms within each process. Their intervals bootstrap repetitions.

The measurement uses an exclusive RTX 6000D, PyTorch 2.13.0+cu130 and CUDA 13.0. Each process records 120 turn and 18 session observations across six arms after warm-up. Earlier measurements remain in Appendix B.5. Context length is unchanged. Installation and PPR propagation are excluded.

Q6. How is one graph readout assigned to layers, and is it updated while decoding?

The graph pools expert evidence across intervened layers. PPR computes one normalized vector per prompt and broadcasts it to those layers, holding it fixed during decoding. Static retains a separate vector per layer. The GLM runs and Qwen graph construction use this implementation.

Max-normalization fixes the largest component at beta for supported prompts. Prompts change the direction, with no separate semantic-relevance gain. Section 3 states the layer assignment and update schedule.

The graph connects tokens by local co-occurrence, tokens to routed experts, and experts by co-activation. Overlapping prompt token IDs form the seed. After 30 PageRank iterations with restart alpha=0.15, we normalize the expert coordinates by their maximum and multiply by beta. Empty seeds stay zero. A nonzero readout receives the same maximum gain even for semantically distant questions. GLM broadcasts the 128-dimensional vector across all 45 MoE layers.

The bias precedes native scoring and selection. GLM retains sigmoid scoring, correction-bias and grouped selection. Neither the graph nor the bias is updated during generation.

Reviewer_oowEOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper implements user memory through biases on frozen MoE router logits, changing expert selection and weighting without replaying history. It proposes static biases derived from historical expert frequencies and dynamic biases obtained through Personalized PageRank.

Critical strengths

1.Clear conceptual framing: “memory as selection” versus “memory as injection.” 2.Honest disclosure of heterogeneous and negative results across backbones.

Critical weaknesses

1.Serious citation errors. The OPPU paper is cited as “Zhaoxuan Tan, Qingkai Liu, and Meng Jiang.” Its actual authors are Zhaoxuan Tan, Qingkai Zeng, Yijun Tian, Zheyuan Liu, Bing Yin, and Meng Jiang. The incorrect author name and incomplete list raise serious concerns about citation verification. The bibliography needs a thorough audit.

2.Unmatched personalization benefit. The headline intrusion comparison uses Qwen3-30B, where the separately reported topic lift is +0.00. Low intrusion may therefore reflect a weak intervention rather than better memory selectivity. Comparisons at equal target benefit are missing.

3.Insufficient statistical reporting. Most behavioral results lack confidence intervals. Domain results are summarized across only two seeds, and co-loading curves lack response-level records and per-point seed counts. The reliability of the reported gains cannot be adequately assessed.

4.Static-memory interference and heterogeneous effects. Static lift falls from +0.326 with zero distractors to +0.076 with two and +0.010 with four. Dynamic memory retains approximately +0.33 through four distractors. This demonstrates static interference, but not a proven capacity limit. Cross-backbone efficacy remains insufficiently explained.

5.Limited theory–experiment connection. The analysis establishes computational properties but does not explain semantic efficacy or interference. The zero-bias equivalence guarantee also does not establish protection on unrelated natural queries: both reported natural-query audits activate the lexical support rule throughout.

6.Restricted baseline coverage. OPPU-style implements only per-user LoRA, while MT-style is a locally inspired approximation. These disclosed limitations restrict comparative conclusions to the tested implementations; relevance-gated retrieval and direct steering baselines are missing.

Questions for authors

See the weakness

2 comments

Citation correction

We have corrected the citation.

Benefit, intrusion and uncertainty

The completed GLM sweep is in the shared table. Its routing and content arms have substantially different benefits. The experiment therefore leaves the equal-benefit question open. Appendix D.1 includes every tested strength and the retrieval settings that abstain.

For the routing arms, the complete held-out strength results are:

betaStatic lift [95% CI]PPR lift [95% CI]
1.5+0.014 [−0.004, 0.047]+0.024 [−0.050, 0.128]
3+0.026 [−0.017, 0.083]+0.040 [−0.094, 0.218]
4.5+0.039 [−0.008, 0.101]+0.028 [−0.131, 0.217]
6+0.051 [−0.004, 0.119]+0.029 [−0.154, 0.246]

Each point uses the same 144 target cases and 384 unrelated cases with five seeds. Static lift increases with strength, although its intervals include zero. PPR is not monotonic. The +0.12 GLM result and +0.083 80B result use the original broader domain-choice protocol. They belong to a different experiment from both this sweep and the submitted 30B intrusion comparison.

We also separate the detector's two denominators:

MethodPer-response intrusionAt least one detection in five
Three-record prompt0.2420.646
Active gated retrieval0.1690.404

The added experiments retain generated answers, per-point sample counts and intervals. We report the original figures under their original protocols and assign the new intervals to the new runs.

For uncertainty, the joint experiment first averages the recorded seeds, then resamples domains, histories within domains and shared query families. The preference control resamples its six topic/history variants after seed averaging. The added 80B attribute table resamples 12 history pairs per attribute and states the sampling seed. Each interval therefore follows the repeated units in its experiment. The numerical supplement provides configurations, response files and run-to-table mappings for recalculation of the added GLM results.

Static interference and the capacity interpretation

We ran static, capped-sum, query-selected vector-bank, PPR and fixed-graph controls at beta=3, with three seeds. The fixed-graph arm uses the same graph as PPR and replaces its query seed with a uniform seed over retained history tokens.

Query conditionPPR liftFixed-graph liftCapped-sum lift
Generic queries, across five loads−0.056 to 0.0000.000 throughout+0.333 at four and seven distractors
Country-constrained queries, eleven distractors+0.037−0.037See full curve in Appendix D.3

PPR's advantage in the original curve does not extend to the new generic queries. The country-constrained queries favor a different ranking. We report the task and composition rule with each curve. The rank statement bounds the dimension of the bias. It gives no particular number of memories at which retention must fail. The cross-backbone comparison also varies model size, expert count and router family together, so it does not isolate the effect of expert count.

The full generic-query curve compares the combination rules directly.

State / distractors024711
Merged static0.0000.0000.0000.000+0.056
Capped sum0.000+0.111+0.333+0.333+0.222
Query-selected bank0.0000.0000.0000.000−0.333
PPR0.0000.000−0.0560.0000.000
Fixed graph0.0000.0000.0000.0000.000

Each load has six generic history–query cases with three seeds. The base target-entity rate is 0.500. Capped sum normalizes each history signature separately, adds the signatures and applies a per-layer norm cap of 2 beta. Merged static estimates one state from concatenated history. The capped-sum mean at four distractors is positive, with an interval of [0.000, 0.667] over histories and queries in this small panel.

We added an existing 80B capped-sum experiment with 32 queries and 24 generations per query. Its three seed-level lifts at eleven distractors are +0.069, +0.056 and +0.062. They start at +0.204, +0.199 and +0.195 with zero distractors. The target effect weakens but remains positive in all three seeds. These curves measure retention under specific combination rules. Section 5.3 includes the GLM controls and 80B endpoints, with the separate protocols and curves in Appendix D.3.

Zero bias on natural queries

PPR applies a nonzero bias on every history–query case in the new grid, including unrelated queries. Its low intrusion occurs with memory active. Token overlap admits those unrelated questions as well as relevant ones. The exact-recovery theorem applies when every routing bias is zero and execution conditions match. Section 3 gives the trigger rule, and Appendix A.6 reports natural-query activation.

We test the zero-bias path with eight tokenizer-disjoint inputs. It reproduces base greedy text and all 151,936 first-token log probabilities. A supported positive control changes the output. This verifies the implementation's OFF branch. Literal token support is too broad to guarantee semantic relevance on natural queries. We removed the implication from route containment to semantic non-intrusion because changing the selection of frozen experts can still increase the probability of an unwanted answer.

Added baselines and a non-lexical evaluator

Relevance-gated retrieval is included in the full sweep. Residual-stream steering uses a fixed 64-case subset of the same task with three seeds. Its strongest tested gain gives +0.042 entity-rate lift against its own zero-gain backend, with zero lexical intrusion.

The retrieval arm uses IDF-cosine similarity, selects up to three positively scored records and sweeps thresholds 0, 0.1, 0.2 and 0.3. It can decline to insert a record. The steering arm takes a history-minus-neutral middle-layer residual direction, normalizes it against the neutral norm and applies gains 0, 0.01, 0.03 and 0.1 during prefill and decoding. Its target rate is 0.167 at zero gain, stays there at 0.01 and 0.03, and reaches 0.208 at 0.1. Its own zero-gain reference avoids attributing a backend difference to the intervention.

For semantic intrusion we used memory-free Qwen3-Next-80B, with method labels withheld:

MethodMean semantic intrusion
Static/PPR, beta=30.000
Three-record prompt0.888
Active retrieval0.493

The semantic evaluator scores more intrusion under active content injection. All frontier judgments parse successfully. The 25 invalid unique judgments affect preference scoring and are excluded there. Appendix D.2 reports the valid-score counts.

The evaluator receives the question, generated answer and scoring rubric. It sees no method label or memory prompt. The intrusion rubric identifies the installed domain. It decodes greedily and judges each unique question/answer/rubric triple once. The score is then assigned to every originating response. There are 22,335 valid judgments out of 22,360 unique judgments. All 30,720 frontier intrusion scores are valid. The semantic evaluator and lexical detector give the same ordering on intrusion. The joint benefit table reports the larger target effects of the active content arms alongside these scores.

Reviewer_r8BzOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper trys to encode user memory as a bias factor that is added to MoE router logits before top-k, so it get the desired effect with nothing required to go into the prompt and no weights needed to get updated. There’s a static version of the design (one vector per layer, derived from expert selection frequencies) and a dynamic one (PageRank over a token-expert graph from history). They test the method on various models. including Qwen3-30B, GLM-4.5-Air, and Qwen3-Next-80B. The author's results show low off-topic intrusion relative to prompt-based and LoRA memory, modest topic/style gains, and the dynamic variant handles multiple loaded memories better.

Critical strengths

Storing memory in the router sounds novel and simply lightweight. The cost is tiny since the backbone stays frozen, and the authors were right to measure things like off-topic intrusion as a separate axis.

Critical weaknesses

The central problem is that intrusion and baselines are only compared on 30B variant. A method that does nothing will trivially show low intrusion. The authors do acknowledge this isn’t a matched comparison, but the paper’s main selling point that I can remember still just rests on exactly that comparison being meaningful.

Beyond that, the evidence is just not complete enough. Effects are small and bounce around across backbones, there are no confidence intervals or seed counts anywhere but could matter in the setting, some raw data is apparently lost, and I cannot find the the beta values used.

The presentation uses too much math but it is not really necessary for it to be clear. The theorems read as restatements of definitions and don’t actually connect to any of the empirical results.

Writing feels like very AI and not easy and direct enough.

Questions for authors
  • Can you show intrusion and benefit at matched strength on a backbone where MoEMory actually does something, like GLM-4.5-Air?
  • Does the always-on static bias degrade general capability?
  • What beta values did you use, and how many samples/seeds back each number?
1 comment

Q3. Which beta values, samples and seeds underlie the added results?

ExperimentStrengthsSampling
GLM joint sweepStatic/PPR beta=1.5/3/4.5/6528 test cases per point; one response per seed; seeds 101/211/307/401/503
Co-loading and preference controlsbeta=3; preference also beta=6Seeds 101/211/307
Same-state capabilityStatic beta=1.5/3/4.5/6 and base384 cases per condition; seeds 101/211/307
Added 80B response attributesStatic beta=312 synthetic history pairs per attribute; three generations per query; seed 0

Routing intervenes in all 45 MoE layers, using GLM's native routing rule. For the joint-sweep intervals we average seeds, then resample domains, histories and shared query families. Appendix D gives decoding settings and denominators. The numerical supplement maps 229 main/control runs and 61,248 response rows to configurations and inputs. Repeated seeds and questions remain repeated observations of those histories, not additional users.

For the joint GLM experiment, decoding is temperature 0.8, top-p 0.9, disabled thinking and a 64-token budget. The preference controls use a 192-token budget. Frontier intervals use 2,000 draws after averaging the five seeds, resampling domains, histories within domains and shared query families. Preference intervals use 4,000 draws over synthetic topic/history variants after seed averaging. This avoids treating repeated generations of the same history and question as independent users. New intervals belong to these added experiments, not retrospectively to the original figure aggregates.

Mathematics and presentation

Section 3 now defines the per-user bias and its placement before giving the static and dynamic constructions. The shifted logits feed both scoring and expert selection. The earlier definition appeared to change only the selected set, although the argument also changed mixture weights. The definition and analysis now use the same intervention.

At a fixed layer input, the top-k membership condition is 2 epsilon < Delta, where Delta is a logit margin. The local rank bound counts continuous control directions. We removed its interpretation as a bound on independently stored memories. Exact base recovery requires zero bias at every routed invocation and matched execution conditions. Appendix C gives the proofs. Target effects and retention are measured in the experiments.

We rewrote the results as connected paragraphs and expanded the main results tables. The original concept diagram and empirical plots are retained. The introduction starts with language and answer-style examples before introducing the formal construction.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.